{"id":479915,"date":"2026-05-15T18:40:16","date_gmt":"2026-05-15T18:40:16","guid":{"rendered":"https:\/\/savepearlharbor.com\/?p=479915"},"modified":"-0001-11-30T00:00:00","modified_gmt":"-0001-11-29T21:00:00","slug":"","status":"publish","type":"post","link":"https:\/\/savepearlharbor.com\/?p=479915","title":{"rendered":"Axera AX650N: \u0430\u0440\u0445\u0438\u0442\u0435\u043a\u0442\u0443\u0440\u0430 Edge ML SoC \u043f\u043e\u0434 CNN, LLM \u0438 VLM"},"content":{"rendered":"<div xmlns=\"http:\/\/www.w3.org\/1999\/xhtml\">\n<figure class=\"full-width \"><img decoding=\"async\" src=\"https:\/\/habrastorage.org\/r\/w1560\/getpro\/habr\/upload_files\/e98\/d1c\/e1d\/e98d1ce1d1d9fd9cf9cbb1f2c7c3bb3c.png\" alt=\"Maix4 HAT\" title=\"Maix4 HAT\" width=\"2048\" height=\"1365\" sizes=\"auto, (max-width: 780px) 100vw, 50vw\" srcset=\"https:\/\/habrastorage.org\/r\/w780\/getpro\/habr\/upload_files\/e98\/d1c\/e1d\/e98d1ce1d1d9fd9cf9cbb1f2c7c3bb3c.png 780w,&#10;       https:\/\/habrastorage.org\/r\/w1560\/getpro\/habr\/upload_files\/e98\/d1c\/e1d\/e98d1ce1d1d9fd9cf9cbb1f2c7c3bb3c.png 781w\" loading=\"lazy\" decode=\"async\"\/><\/p>\n<div><figcaption>Maix4 HAT<\/figcaption><\/div>\n<\/figure>\n<p>\u0411\u043e\u043b\u044c\u0448\u0438\u043d\u0441\u0442\u0432\u043e \u0437\u0430\u0434\u0430\u0447 \u0441\u043e\u0432\u0440\u0435\u043c\u0435\u043d\u043d\u043e\u0439 \u0440\u043e\u0431\u043e\u0442\u043e\u0442\u0435\u0445\u043d\u0438\u043a\u0438 \u0442\u0430\u043a \u0438\u043b\u0438 \u0438\u043d\u0430\u0447\u0435 \u0437\u0430\u0432\u044f\u0437\u0430\u043d\u044b \u043d\u0430 \u043d\u0435\u0439\u0440\u043e\u043d\u043d\u044b\u0445 \u0441\u0435\u0442\u044f\u0445: \u0434\u0435\u0442\u0435\u043a\u0446\u0438\u044f \u043e\u0431\u044a\u0435\u043a\u0442\u043e\u0432, \u043e\u0446\u0435\u043d\u043a\u0430 \u0433\u043b\u0443\u0431\u0438\u043d\u044b, \u043b\u043e\u043a\u0430\u043b\u0438\u0437\u0430\u0446\u0438\u044f, \u043f\u043b\u0430\u043d\u0438\u0440\u043e\u0432\u0430\u043d\u0438\u0435. \u0412\u0441\u0451 \u044d\u0442\u043e \u0440\u0435\u0441\u0443\u0440\u0441\u043e\u0451\u043c\u043a\u043e, \u0438 \u0432\u043e\u043f\u0440\u043e\u0441 \u0432\u044b\u0431\u043e\u0440\u0430 \u043a\u043e\u043c\u043f\u0430\u043a\u0442\u043d\u043e\u0433\u043e \u0432\u044b\u0447\u0438\u0441\u043b\u0438\u0442\u0435\u043b\u044f (\u0434\u043e\u0441\u0442\u0430\u0442\u043e\u0447\u043d\u043e \u0447\u0430\u0441\u0442\u043e \u0430\u043b\u0433\u043e\u0440\u0438\u0442\u043c\u044b \u0434\u043e\u043b\u0436\u043d\u044b \u0440\u0430\u0431\u043e\u0442\u0430\u0442\u044c \u043b\u043e\u043a\u0430\u043b\u044c\u043d\u043e) \u0432\u0441\u0442\u0430\u0435\u0442 \u0434\u043e\u0432\u043e\u043b\u044c\u043d\u043e \u043e\u0441\u0442\u0440\u043e. \u041d\u0430 \u043f\u0440\u0430\u043a\u0442\u0438\u043a\u0435 \u0432\u044b\u0431\u043e\u0440 \u0441\u0432\u043e\u0434\u0438\u0442\u0441\u044f \u043a \u0442\u0440\u0451\u043c \u043a\u043b\u0430\u0441\u0441\u0430\u043c \u0443\u0441\u0442\u0440\u043e\u0439\u0441\u0442\u0432: <strong>NVIDIA Jetson<\/strong>, \u0432\u043d\u0435\u0448\u043d\u0438\u0439 \u0443\u0441\u043a\u043e\u0440\u0438\u0442\u0435\u043b\u044c (\u043e\u0434\u0438\u043d \u0438\u0437 \u0441\u0430\u043c\u044b\u0445 \u043f\u043e\u043f\u0443\u043b\u044f\u0440\u043d\u044b\u0445 \u2014 Hailo) \u0438 \u043a\u0438\u0442\u0430\u0439\u0441\u043a\u0438\u0439 (\u043d\u0435 \u0432\u0441\u0435\u0433\u0434\u0430, \u043a\u043e\u043d\u0435\u0447\u043d\u043e, \u043d\u043e \u0432 \u0441\u043e\u0432\u0440\u0435\u043c\u0435\u043d\u043d\u044b\u0445 \u0440\u0435\u0430\u043b\u0438\u044f\u0445 \u043e\u0431\u044b\u0447\u043d\u043e \u043a\u0438\u0442\u0430\u0439\u0441\u043a\u0438\u0439) SoC \u0441 \u0438\u043d\u0442\u0435\u0433\u0440\u0438\u0440\u043e\u0432\u0430\u043d\u043d\u044b\u043c NPU. \u0412 \u044d\u0442\u043e\u0439 \u0441\u0442\u0430\u0442\u044c\u0435 \u044f \u0440\u0430\u0441\u0441\u043c\u043e\u0442\u0440\u044e \u043f\u0440\u0435\u0434\u0441\u0442\u0430\u0432\u0438\u0442\u0435\u043b\u044f \u0442\u0440\u0435\u0442\u044c\u0435\u0433\u043e \u043a\u043b\u0430\u0441\u0441\u0430 \u2014 <strong>Axera AX650N<\/strong>, \u0430 <strong>NVIDIA Jetson<\/strong> \u0431\u0443\u0434\u0435\u0442 \u0438\u0441\u043f\u043e\u043b\u044c\u0437\u043e\u0432\u0430\u0442\u044c\u0441\u044f \u0434\u043b\u044f \u0441\u0440\u0430\u0432\u043d\u0435\u043d\u0438\u044f, \u0442\u0430\u043a \u043a\u0430\u043a \u044d\u0442\u043e \u0435\u0434\u0438\u043d\u0441\u0442\u0432\u0435\u043d\u043d\u043e\u0435 \u043c\u0430\u0441\u0441\u043e\u0432\u043e\u0435 edge-\u0440\u0435\u0448\u0435\u043d\u0438\u0435 \u0441 \u0443\u043d\u0438\u0432\u0435\u0440\u0441\u0430\u043b\u044c\u043d\u044b\u043c\u0438 \u0432\u044b\u0447\u0438\u0441\u043b\u0438\u0442\u0435\u043b\u044c\u043d\u044b\u043c\u0438 \u044f\u0434\u0440\u0430\u043c\u0438 (CUDA).<\/p>\n<p>\u042d\u0442\u043e \u043f\u0435\u0440\u0432\u0430\u044f \u0447\u0430\u0441\u0442\u044c \u0446\u0438\u043a\u043b\u0430. \u0417\u0434\u0435\u0441\u044c \u044f \u0440\u0430\u0437\u0431\u0435\u0440\u0443 \u0430\u043f\u043f\u0430\u0440\u0430\u0442\u043d\u0443\u044e \u0430\u0440\u0445\u0438\u0442\u0435\u043a\u0442\u0443\u0440\u0443 \u0441\u0430\u043c\u043e\u0433\u043e <strong>AX650N<\/strong> \u2014 CPU, NPU, DSP, ISP, \u043f\u0430\u043c\u044f\u0442\u044c \u2014 \u0438 \u043f\u043e\u0434\u0435\u043b\u044e\u0441\u044c \u0440\u0435\u0437\u0443\u043b\u044c\u0442\u0430\u0442\u0430\u043c\u0438 \u043f\u0435\u0440\u0432\u044b\u0445 \u0442\u0435\u0441\u0442\u043e\u0432: YOLO, Depth Anything, SuperPoint \u0438 \u043c\u0443\u043b\u044c\u0442\u0438\u043c\u043e\u0434\u0430\u043b\u044c\u043d\u044b\u0439 Qwen3. \u041f\u043e\u0434\u0440\u043e\u0431\u043d\u044b\u0435 \u0431\u0435\u043d\u0447\u043c\u0430\u0440\u043a\u0438 \u0438 \u0441\u0440\u0430\u0432\u043d\u0435\u043d\u0438\u044f \u2014 \u0432\u043e \u0432\u0442\u043e\u0440\u043e\u0439 \u0447\u0430\u0441\u0442\u0438.<\/p>\n<p>\u042f \u0442\u0435\u0441\u0442\u0438\u0440\u043e\u0432\u0430\u043b <strong>AX650N<\/strong> \u0432 \u0440\u0430\u043c\u043a\u0430\u0445 \u0433\u043e\u0442\u043e\u0432\u043e\u0433\u043e \u0443\u0441\u0442\u0440\u043e\u0439\u0441\u0442\u0432\u0430 \u043e\u0442 <a href=\"https:\/\/wiki.sipeed.com\/hardware\/en\/maixIV\/m4chat\/intro.html\" rel=\"noopener noreferrer nofollow\"><strong>Sipeed \u2014 Maix4 Hat<\/strong><\/a>. \u041e\u043d \u0441\u043e\u0441\u0442\u043e\u0438\u0442 \u0438\u0437 \u0434\u0432\u0443\u0445 \u0447\u0430\u0441\u0442\u0435\u0439: <strong>SoM<\/strong>, \u043d\u0430 \u043a\u043e\u0442\u043e\u0440\u043e\u043c \u0440\u0430\u0441\u043f\u043e\u043b\u043e\u0436\u0435\u043d\u044b <strong>SoC<\/strong> \u0438 <strong>8 GB RAM<\/strong> <em>(2&#215;4 GB, \u0442\u0430\u043a \u043a\u0430\u043a \u0443 <\/em><strong><em>AX650N<\/em><\/strong><em> \u0434\u0432\u0430 \u043e\u0442\u0434\u0435\u043b\u044c\u043d\u044b\u0445 DDR-\u043a\u043e\u043d\u0442\u0440\u043e\u043b\u043b\u0435\u0440\u0430)<\/em>, \u0438 <strong>baseboard<\/strong> \u043e\u0442 Sipeed \u0441 \u043c\u0438\u043d\u0438\u043c\u0430\u043b\u044c\u043d\u044b\u043c \u043a\u043e\u043b\u0438\u0447\u0435\u0441\u0442\u0432\u043e\u043c \u0438\u043d\u0442\u0435\u0440\u0444\u0435\u0439\u0441\u043e\u0432. \u0421\u043a\u0440\u043e\u043c\u043d\u043e\u0441\u0442\u044c \u0438\u043d\u0442\u0435\u0440\u0444\u0435\u0439\u0441\u043e\u0432 \u043e\u0431\u044a\u044f\u0441\u043d\u044f\u0435\u0442\u0441\u044f \u043f\u0440\u043e\u0441\u0442\u043e: baseboard \u2014 \u044d\u0442\u043e HAT \u0434\u043b\u044f Raspberry Pi 5, \u043f\u043e\u0434\u043a\u043b\u044e\u0447\u0430\u044e\u0449\u0438\u0439\u0441\u044f \u043f\u043e PCIe 2.0. \u0412 \u0442\u0430\u043a\u043e\u0439 \u043a\u043e\u043d\u0444\u0438\u0433\u0443\u0440\u0430\u0446\u0438\u0438 AX650N \u0440\u0430\u0431\u043e\u0442\u0430\u0435\u0442 \u043a\u0430\u043a \u0432\u043d\u0435\u0448\u043d\u0438\u0439 ML-\u0443\u0441\u043a\u043e\u0440\u0438\u0442\u0435\u043b\u044c, \u0430\u043d\u0430\u043b\u043e\u0433\u0438\u0447\u043d\u043e <strong>Hailo<\/strong>. \u0412 \u0440\u0430\u043c\u043a\u0430\u0445 \u044d\u0442\u043e\u0439 \u0438 \u043f\u043e\u0441\u043b\u0435\u0434\u0443\u044e\u0449\u0438\u0445 \u0441\u0442\u0430\u0442\u0435\u0439 \u044f \u0431\u0443\u0434\u0443 \u0438\u0441\u043f\u043e\u043b\u044c\u0437\u043e\u0432\u0430\u0442\u044c Maix4 Hat \u043a\u0430\u043a \u0441\u0430\u043c\u043e\u0441\u0442\u043e\u044f\u0442\u0435\u043b\u044c\u043d\u044b\u0439 \u043c\u0438\u043a\u0440\u043e\u043a\u043e\u043c\u043f\u044c\u044e\u0442\u0435\u0440.<\/p>\n<p>\u0423 Sipeed \u0435\u0441\u0442\u044c \u0438 \u043f\u043e\u043b\u043d\u043e\u0446\u0435\u043d\u043d\u044b\u0439 baseboard \u0434\u043b\u044f <strong>AX650N<\/strong> \u2014 <strong>M4N Dock<\/strong>. \u0412 \u043d\u0451\u043c \u0443\u0436\u0435 \u0432\u044b\u0432\u0435\u0434\u0435\u043d Ethernet, PCIe \u0438 2 \u0447\u0435\u0442\u044b\u0440\u0451\u0445-\u043b\u0435\u0439\u043d\u043e\u0432\u044b\u0445 MIPI CSI \u0434\u043b\u044f \u043a\u0430\u043c\u0435\u0440. \u041d\u043e \u0441\u0442\u043e\u0438\u0442 \u043e\u043d \u043f\u0440\u0438\u043c\u0435\u0440\u043d\u043e \u0432\u0434\u0432\u043e\u0435 \u0434\u043e\u0440\u043e\u0436\u0435 HAT-\u0432\u0435\u0440\u0441\u0438\u0438 (45 \u0442\u044b\u0441\u044f\u0447 \u0440\u0443\u0431\u043b\u0435\u0439). \u0412 \u043f\u043b\u0430\u043d\u0435 \u043f\u0440\u043e\u0438\u0437\u0432\u043e\u0434\u0438\u0442\u0435\u043b\u044c\u043d\u043e\u0441\u0442\u0438 \u0440\u0430\u0437\u043d\u0438\u0446\u044b \u043d\u0435\u0442 \u2014 <strong>SoM \u043e\u0434\u0438\u043d\u0430\u043a\u043e\u0432\u044b\u0439<\/strong>.<\/p>\n<figure class=\"full-width \"><img decoding=\"async\" src=\"https:\/\/habrastorage.org\/r\/w1560\/getpro\/habr\/upload_files\/0e6\/500\/487\/0e65004870fa0892d519f31519b2a283.png\" width=\"949\" height=\"712\" sizes=\"auto, (max-width: 780px) 100vw, 50vw\" srcset=\"https:\/\/habrastorage.org\/r\/w780\/getpro\/habr\/upload_files\/0e6\/500\/487\/0e65004870fa0892d519f31519b2a283.png 780w,&#10;       https:\/\/habrastorage.org\/r\/w1560\/getpro\/habr\/upload_files\/0e6\/500\/487\/0e65004870fa0892d519f31519b2a283.png 781w\" loading=\"lazy\" decode=\"async\"\/><\/figure>\n<p>\u0417\u0430\u044f\u0432\u043b\u0435\u043d\u043d\u044b\u0435 <strong>18 TOPS@INT8<\/strong> \u2014 \u043c\u0430\u0440\u043a\u0435\u0442\u0438\u043d\u0433\u043e\u0432\u043e\u0435 \u0447\u0438\u0441\u043b\u043e, \u043a\u043e\u0442\u043e\u0440\u043e\u0435 \u0432\u043a\u043b\u044e\u0447\u0430\u0435\u0442 \u043f\u0440\u043e\u0438\u0437\u0432\u043e\u0434\u0438\u0442\u0435\u043b\u044c\u043d\u043e\u0441\u0442\u044c \u043d\u0435\u0441\u043a\u043e\u043b\u044c\u043a\u0438\u0445 \u043f\u043e\u0434\u0441\u0438\u0441\u0442\u0435\u043c \u0441\u0440\u0430\u0437\u0443. \u0412 \u0434\u0430\u0442\u0430\u0448\u0438\u0442\u0435 \u0434\u043b\u044f \u0441\u0430\u043c\u043e\u0433\u043e NPU \u0443\u043a\u0430\u0437\u0430\u043d\u043e <strong>10.8 TOPS@INT8<\/strong>. \u0414\u0430\u043b\u0435\u0435 \u0432 \u0441\u0442\u0430\u0442\u044c\u0435 \u0431\u0443\u0434\u0435\u0442 \u0440\u0430\u0437\u043e\u0431\u0440\u0430\u043d\u043e, \u043e\u0442\u043a\u0443\u0434\u0430 \u0431\u0435\u0440\u0451\u0442\u0441\u044f \u0440\u0430\u0437\u043d\u0438\u0446\u0430 \u0438 \u043a\u0443\u0434\u0430 \u0434\u0435\u043b\u0438\u0441\u044c \u0435\u0449\u0451 <strong>7 TOPS<\/strong>. \u0412 \u043b\u044e\u0431\u043e\u043c \u0441\u043b\u0443\u0447\u0430\u0435, \u0434\u0430\u0436\u0435 <strong>10 TOPS<\/strong> \u0432\u043f\u043e\u043b\u043d\u0435 \u0434\u043e\u0441\u0442\u0430\u0442\u043e\u0447\u043d\u043e \u0434\u043b\u044f \u0438\u043d\u0444\u0435\u0440\u0435\u043d\u0441\u0430 \u0441\u0432\u0451\u0440\u0442\u043e\u0447\u043d\u044b\u0445 \u043c\u043e\u0434\u0435\u043b\u0435\u0439 (\u043d\u0430\u043f\u0440\u0438\u043c\u0435\u0440, YOLO), \u043d\u043e \u0432\u043e\u0442 \u043d\u0430\u0441\u043a\u043e\u043b\u044c\u043a\u043e \u044d\u0442\u043e\u0442 \u0447\u0438\u043f \u0441\u043f\u0440\u0430\u0432\u0438\u0442\u0441\u044f \u0441 \u0431\u043e\u043b\u0435\u0435 \u0442\u044f\u0436\u0451\u043b\u044b\u043c\u0438 \u0437\u0430\u0434\u0430\u0447\u0430\u043c\u0438 \u2014 \u0432\u0438\u0437\u0443\u0430\u043b\u044c\u043d\u044b\u043c\u0438 \u0442\u0440\u0430\u043d\u0441\u0444\u043e\u0440\u043c\u0435\u0440\u0430\u043c\u0438 \u0438 LLM\/VLM \u2014 \u043f\u0440\u0435\u0434\u0441\u0442\u043e\u0438\u0442 \u0432\u044b\u044f\u0441\u043d\u0438\u0442\u044c.<\/p>\n<h3>\u0427\u0442\u043e \u0432\u043d\u0443\u0442\u0440\u0438: \u0430\u0440\u0445\u0438\u0442\u0435\u043a\u0442\u0443\u0440\u0430 AX650N<\/h3>\n<p>\u0427\u0430\u0441\u0442\u044c \u0434\u043e\u043a\u0443\u043c\u0435\u043d\u0442\u0430\u0446\u0438\u0438 Sipeed<a href=\"https:\/\/dl.sipeed.com\/shareURL\/MaixIV\/M4N-Dock\" rel=\"noopener noreferrer nofollow\"> \u0432\u044b\u043b\u043e\u0436\u0438\u043b\u0438 \u0443 \u0441\u0435\u0431\u044f<\/a>. \u041d\u0430 \u043e\u0441\u043d\u043e\u0432\u0435 \u0434\u043e\u0441\u0442\u0443\u043f\u043d\u044b\u0445 \u0434\u0430\u0442\u0430\u0448\u0438\u0442\u043e\u0432 \u0438 \u0442\u0435\u0441\u0442\u043e\u0432 \u043d\u0430 \u0440\u0435\u0430\u043b\u044c\u043d\u043e\u043c \u0436\u0435\u043b\u0435\u0437\u0435 \u044f \u0434\u0430\u043b\u0435\u0435 \u043e\u043f\u0438\u0441\u0430\u043b \u043e\u0431\u0449\u0443\u044e \u0430\u0440\u0445\u0438\u0442\u0435\u043a\u0442\u0443\u0440\u0443 \u0438 \u043e\u0441\u043e\u0431\u0435\u043d\u043d\u043e\u0441\u0442\u0438 \u0440\u0430\u0431\u043e\u0442\u044b NPU.<\/p>\n<p><strong>AX650N<\/strong> \u043e\u0431\u044a\u0435\u0434\u0438\u043d\u044f\u0435\u0442 \u0432\u043e\u0441\u0435\u043c\u044c \u044f\u0434\u0435\u0440 <strong>Cortex-A55<\/strong>, \u0441\u043e\u0431\u0441\u0442\u0432\u0435\u043d\u043d\u044b\u0439 <strong>NPU<\/strong> Axera Neutron, <strong>AI-ISP<\/strong> \u0441 \u0430\u043f\u043f\u0430\u0440\u0430\u0442\u043d\u043e\u0439 \u043d\u0435\u0439\u0440\u043e\u0441\u0435\u0442\u0435\u0432\u043e\u0439 \u043e\u0431\u0440\u0430\u0431\u043e\u0442\u043a\u043e\u0439, \u0434\u0432\u043e\u0439\u043d\u043e\u0439 <strong>DSP<\/strong> \u043d\u0430 \u044f\u0434\u0440\u0430\u0445 Tensilica Vision Q7, \u0430\u043f\u043f\u0430\u0440\u0430\u0442\u043d\u044b\u0435 \u0432\u0438\u0434\u0435\u043e\u043a\u043e\u0434\u0435\u043a\u0438 \u0438 \u0446\u0435\u043b\u044b\u0439 \u043d\u0430\u0431\u043e\u0440 \u0434\u043e\u043f\u043e\u043b\u043d\u0438\u0442\u0435\u043b\u044c\u043d\u044b\u0445 \u0441\u043f\u0435\u0446\u0438\u0430\u043b\u0438\u0437\u0438\u0440\u043e\u0432\u0430\u043d\u043d\u044b\u0445 Vision-\u043c\u043e\u0434\u0443\u043b\u0435\u0439. <strong>GPU<\/strong> \u043d\u0435\u0442 \u0432\u043e\u043e\u0431\u0449\u0435 \u2014 \u0438 \u044d\u0442\u043e \u043f\u0440\u0438\u043d\u0446\u0438\u043f\u0438\u0430\u043b\u044c\u043d\u044b\u0439 \u0432\u044b\u0431\u043e\u0440 \u0430\u0440\u0445\u0438\u0442\u0435\u043a\u0442\u0443\u0440\u044b (\u0438\u043b\u0438 \u043f\u0440\u043e\u0441\u0442\u043e \u0443\u0434\u0435\u0448\u0435\u0432\u043b\u0435\u043d\u0438\u0435), \u043a \u043e\u0431\u0441\u0443\u0436\u0434\u0435\u043d\u0438\u044e \u043a\u043e\u0442\u043e\u0440\u043e\u0433\u043e \u044f \u0432\u0435\u0440\u043d\u0443\u0441\u044c \u0432 \u043a\u043e\u043d\u0446\u0435 \u0441\u0442\u0430\u0442\u044c\u0438.<\/p>\n<h3>CPU<\/h3>\n<p>\u0412\u043e\u0441\u0435\u043c\u044c \u044f\u0434\u0435\u0440 <strong>Cortex-A55<\/strong> \u043e\u0440\u0433\u0430\u043d\u0438\u0437\u043e\u0432\u0430\u043d\u044b \u0432 \u0435\u0434\u0438\u043d\u044b\u0439 \u043a\u043b\u0430\u0441\u0442\u0435\u0440 \u0441 \u0442\u0440\u0451\u0445\u0443\u0440\u043e\u0432\u043d\u0435\u0432\u044b\u043c \u043a\u044d\u0448\u0435\u043c:<\/p>\n<ul>\n<li>\n<p><strong>L1<\/strong>: 32 KB I-Cache + 32 KB D-Cache \u043d\u0430 \u043a\u0430\u0436\u0434\u043e\u0435 \u044f\u0434\u0440\u043e<\/p>\n<\/li>\n<li>\n<p><strong>L2<\/strong>: 64 KB \u043d\u0430 \u043a\u0430\u0436\u0434\u043e\u0435 \u044f\u0434\u0440\u043e<\/p>\n<\/li>\n<li>\n<p><strong>L3<\/strong>: \u043e\u0431\u0449\u0438\u0439 512 KB \u043d\u0430 \u0432\u0435\u0441\u044c \u043a\u043b\u0430\u0441\u0442\u0435\u0440<\/p>\n<\/li>\n<li>\n<p><strong>\u0427\u0430\u0441\u0442\u043e\u0442\u0430<\/strong>: \u0434\u043e 1.7 GHz \u043f\u0440\u0438 \u043d\u0430\u043f\u0440\u044f\u0436\u0435\u043d\u0438\u0438 1.0 V; \u0434\u043e 1.1 GHz \u043f\u0440\u0438 0.8 V<\/p>\n<\/li>\n<li>\n<p><strong>SIMD\/FPU<\/strong>: \u043f\u043e\u0434\u0434\u0435\u0440\u0436\u043a\u0430 NEON \u0438 \u0438\u043d\u0442\u0435\u0433\u0440\u0438\u0440\u043e\u0432\u0430\u043d\u043d\u044b\u0439 FPU<\/p>\n<\/li>\n<\/ul>\n<p><strong>Cortex-A55<\/strong> \u2014 \u043d\u0435 \u0441\u0430\u043c\u044b\u0435 \u043c\u043e\u0449\u043d\u044b\u0435 \u044f\u0434\u0440\u0430 \u0434\u043b\u044f 2026 \u0433\u043e\u0434\u0430 <em>(\u043d\u0430\u043f\u0440\u0438\u043c\u0435\u0440, \u0443 RK3588 4xA76 + 4xA55)<\/em>, \u043d\u043e \u0434\u043b\u044f <strong>Vision-\u043f\u0430\u0439\u043f\u043b\u0430\u0439\u043d\u043e\u0432<\/strong> \u044d\u0442\u043e\u0433\u043e \u043e\u0431\u044b\u0447\u043d\u043e \u0434\u043e\u0441\u0442\u0430\u0442\u043e\u0447\u043d\u043e. CPU \u0437\u0434\u0435\u0441\u044c \u0437\u0430\u043d\u0438\u043c\u0430\u0435\u0442\u0441\u044f \u043e\u0440\u043a\u0435\u0441\u0442\u0440\u0430\u0446\u0438\u0435\u0439: \u043d\u0430\u0441\u0442\u0440\u0430\u0438\u0432\u0430\u0435\u0442 DMA, \u043a\u043e\u043d\u0444\u0438\u0433\u0443\u0440\u0438\u0440\u0443\u0435\u0442 ISP \u0438 NPU, \u0437\u0430\u043f\u0443\u0441\u043a\u0430\u0435\u0442 MIPI CSI \u2014 \u0432\u0441\u0435 \u0440\u0435\u0441\u0443\u0440\u0441\u043e\u0451\u043c\u043a\u0438\u0435 \u043e\u043f\u0435\u0440\u0430\u0446\u0438\u0438 \u0432\u044b\u043f\u043e\u043b\u043d\u044f\u044e\u0442\u0441\u044f \u0430\u043f\u043f\u0430\u0440\u0430\u0442\u043d\u043e. \u0415\u0434\u0438\u043d\u0441\u0442\u0432\u0435\u043d\u043d\u043e\u0435 \u043c\u0435\u0441\u0442\u043e, \u0433\u0434\u0435 \u043f\u0440\u043e\u0446\u0435\u0441\u0441\u043e\u0440 \u0440\u0435\u0430\u043b\u044c\u043d\u043e \u043d\u0430\u0433\u0440\u0443\u0436\u0430\u0435\u0442\u0441\u044f \u2014 \u043f\u043e\u0441\u0442\u043f\u0440\u043e\u0446\u0435\u0441\u0441\u0438\u043d\u0433 \u0440\u0435\u0437\u0443\u043b\u044c\u0442\u0430\u0442\u043e\u0432 \u043c\u043e\u0434\u0435\u043b\u0435\u0439 <em>(NMS, confidence filtering)<\/em>. \u041f\u043e\u044d\u0442\u043e\u043c\u0443 \u043d\u0435\u043e\u0431\u0445\u043e\u0434\u0438\u043c\u043e \u043c\u0430\u043a\u0441\u0438\u043c\u0430\u043b\u044c\u043d\u0443\u044e \u0447\u0430\u0441\u0442\u044c \u043f\u043e\u0441\u0442\u043f\u0440\u043e\u0446\u0435\u0441\u0441\u0438\u043d\u0433\u0430 \u043f\u0435\u0440\u0435\u043d\u043e\u0441\u0438\u0442\u044c \u043d\u0430 NPU \u0438\u043b\u0438 DSP, \u0447\u0442\u043e\u0431\u044b CPU \u043d\u0435 \u0441\u0442\u0430\u043b \u0443\u0437\u043a\u0438\u043c \u0433\u043e\u0440\u043b\u044b\u0448\u043a\u043e\u043c \u0432 \u0438\u0442\u043e\u0433\u043e\u0432\u043e\u0439 \u0441\u043a\u043e\u0440\u043e\u0441\u0442\u0438 \u0440\u0430\u0431\u043e\u0442\u044b \u043c\u043e\u0434\u0435\u043b\u0438.<\/p>\n<h3>\u041f\u043e\u0434\u0441\u0438\u0441\u0442\u0435\u043c\u0430 \u043f\u0430\u043c\u044f\u0442\u0438: \u0434\u0432\u043e\u0439\u043d\u043e\u0439 DDR-\u043a\u043e\u043d\u0442\u0440\u043e\u043b\u043b\u0435\u0440<\/h3>\n<p>AX650N \u043e\u0441\u043d\u0430\u0449\u0451\u043d <strong>\u0434\u0432\u0443\u043c\u044f \u043d\u0435\u0437\u0430\u0432\u0438\u0441\u0438\u043c\u044b\u043c\u0438 32-\u0431\u0438\u0442\u043d\u044b\u043c\u0438 \u043a\u043e\u043d\u0442\u0440\u043e\u043b\u043b\u0435\u0440\u0430\u043c\u0438 LPDDR4\/LPDDR4x<\/strong> \u2014 DDR0 \u0438 DDR1. \u041a\u0430\u0436\u0434\u044b\u0439 \u043a\u043e\u043d\u0442\u0440\u043e\u043b\u043b\u0435\u0440 \u0438\u043c\u0435\u0435\u0442 \u0441\u043e\u0431\u0441\u0442\u0432\u0435\u043d\u043d\u0443\u044e \u0444\u0438\u0437\u0438\u0447\u0435\u0441\u043a\u0443\u044e \u0448\u0438\u043d\u0443 \u0434\u0430\u043d\u043d\u044b\u0445 (32 \u0431\u0438\u0442\u0430), \u0441\u043e\u0431\u0441\u0442\u0432\u0435\u043d\u043d\u044b\u0439 \u043d\u0430\u0431\u043e\u0440 \u0441\u0438\u0433\u043d\u0430\u043b\u043e\u0432 CA, CK, CS, DQS, \u0430 \u0438\u0445 \u0430\u0434\u0440\u0435\u0441\u043d\u044b\u0435 \u043f\u0440\u043e\u0441\u0442\u0440\u0430\u043d\u0441\u0442\u0432\u0430 \u0432 \u0441\u0438\u0441\u0442\u0435\u043c\u043d\u043e\u0439 \u043a\u0430\u0440\u0442\u0435 \u0430\u0434\u0440\u0435\u0441\u043e\u0432 \u0440\u0430\u0437\u043d\u0435\u0441\u0435\u043d\u044b: DDR0-\u0441\u0438\u0441\u0442\u0435\u043c\u0430 \u0431\u0430\u0437\u0438\u0440\u0443\u0435\u0442\u0441\u044f \u043f\u043e \u0430\u0434\u0440\u0435\u0441\u0443 0x8000000, DDR1-\u0441\u0438\u0441\u0442\u0435\u043c\u0430 \u2014 \u043f\u043e 0xC000000.<\/p>\n<figure class=\"full-width \"><img decoding=\"async\" src=\"https:\/\/habrastorage.org\/r\/w1560\/getpro\/habr\/upload_files\/b9d\/c9a\/710\/b9dc9a71009cb98feeac0e81882c5328.png\" width=\"720\" height=\"540\" sizes=\"auto, (max-width: 780px) 100vw, 50vw\" srcset=\"https:\/\/habrastorage.org\/r\/w780\/getpro\/habr\/upload_files\/b9d\/c9a\/710\/b9dc9a71009cb98feeac0e81882c5328.png 780w,&#10;       https:\/\/habrastorage.org\/r\/w1560\/getpro\/habr\/upload_files\/b9d\/c9a\/710\/b9dc9a71009cb98feeac0e81882c5328.png 781w\" loading=\"lazy\" decode=\"async\"\/><\/figure>\n<p>\u0414\u0432\u0430 \u043a\u043e\u043d\u0442\u0440\u043e\u043b\u043b\u0435\u0440\u0430 \u043d\u0435 \u043e\u0431\u0440\u0430\u0437\u0443\u044e\u0442 \u0435\u0434\u0438\u043d\u0443\u044e 64-\u0431\u0438\u0442\u043d\u0443\u044e \u0448\u0438\u043d\u0443 (\u043a\u0430\u043a \u0432 \u043a\u043b\u0430\u0441\u0441\u0438\u0447\u0435\u0441\u043a\u043e\u043c \u201cdual-channel ddr\u201d), \u0430 \u0440\u0430\u0431\u043e\u0442\u0430\u044e\u0442 <strong>\u043f\u0430\u0440\u0430\u043b\u043b\u0435\u043b\u044c\u043d\u043e \u0441 \u0440\u0430\u0437\u0434\u0435\u043b\u044c\u043d\u044b\u043c\u0438 \u0430\u0434\u0440\u0435\u0441\u043d\u044b\u043c\u0438 \u043f\u0440\u043e\u0441\u0442\u0440\u0430\u043d\u0441\u0442\u0432\u0430\u043c\u0438<\/strong>. \u042d\u0442\u043e \u043e\u0437\u043d\u0430\u0447\u0430\u0435\u0442, \u0447\u0442\u043e NPU, VDSP, ISP \u0438 CPU \u043c\u043e\u0433\u0443\u0442 \u0438\u043d\u0438\u0446\u0438\u0438\u0440\u043e\u0432\u0430\u0442\u044c \u0442\u0440\u0430\u043d\u0437\u0430\u043a\u0446\u0438\u0438 \u043a \u0440\u0430\u0437\u043d\u044b\u043c \u043a\u043e\u043d\u0442\u0440\u043e\u043b\u043b\u0435\u0440\u0430\u043c \u043e\u0434\u043d\u043e\u0432\u0440\u0435\u043c\u0435\u043d\u043d\u043e, \u043d\u0435 \u0441\u043e\u0437\u0434\u0430\u0432\u0430\u044f \u043a\u043e\u043d\u043a\u0443\u0440\u0435\u043d\u0446\u0438\u0438 \u0437\u0430 \u043e\u0434\u043d\u0443 \u0448\u0438\u043d\u0443. \u041f\u0440\u0438 \u0441\u043a\u043e\u0440\u043e\u0441\u0442\u0438 4266 Mbps \u043a\u0430\u0436\u0434\u044b\u0439 \u043a\u043e\u043d\u0442\u0440\u043e\u043b\u043b\u0435\u0440 \u043e\u0431\u0435\u0441\u043f\u0435\u0447\u0438\u0432\u0430\u0435\u0442 <strong>~17 GB\/s<\/strong>, \u0441\u0443\u043c\u043c\u0430\u0440\u043d\u043e \u2014 \u0434\u043e <strong>~34 GB\/s<\/strong> \u0442\u0435\u043e\u0440\u0435\u0442\u0438\u0447\u0435\u0441\u043a\u043e\u0439 \u043f\u0440\u043e\u043f\u0443\u0441\u043a\u043d\u043e\u0439 \u0441\u043f\u043e\u0441\u043e\u0431\u043d\u043e\u0441\u0442\u0438.<\/p>\n<p>\u041c\u0430\u043a\u0441\u0438\u043c\u0430\u043b\u044c\u043d\u044b\u0439 \u043e\u0431\u044a\u0451\u043c: \u0434\u043e 16 GB \u043d\u0430 \u043a\u0430\u043d\u0430\u043b (\u0434\u043e 32 GB \u0441\u0443\u043c\u043c\u0430\u0440\u043d\u043e). \u041f\u0440\u0438 \u044d\u0442\u043e\u043c \u0447\u0430\u0441\u0442\u044c \u043f\u0430\u043c\u044f\u0442\u0438 \u0440\u0435\u0437\u0435\u0440\u0432\u0438\u0440\u0443\u0435\u0442\u0441\u044f \u043a\u0430\u043a <strong>CMM (Contiguous Media Memory)<\/strong> \u2014 \u043d\u0435\u043f\u0440\u0435\u0440\u044b\u0432\u043d\u044b\u0439 \u0440\u0435\u0433\u0438\u043e\u043d, \u0434\u043e\u0441\u0442\u0443\u043f\u043d\u044b\u0439 NPU, VPU, ISP \u0438 DSP \u043d\u0430\u043f\u0440\u044f\u043c\u0443\u044e, \u043c\u0438\u043d\u0443\u044f MMU \u044f\u0434\u0440\u0430 Linux. \u041e\u0442\u043d\u043e\u0448\u0435\u043d\u0438\u0435 CMM \u043a Linux-\u043f\u0430\u043c\u044f\u0442\u0438 \u043a\u043e\u043d\u0444\u0438\u0433\u0443\u0440\u0438\u0440\u0443\u0435\u0442\u0441\u044f \u043f\u0440\u043e\u0433\u0440\u0430\u043c\u043c\u043d\u043e. \u041f\u043e \u0443\u043c\u043e\u043b\u0447\u0430\u043d\u0438\u044e \u043d\u0430 <strong>Maix4 HAT<\/strong> \u043f\u043e\u0434 \u0441\u0438\u0441\u0442\u0435\u043c\u0443 \u0432\u044b\u0434\u0435\u043b\u044f\u0435\u0442\u0441\u044f <strong>2 GB<\/strong>, \u0430 \u043f\u043e\u0434 CMM \u043e\u0441\u0442\u0430\u0432\u0448\u0438\u0435\u0441\u044f <strong>6 GB<\/strong>.<\/p>\n<h3>NPU \u2014 Axera Neutron<\/h3>\n<p>\u0421\u0440\u0430\u0437\u0443 \u0445\u043e\u0447\u0443 \u043e\u0431\u0440\u0430\u0442\u0438\u0442\u044c \u0432\u043d\u0438\u043c\u0430\u043d\u0438\u0435, \u0447\u0442\u043e \u0431\u043e\u043b\u044c\u0448\u0438\u043d\u0441\u0442\u0432\u043e \u041f\u041e \u0434\u043b\u044f NPU (\u043d\u0438\u0437\u043a\u043e\u0443\u0440\u043e\u0432\u043d\u0435\u0432\u044b\u0435 \u0434\u0440\u0430\u0439\u0432\u0435\u0440\u0430 \u0438 SDK \u0434\u043b\u044f \u043f\u043e\u0434\u0433\u043e\u0442\u043e\u0432\u043a\u0438 \u043c\u043e\u0434\u0435\u043b\u0435\u0439) \u044f\u0432\u043b\u044f\u0435\u0442\u0441\u044f \u043f\u0440\u043e\u043f\u0440\u0438\u0435\u0442\u0430\u0440\u043d\u044b\u043c, \u043f\u043e\u044d\u0442\u043e\u043c\u0443 \u0432\u0441\u0451 \u043e\u043f\u0438\u0441\u0430\u043d\u043d\u043e\u0435 \u0434\u0430\u043b\u044c\u0448\u0435 \u2014 \u043b\u043e\u0433\u0438\u0447\u0435\u0441\u043a\u0438\u0435 \u0432\u044b\u0432\u043e\u0434\u044b \u043d\u0430 \u043e\u0441\u043d\u043e\u0432\u0435 \u0434\u043e\u0441\u0442\u0443\u043f\u043d\u043e\u0439 \u0434\u043e\u043a\u0443\u043c\u0435\u043d\u0442\u0430\u0446\u0438\u0438.<\/p>\n<p>\u0417\u0430\u044f\u0432\u043b\u0435\u043d\u043d\u044b\u0435 <strong>72 TOPS@INT4 \/ 18 TOPS@INT8<\/strong> \u2014 \u044d\u0442\u043e \u043c\u0430\u0440\u043a\u0435\u0442\u0438\u043d\u0433\u043e\u0432\u0430\u044f \u0441\u0443\u043c\u043c\u0430 \u043f\u043e \u043d\u0435\u0441\u043a\u043e\u043b\u044c\u043a\u0438\u043c \u043f\u043e\u0434\u0441\u0438\u0441\u0442\u0435\u043c\u0430\u043c \u0441\u0440\u0430\u0437\u0443. \u0412 \u0434\u0430\u0442\u0430\u0448\u0438\u0442\u0435 \u043e\u0442\u0434\u0435\u043b\u044c\u043d\u043e \u0434\u043b\u044f NPU \u0443\u043a\u0430\u0437\u0430\u043d\u043e: <strong>43.2 TOPS@INT4 \/ 10.8 TOPS@INT8<\/strong>. \u041a\u0430\u043a \u0438\u043c\u0435\u043d\u043d\u043e \u0431\u044b\u043b\u0438 \u043f\u043e\u043b\u0443\u0447\u0435\u043d\u044b <strong>18 TOPS@INT8<\/strong> \u043d\u0435\u043f\u043e\u043d\u044f\u0442\u043d\u043e, \u043d\u043e \u043e\u0447\u0435\u0432\u0438\u0434\u043d\u043e, \u0447\u0442\u043e \u0432 \u043d\u0438\u0445 \u0432\u0445\u043e\u0434\u0438\u0442 \u0441\u0443\u043c\u043c\u0430 NPU, DSP, MAU \u0438 \u0434\u0440\u0443\u0433\u0438\u0445 IP-\u0431\u043b\u043e\u043a\u043e\u0432. \u041f\u0440\u0438 \u044d\u0442\u043e\u043c, \u043d\u0430\u043f\u0440\u0438\u043c\u0435\u0440, \u0434\u0432\u0443\u0445\u044a\u044f\u0434\u0435\u0440\u043d\u044b\u0439 DSP \u043d\u0435 \u0443\u0447\u0430\u0441\u0442\u0432\u0443\u0435\u0442 \u0432 \u0438\u043d\u0444\u0435\u0440\u0435\u043d\u0441\u0435 \u043c\u043e\u0434\u0435\u043b\u0435\u0439. \u042d\u0442\u043e \u0431\u044b\u043b\u043e \u043f\u0440\u043e\u0432\u0435\u0440\u0435\u043d\u043e \u043f\u0440\u0438 \u0437\u0430\u043f\u0443\u0449\u0435\u043d\u043d\u043e\u0439 YOLO \u0447\u0435\u0440\u0435\u0437 \u0434\u0435\u0431\u0430\u0433 \u0438\u043d\u0444\u043e\u0440\u043c\u0430\u0446\u0438\u044e \u043e\u0442 \u0434\u0440\u0430\u0439\u0432\u0435\u0440\u0430 DSP \u0432 <em>sysfs<\/em> (<em>\/proc\/ax_proc\/dspX\/info<\/em>).<\/p>\n<p>NPU \u043f\u043e\u0434\u0434\u0435\u0440\u0436\u0438\u0432\u0430\u0435\u0442 \u0448\u0438\u0440\u043e\u043a\u0438\u0439 \u043d\u0430\u0431\u043e\u0440 \u043a\u0432\u0430\u043d\u0442\u043e\u0432\u0430\u043d\u0438\u044f (\u043f\u0440\u0438\u0447\u0451\u043c \u0438\u043c\u0435\u0435\u0442\u0441\u044f \u043f\u043e\u0434\u0434\u0435\u0440\u0436\u043a\u0430 \u0441\u043c\u0435\u0448\u0430\u043d\u043d\u044b\u0445 \u0442\u0438\u043f\u043e\u0432 \u043c\u0435\u0436\u0434\u0443 \u0441\u043b\u043e\u044f\u043c\u0438): \u0446\u0435\u043b\u043e\u0447\u0438\u0441\u043b\u0435\u043d\u043d\u044b\u0435 INT4\/INT8\/INT16 \u0438 \u0444\u043b\u043e\u0430\u0442\u044b FP16\/FP32. \u0418 \u044d\u0442\u043e \u043d\u0430 \u0441\u0430\u043c\u043e\u043c \u0434\u0435\u043b\u0435 \u043e\u0447\u0435\u043d\u044c \u0432\u0430\u0436\u043d\u043e, \u043f\u043e\u0442\u043e\u043c\u0443 \u0447\u0442\u043e \u0435\u0441\u043b\u0438 \u0442\u0440\u0435\u0431\u0443\u0435\u0442\u0441\u044f \u0434\u0435\u0442\u0435\u043a\u0442\u0438\u0440\u043e\u0432\u0430\u0442\u044c \u043e\u0431\u044a\u0435\u043a\u0442\u044b \u0441 \u0432\u044b\u0441\u043e\u043a\u043e\u0439 \u0442\u043e\u0447\u043d\u043e\u0441\u0442\u044c\u044e \u0433\u0440\u0430\u043d\u0438\u0446, \u0442\u043e \u0432 INT8 \u043a\u0432\u0430\u043d\u0442\u043e\u0432\u0430\u043d\u0438\u0438 YOLO \u043c\u043e\u0434\u0435\u043b\u0438 \u043c\u043e\u0436\u0435\u0442 \u043d\u0435 \u0445\u0432\u0430\u0442\u0430\u0442\u044c.<\/p>\n<p>\u0417\u0430\u044f\u0432\u043b\u044f\u0435\u0442\u0441\u044f, \u0447\u0442\u043e <strong>INT4 \u0432 4 \u0440\u0430\u0437\u0430<\/strong> \u0431\u044b\u0441\u0442\u0440\u0435\u0435, \u0447\u0435\u043c INT8 (10.8 vs 43.2). \u041e\u0431\u044b\u0447\u043d\u043e INT4 \u0434\u0430\u0451\u0442 \u0442\u0435\u043e\u0440\u0435\u0442\u0438\u0447\u0435\u0441\u043a\u0438\u0439 \u043f\u0440\u0438\u0440\u043e\u0441\u0442 \u043d\u0435 \u0431\u043e\u043b\u0435\u0435 \u0447\u0435\u043c \u0432 2 \u0440\u0430\u0437\u0430 \u043f\u043e \u0441\u0440\u0430\u0432\u043d\u0435\u043d\u0438\u044e \u0441 INT8 <a href=\"https:\/\/clehaxze.tw\/gemlog\/2024\/02-14-benchmarking-rk3588-npu-matrix-multiplcation-performance-ep2.gmi\" rel=\"noopener noreferrer nofollow\">(\u043d\u0430\u043f\u0440\u0438\u043c\u0435\u0440, \u0434\u043b\u044f RK3588)<\/a>. \u0418\u0437 \u044d\u0442\u043e\u0433\u043e \u043c\u043e\u0436\u043d\u043e \u0441\u0434\u0435\u043b\u0430\u0442\u044c \u0432\u044b\u0432\u043e\u0434, \u0447\u0442\u043e \u043f\u043e\u0434\u0434\u0435\u0440\u0436\u043a\u0430 INT4 \u0440\u0435\u0430\u043b\u0438\u0437\u043e\u0432\u0430\u043d\u0430 \u0430\u043f\u043f\u0430\u0440\u0430\u0442\u043d\u043e. \u0421\u043a\u043e\u0440\u0435\u0435 \u0432\u0441\u0435\u0433\u043e \u0434\u0432\u0430 INT4 \u0443\u043f\u0430\u043a\u043e\u0432\u044b\u0432\u0430\u044e\u0442\u0441\u044f \u0432 \u043e\u0434\u043d\u0443 INT8 MAC \u0434\u043e\u0440\u043e\u0436\u043a\u0443, \u043f\u0440\u0438 \u044d\u0442\u043e\u043c \u0432 \u043e\u0442\u043b\u0438\u0447\u0438\u0435 \u043e\u0442 \u0442\u043e\u0433\u043e \u0436\u0435 RK3588 \u0441\u0447\u0438\u0442\u0430\u044e\u0442\u0441\u044f \u043f\u0430\u0440\u0430\u043b\u043b\u0435\u043b\u044c\u043d\u043e (\u0434\u0432\u0430 INT4 \u043f\u0435\u0440\u0435\u043c\u043d\u043e\u0436\u0435\u043d\u0438\u044f \u0437\u0430 \u043e\u0434\u0438\u043d \u0442\u0430\u043a\u0442). \u041a\u0432\u0430\u043d\u0442\u043e\u0432\u0430\u043d\u0438\u0435 \u0432 INT4 \u043e\u0431\u044b\u0447\u043d\u043e \u043d\u0435 \u0438\u0441\u043f\u043e\u043b\u044c\u0437\u0443\u0435\u0442\u0441\u044f \u0434\u043b\u044f \u0441\u0432\u0451\u0440\u0442\u043e\u0447\u043d\u044b\u0445 \u0432\u0438\u0437\u0443\u0430\u043b\u044c\u043d\u044b\u0445 \u043c\u043e\u0434\u0435\u043b\u0435\u0439, \u043d\u043e \u0432\u043e\u0442 \u0432\u0435\u0441\u0430 <strong>LLM<\/strong> \u0430\u043a\u0442\u0438\u0432\u043d\u043e \u043a\u043e\u043d\u0432\u0435\u0440\u0442\u0438\u0440\u0443\u044e\u0442 \u0432 <strong>INT4<\/strong>.<\/p>\n<p>\u041f\u043e \u043a\u0430\u0440\u0442\u0435 \u0430\u0434\u0440\u0435\u0441\u043e\u0432 \u043c\u043e\u0436\u043d\u043e \u0447\u0430\u0441\u0442\u0438\u0447\u043d\u043e \u0432\u043e\u0441\u0441\u0442\u0430\u043d\u043e\u0432\u0438\u0442\u044c \u0442\u043e\u043f\u043e\u043b\u043e\u0433\u0438\u044e \u043f\u043e\u0434\u0441\u0438\u0441\u0442\u0435\u043c\u044b NPU:\u00a0<\/p>\n<figure class=\"full-width \"><img decoding=\"async\" src=\"https:\/\/habrastorage.org\/r\/w1560\/getpro\/habr\/upload_files\/fbf\/d8c\/9a2\/fbfd8c9a260b3d8ee4fca3baba42dd2f.png\" width=\"816\" height=\"756\" sizes=\"auto, (max-width: 780px) 100vw, 50vw\" srcset=\"https:\/\/habrastorage.org\/r\/w780\/getpro\/habr\/upload_files\/fbf\/d8c\/9a2\/fbfd8c9a260b3d8ee4fca3baba42dd2f.png 780w,&#10;       https:\/\/habrastorage.org\/r\/w1560\/getpro\/habr\/upload_files\/fbf\/d8c\/9a2\/fbfd8c9a260b3d8ee4fca3baba42dd2f.png 781w\" loading=\"lazy\" decode=\"async\"\/><\/figure>\n<div>\n<div class=\"table\">\n<table>\n<tbody>\n<tr>\n<td>\n<p align=\"left\"><strong>\u0411\u043b\u043e\u043a<\/strong><\/p>\n<\/td>\n<td>\n<p align=\"left\"><strong>\u0410\u0434\u0440\u0435\u0441<\/strong><\/p>\n<\/td>\n<td>\n<p align=\"left\"><strong>\u041a\u043e\u043b-\u0432\u043e<\/strong><\/p>\n<\/td>\n<td>\n<p align=\"left\"><strong>\u041d\u0430\u0437\u043d\u0430\u0447\u0435\u043d\u0438\u0435<\/strong><\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p align=\"left\">OCM<\/p>\n<\/td>\n<td>\n<p align=\"left\">0x14000000\u20130x14AFFFFF<\/p>\n<\/td>\n<td>\n<p align=\"left\">1<\/p>\n<\/td>\n<td>\n<p align=\"left\">On-Chip Memory, ~11.5 MB<\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p align=\"left\">SDMA<\/p>\n<\/td>\n<td>\n<p align=\"left\">0x16000000\u20130x162FFFFF<\/p>\n<\/td>\n<td>\n<p align=\"left\">3<\/p>\n<\/td>\n<td>\n<p align=\"left\">System DMA \u043c\u0435\u0436\u0434\u0443 CMM \u0438 OCM<\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p align=\"left\">EU0\u2013EU12<\/p>\n<\/td>\n<td>\n<p align=\"left\">0x16400000\u20130x164CFFFF<\/p>\n<\/td>\n<td>\n<p align=\"left\">13<\/p>\n<\/td>\n<td>\n<p align=\"left\">\u0418\u0441\u043f\u043e\u043b\u043d\u0438\u0442\u0435\u043b\u044c\u043d\u044b\u0435 \u0431\u043b\u043e\u043a\u0438<\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p align=\"left\">sync<\/p>\n<\/td>\n<td>\n<p align=\"left\">0x164D0000<\/p>\n<\/td>\n<td>\n<p align=\"left\">1<\/p>\n<\/td>\n<td>\n<p align=\"left\">\u0421\u0438\u043d\u0445\u0440\u043e\u043d\u0438\u0437\u0430\u0446\u0438\u044f \u043c\u0435\u0436\u0434\u0443 EU<\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p align=\"left\">fab<\/p>\n<\/td>\n<td>\n<p align=\"left\">0x164E0000<\/p>\n<\/td>\n<td>\n<p align=\"left\">1<\/p>\n<\/td>\n<td>\n<p align=\"left\">\u041c\u0435\u0436\u0441\u043e\u0435\u0434\u0438\u043d\u0438\u0442\u0435\u043b\u044c\u043d\u0430\u044f \u0444\u0430\u0431\u0440\u0438\u043a\u0430<\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p align=\"left\">fbcdc<\/p>\n<\/td>\n<td>\n<p align=\"left\">0x16500000<\/p>\n<\/td>\n<td>\n<p align=\"left\">1<\/p>\n<\/td>\n<td>\n<p align=\"left\">Frame Buffer Compress\/Decompress<\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<\/div>\n<p>\u0422\u0440\u0438\u043d\u0430\u0434\u0446\u0430\u0442\u044c EU \u0440\u0430\u0437\u0434\u0435\u043b\u0435\u043d\u044b \u043d\u0430 <strong>\u0442\u0440\u0438 \u0433\u0440\u0443\u043f\u043f\u044b vNPU<\/strong>. \u0418\u0437\u043d\u0430\u0447\u0430\u043b\u044c\u043d\u043e \u044f \u043f\u0440\u0435\u0434\u043f\u043e\u043b\u043e\u0436\u0438\u043b \u0440\u0430\u0441\u043f\u0440\u0435\u0434\u0435\u043b\u0435\u043d\u0438\u0435:<\/p>\n<ul>\n<li>\n<p><strong>Convolution Unit (3x)<\/strong> \u2014 \u0432\u044b\u043f\u043e\u043b\u043d\u044f\u0435\u0442 \u0432\u0441\u0435 \u0432\u0438\u0434\u044b \u0441\u0432\u0451\u0440\u0442\u043e\u043a: standard, depthwise, grouped, dilated, transposed convolution<\/p>\n<\/li>\n<li>\n<p><strong>Computer Vision Unit (3x)<\/strong> \u2014 \u043d\u043e\u0440\u043c\u0430\u043b\u0438\u0437\u0430\u0446\u0438\u044f \u0438\u0437\u043e\u0431\u0440\u0430\u0436\u0435\u043d\u0438\u0439, \u0440\u0435\u0441\u0430\u0439\u0437, remap\/warp<\/p>\n<\/li>\n<li>\n<p><strong>Tensor Unit (3x)<\/strong> \u2014 \u0444\u0443\u043d\u043a\u0446\u0438\u0438 \u0430\u043a\u0442\u0438\u0432\u0430\u0446\u0438\u0438, pooling, elementwise-\u043e\u043f\u0435\u0440\u0430\u0446\u0438\u0438, reduce<\/p>\n<\/li>\n<\/ul>\n<figure class=\"full-width \"><img decoding=\"async\" src=\"https:\/\/habrastorage.org\/r\/w1560\/getpro\/habr\/upload_files\/8f8\/12f\/fea\/8f812ffea64b4ed3a512b8f9ae987e8d.png\" width=\"1566\" height=\"792\" sizes=\"auto, (max-width: 780px) 100vw, 50vw\" srcset=\"https:\/\/habrastorage.org\/r\/w780\/getpro\/habr\/upload_files\/8f8\/12f\/fea\/8f812ffea64b4ed3a512b8f9ae987e8d.png 780w,&#10;       https:\/\/habrastorage.org\/r\/w1560\/getpro\/habr\/upload_files\/8f8\/12f\/fea\/8f812ffea64b4ed3a512b8f9ae987e8d.png 781w\" loading=\"lazy\" decode=\"async\"\/><\/figure>\n<p>\u0417\u0430\u0442\u0435\u043c \u0432 \u043a\u0430\u0447\u0435\u0441\u0442\u0432\u0435 \u044d\u043a\u0441\u043f\u0435\u0440\u0438\u043c\u0435\u043d\u0442\u0430 \u044f \u044d\u043a\u0441\u043f\u043e\u0440\u0442\u0438\u0440\u043e\u0432\u0430\u043b \u043e\u0434\u043d\u0443 \u0438 \u0442\u0443 \u0436\u0435 YOLO-\u043c\u043e\u0434\u0435\u043b\u044c \u0447\u0435\u0440\u0435\u0437 Pulsar 2 <em>(\u0438\u043d\u0441\u0442\u0440\u0443\u043c\u0435\u043d\u0442 \u043f\u043e\u0434\u0433\u043e\u0442\u043e\u0432\u043a\u0438 \u043c\u043e\u0434\u0435\u043b\u0435\u0439 \u0434\u043b\u044f Axera NPU)<\/em> \u043f\u043e\u0434 \u0442\u0440\u0438 \u0440\u0430\u0437\u043d\u044b\u0445 \u0440\u0435\u0436\u0438\u043c\u0430 \u0440\u0430\u0431\u043e\u0442\u044b. Pulsar 2 \u043c\u043e\u0436\u0435\u0442 \u0433\u0435\u043d\u0435\u0440\u0438\u0440\u043e\u0432\u0430\u0442\u044c \u0442\u0440\u0435\u0439\u0441 \u043c\u043e\u0434\u0435\u043b\u0438, \u0438\u0437 \u043a\u043e\u0442\u043e\u0440\u043e\u0433\u043e \u0432\u0438\u0434\u043d\u043e \u0432 \u043a\u0430\u043a\u043e\u043c \u043f\u043e\u0440\u044f\u0434\u043a\u0435 \u0447\u0442\u043e \u0432\u044b\u043f\u043e\u043b\u043d\u044f\u0435\u0442\u0441\u044f \u0438 \u043d\u0430 \u043a\u0430\u043a\u0438\u0445 EU.<\/p>\n<p>\u0412 \u0440\u0435\u0436\u0438\u043c\u0435 <strong>NPU1<\/strong> (\u043e\u0434\u043d\u043e \u044f\u0434\u0440\u043e \u0438\u0437 \u0442\u0440\u0451\u0445) \u0442\u0440\u0435\u0439\u0441 \u0434\u0435\u043c\u043e\u043d\u0441\u0442\u0440\u0438\u0440\u0443\u0435\u0442 \u043c\u0438\u043d\u0438\u043c\u0430\u043b\u044c\u043d\u044b\u0439 \u043d\u0430\u0431\u043e\u0440 \u0430\u043a\u0442\u0438\u0432\u043d\u044b\u0445 \u0442\u0440\u0435\u043a\u043e\u0432: \u043e\u0434\u0438\u043d teng, \u0434\u0432\u0430 conv, \u043e\u0434\u0438\u043d sdma \u0438 \u043e\u0442\u0434\u0435\u043b\u044c\u043d\u044b\u0439 cv.<\/p>\n<figure class=\"full-width \"><img decoding=\"async\" src=\"https:\/\/habrastorage.org\/r\/w1560\/getpro\/habr\/upload_files\/67c\/c50\/876\/67cc50876d7b36e3093da9f3b33bf37c.png\" width=\"1630\" height=\"316\" sizes=\"auto, (max-width: 780px) 100vw, 50vw\" srcset=\"https:\/\/habrastorage.org\/r\/w780\/getpro\/habr\/upload_files\/67c\/c50\/876\/67cc50876d7b36e3093da9f3b33bf37c.png 780w,&#10;       https:\/\/habrastorage.org\/r\/w1560\/getpro\/habr\/upload_files\/67c\/c50\/876\/67cc50876d7b36e3093da9f3b33bf37c.png 781w\" loading=\"lazy\" decode=\"async\"\/><\/figure>\n<p>\u0412 \u0440\u0435\u0436\u0438\u043c\u0435 <strong>NPU2<\/strong> (Big-Little, \u0434\u0432\u0430 \u0432\u0438\u0440\u0442\u0443\u0430\u043b\u044c\u043d\u044b\u0445 \u044f\u0434\u0440\u0430) \u043a\u043e\u043b\u0438\u0447\u0435\u0441\u0442\u0432\u043e \u0430\u043a\u0442\u0438\u0432\u043d\u044b\u0445 \u0442\u0440\u0435\u043a\u043e\u0432 \u0443\u0434\u0432\u0430\u0438\u0432\u0430\u0435\u0442\u0441\u044f \u0441\u0438\u043c\u043c\u0435\u0442\u0440\u0438\u0447\u043d\u043e: \u043f\u043e\u044f\u0432\u043b\u044f\u044e\u0442\u0441\u044f teng4\/teng5, conv0-3, sdma8\/sdma9. \u041f\u0430\u0442\u0442\u0435\u0440\u043d \u043f\u043e\u0434\u0442\u0432\u0435\u0440\u0436\u0434\u0430\u0435\u0442, \u0447\u0442\u043e \u0432\u0442\u043e\u0440\u043e\u0439 vNPU \u0434\u043e\u0431\u0430\u0432\u043b\u044f\u0435\u0442 \u0430\u043d\u0430\u043b\u043e\u0433\u0438\u0447\u043d\u044b\u0439 \u043a\u043e\u043c\u043f\u043b\u0435\u043a\u0442 EU, \u0430 \u043d\u0435 \u00ab\u0434\u043e\u0441\u0442\u0440\u0430\u0438\u0432\u0430\u0435\u0442\u00bb \u043f\u0443\u043b \u043a\u0430\u043a\u043e\u0433\u043e-\u0442\u043e \u043e\u0434\u043d\u043e\u0433\u043e \u0442\u0438\u043f\u0430.<\/p>\n<figure class=\"full-width \"><img decoding=\"async\" src=\"https:\/\/habrastorage.org\/r\/w1560\/getpro\/habr\/upload_files\/ae0\/f0c\/d76\/ae0f0cd76c27d717307cdbd8f5f6ec8c.png\" width=\"1635\" height=\"392\" sizes=\"auto, (max-width: 780px) 100vw, 50vw\" srcset=\"https:\/\/habrastorage.org\/r\/w780\/getpro\/habr\/upload_files\/ae0\/f0c\/d76\/ae0f0cd76c27d717307cdbd8f5f6ec8c.png 780w,&#10;       https:\/\/habrastorage.org\/r\/w1560\/getpro\/habr\/upload_files\/ae0\/f0c\/d76\/ae0f0cd76c27d717307cdbd8f5f6ec8c.png 781w\" loading=\"lazy\" decode=\"async\"\/><\/figure>\n<p>\u0412 \u0440\u0435\u0436\u0438\u043c\u0435 <strong>NPU3<\/strong> (\u0432\u0441\u0435 \u0442\u0440\u0438 \u044f\u0434\u0440\u0430 \u043e\u0431\u044a\u0435\u0434\u0438\u043d\u0435\u043d\u044b) \u0442\u0440\u0435\u0439\u0441 \u043d\u0430\u0438\u0431\u043e\u043b\u0435\u0435 \u043f\u043e\u043b\u043d\u044b\u0439: teng6\/teng7\/teng8, conv0-5, sdma12-14.\u00a0<\/p>\n<figure class=\"full-width \"><img decoding=\"async\" src=\"https:\/\/habrastorage.org\/r\/w1560\/getpro\/habr\/upload_files\/6a0\/f40\/b28\/6a0f40b28c840d5d70df965d7732a992.png\" width=\"1631\" height=\"512\" sizes=\"auto, (max-width: 780px) 100vw, 50vw\" srcset=\"https:\/\/habrastorage.org\/r\/w780\/getpro\/habr\/upload_files\/6a0\/f40\/b28\/6a0f40b28c840d5d70df965d7732a992.png 780w,&#10;       https:\/\/habrastorage.org\/r\/w1560\/getpro\/habr\/upload_files\/6a0\/f40\/b28\/6a0f40b28c840d5d70df965d7732a992.png 781w\" loading=\"lazy\" decode=\"async\"\/><\/figure>\n<p>\u0418\u0437 \u044d\u0442\u043e\u0433\u043e \u043c\u043e\u0436\u043d\u043e \u0441\u0434\u0435\u043b\u0430\u0442\u044c \u0432\u044b\u0432\u043e\u0434, \u0447\u0442\u043e \u043a\u0430\u0436\u0434\u044b\u0439 <strong>vNPU<\/strong> \u0441\u043e\u0434\u0435\u0440\u0436\u0438\u0442 <strong>2 Convolution EU + 1 Tensor EU + 1 SDMA<\/strong>, \u0430 \u0432\u0441\u0435\u0433\u043e <strong>Conv EU<\/strong> \u2014\u00a0 <strong>6<\/strong>, \u0430 \u043d\u0435 3, \u043a\u0430\u043a \u044f \u043f\u0440\u0435\u0434\u043f\u043e\u043b\u043e\u0436\u0438\u043b \u0438\u0437\u043d\u0430\u0447\u0430\u043b\u044c\u043d\u043e. \u0422\u0430\u043a\u0436\u0435 \u043c\u043e\u0436\u043d\u043e \u0437\u0430\u043c\u0435\u0442\u0438\u0442\u044c, \u0447\u0442\u043e \u0432\u0441\u0435 \u0432\u044b\u0447\u0438\u0441\u043b\u0435\u043d\u0438\u044f \u043c\u043e\u0434\u0435\u043b\u0438 \u0440\u0430\u0437\u0434\u0435\u043b\u044f\u044e\u0442\u0441\u044f \u043c\u0435\u0436\u0434\u0443 NPU EU  \u2014 \u0437\u0430\u0434\u0430\u0447 \u0434\u043b\u044f MAU \u0438\u043b\u0438 DSP \u043d\u0435\u0442. \u042d\u0442\u043e \u0435\u0449\u0451 \u0440\u0430\u0437 \u043f\u043e\u0434\u0442\u0432\u0435\u0440\u0436\u0434\u0430\u0435\u0442, \u0447\u0442\u043e \u043d\u0430 \u0438\u043d\u0444\u0435\u0440\u0435\u043d\u0441 \u0441\u0442\u0430\u043d\u0434\u0430\u0440\u0442\u043d\u044b\u043c\u0438 \u0441\u0440\u0435\u0434\u0441\u0442\u0432\u0430\u043c\u0438 \u0432\u044b\u0434\u0435\u043b\u044f\u0435\u0442\u0441\u044f \u043c\u0430\u043a\u0441\u0438\u043c\u0443\u043c <strong>10.8 TOPS<\/strong>.<\/p>\n<p>\u0415\u0441\u043b\u0438 \u043f\u043e\u0441\u0447\u0438\u0442\u0430\u0442\u044c \u043a\u043e\u043b\u0438\u0447\u0435\u0441\u0442\u0432\u043e <strong>EU<\/strong> \u0434\u043b\u044f <strong>NPU3<\/strong> \u0440\u0435\u0436\u0438\u043c\u0430, \u0442\u043e \u0438\u0445 \u043f\u043e\u043b\u0443\u0447\u0438\u0442\u0441\u044f 12, \u0445\u043e\u0442\u044f \u043f\u043e \u043a\u0430\u0440\u0442\u0435 \u0430\u0434\u0440\u0435\u0441\u043e\u0432 \u0438\u0445 \u0434\u043e\u043b\u0436\u043d\u043e \u0431\u044b\u0442\u044c 13. \u0421\u043d\u0430\u0447\u0430\u043b\u0430 \u044f \u043f\u0440\u0435\u0434\u043f\u043e\u043b\u043e\u0436\u0438\u043b, \u0447\u0442\u043e \u043e\u0434\u0438\u043d \u043e\u0441\u0442\u0430\u0432\u0448\u0438\u0439\u0441\u044f EU \u2014 \u044d\u0442\u043e <strong>MAU<\/strong> <em>(Matrix Operation Unit)<\/em>, \u043e\u0442\u0434\u0435\u043b\u044c\u043d\u044b\u0439 \u0431\u043b\u043e\u043a \u0434\u043b\u044f \u043f\u0435\u0440\u0435\u043c\u043d\u043e\u0436\u0435\u043d\u0438\u044f \u043c\u0430\u0442\u0440\u0438\u0446. \u0412 \u0442\u0440\u0435\u0439\u0441\u0435 \u0435\u0433\u043e \u043d\u0435\u0442, \u043f\u043e\u0442\u043e\u043c\u0443 \u0447\u0442\u043e \u0434\u043b\u044f <strong>YOLO<\/strong> \u043e\u043d \u0432 \u0446\u0435\u043b\u043e\u043c \u0438 \u043d\u0435 \u043d\u0443\u0436\u0435\u043d, \u0432\u0441\u0451 \u043b\u043e\u0436\u0438\u0442\u0441\u044f \u043d\u0430 <strong>Conv<\/strong> \u0438 <strong>Tensor<\/strong> EU. \u0414\u0430\u0442\u0430\u0448\u0438\u0442 \u043a\u043e\u0441\u0432\u0435\u043d\u043d\u043e \u044d\u0442\u043e \u043f\u043e\u0434\u0442\u0432\u0435\u0440\u0436\u0434\u0430\u0435\u0442, \u043f\u043e\u0442\u043e\u043c\u0443 \u0447\u0442\u043e <strong>MAU<\/strong> \u043e\u043f\u0438\u0441\u0430\u043d \u0432 \u0447\u0430\u0441\u0442\u0438, \u043f\u043e\u0441\u0432\u044f\u0449\u0435\u043d\u043d\u043e\u0439 <strong>NPU<\/strong>:<\/p>\n<figure class=\"full-width \"><img decoding=\"async\" src=\"https:\/\/habrastorage.org\/r\/w1560\/getpro\/habr\/upload_files\/786\/c32\/cdc\/786c32cdc0fcab73241572adf4d7c502.jpeg\" width=\"812\" height=\"871\" sizes=\"auto, (max-width: 780px) 100vw, 50vw\" srcset=\"https:\/\/habrastorage.org\/r\/w780\/getpro\/habr\/upload_files\/786\/c32\/cdc\/786c32cdc0fcab73241572adf4d7c502.jpeg 780w,&#10;       https:\/\/habrastorage.org\/r\/w1560\/getpro\/habr\/upload_files\/786\/c32\/cdc\/786c32cdc0fcab73241572adf4d7c502.jpeg 781w\" loading=\"lazy\" decode=\"async\"\/><\/figure>\n<p>\u0422\u043e\u0447\u043d\u043e\u0433\u043e \u043a\u043e\u043b\u0438\u0447\u0435\u0441\u0442\u0432\u0430 <strong>MAC-\u0431\u043b\u043e\u043a\u043e\u0432<\/strong> \u0432 \u043a\u0430\u0436\u0434\u043e\u043c EU \u043d\u0435 \u0443\u043a\u0430\u0437\u0430\u043d\u043e, \u043d\u043e \u0437\u043d\u0430\u044f \u0447\u0442\u043e \u043e\u0431\u0449\u0430\u044f \u043f\u0438\u043a\u043e\u0432\u0430\u044f \u043f\u0440\u043e\u0438\u0437\u0432\u043e\u0434\u0438\u0442\u0435\u043b\u044c\u043d\u043e\u0441\u0442\u044c <strong>10.8 TOPS<\/strong> \u043c\u043e\u0436\u043d\u043e \u043d\u0430\u0439\u0442\u0438 \u043a\u043e\u043b\u0438\u0447\u0435\u0441\u0442\u0432\u043e <strong>MAC<\/strong> \u0432 \u043a\u0430\u0436\u0434\u043e\u043c <strong>EU<\/strong>, \u0435\u0441\u043b\u0438 \u043f\u0440\u0435\u0434\u043f\u043e\u043b\u043e\u0436\u0438\u0442\u044c, \u0447\u0442\u043e \u043e\u043d\u043e \u043e\u0434\u0438\u043d\u0430\u043a\u043e\u0432\u043e \u0434\u043b\u044f \u043a\u0430\u0436\u0434\u043e\u0433\u043e. \u0412 \u0442\u0430\u043a\u043e\u043c \u0441\u043b\u0443\u0447\u0430\u0435 \u043f\u043e\u043b\u0443\u0447\u0430\u0435\u0442\u0441\u044f:<\/p>\n<p><img decoding=\"async\" class=\"formula\" source=\"\\text{MAC \u043d\u0430 EU} = \\frac{\\text{TOPS} \\times 10^{12}}{2 \\times \\text{EU\\_COUNT} \\times \\text{FREQ\\_HZ} \\times 10^6} = \\frac{10.8 \\times 10^{12}}{2 \\times 13 \\times 800 \\times 10^6} \\approx 519\" alt=\"\\text{MAC \u043d\u0430 EU} = \\frac{\\text{TOPS} \\times 10^{12}}{2 \\times \\text{EU\\_COUNT} \\times \\text{FREQ\\_HZ} \\times 10^6} = \\frac{10.8 \\times 10^{12}}{2 \\times 13 \\times 800 \\times 10^6} \\approx 519\" src=\"https:\/\/habrastorage.org\/getpro\/habr\/formulas\/2\/2b\/2b5\/2b59bdda43425c68d23dc0954b83f2a5.svg\" width=\"648\" height=\"40\" data-width=\"81.381\" data-height=\"5.785\" data-vertical-align=\"-2.327\" sizes=\"auto, (max-width: 780px) 100vw, 50vw\" srcset=\"https:\/\/habrastorage.org\/getpro\/habr\/formulas\/2\/2b\/2b5\/2b59bdda43425c68d23dc0954b83f2a5.svg 780w,&#10;       https:\/\/habrastorage.org\/getpro\/habr\/formulas\/2\/2b\/2b5\/2b59bdda43425c68d23dc0954b83f2a5.svg 781w\" loading=\"lazy\" decode=\"async\"\/><\/p>\n<p>\u0422\u043e\u0433\u0434\u0430 \u043f\u0440\u043e\u0438\u0437\u0432\u043e\u0434\u0438\u0442\u0435\u043b\u044c\u043d\u043e\u0441\u0442\u044c \u043a\u0430\u0436\u0434\u043e\u0433\u043e EU \u043d\u0435\u043c\u043d\u043e\u0433\u043e \u0431\u043e\u043b\u044c\u0448\u0435 <strong>0.8 TOPS<\/strong>. \u0412\u043e\u0437\u043c\u043e\u0436\u043d\u043e \u0432 \u043a\u0430\u0436\u0434\u043e\u043c <strong>EU<\/strong> \u0431\u043e\u043b\u0435\u0435 \u043a\u0440\u0443\u0433\u043b\u043e\u0435 \u0437\u043d\u0430\u0447\u0435\u043d\u0438\u0435 <strong>MAC<\/strong>, \u0430 \u0438\u043c\u0435\u043d\u043d\u043e \u043f\u043e <strong>512 MAC<\/strong>, \u043d\u043e \u043e\u043d\u0438 \u043d\u0435\u0440\u0430\u0432\u043d\u043e\u043c\u0435\u0440\u043d\u043e \u0440\u0430\u0441\u043f\u0440\u0435\u0434\u0435\u043b\u0435\u043d\u044b \u0438 \u0432 \u043a\u0430\u043a\u0438\u0445-\u0442\u043e \u043e\u0442\u0434\u0435\u043b\u044c\u043d\u044b\u0445 <strong>EU<\/strong> \u0431\u043e\u043b\u044c\u0448\u0435 <strong>MAC<\/strong>. \u041b\u0438\u0431\u043e \u0447\u0430\u0441\u0442\u043e\u0442\u0430 <strong>NPU<\/strong> \u043d\u0435\u043c\u043d\u043e\u0433\u043e \u0432\u044b\u0448\u0435, \u0447\u0435\u043c <strong>800 MHz<\/strong> <em>(\u043d\u0430\u043f\u0440\u0438\u043c\u0435\u0440, 810 MHz \u0434\u0430\u0451\u0442 \u043f\u0440\u0430\u043a\u0442\u0438\u0447\u0435\u0441\u043a\u0438 10.8 TOPS)<\/em>. \u0417\u043d\u0430\u0447\u0435\u043d\u0438\u0435 \u0432 <strong>800 MHz<\/strong> \u044f \u0432\u0437\u044f\u043b \u0438\u0437 \u0434\u0435\u0431\u0430\u0433\u0430 <em>sysfs NPU \u0434\u0440\u0430\u0439\u0432\u0435\u0440\u0430<\/em> (\u043e\u043d\u0430 \u0432\u0441\u0435\u0433\u0434\u0430 \u0444\u0438\u043a\u0441\u0438\u0440\u043e\u0432\u0430\u043d\u043d\u0430\u044f \u0438 \u0443\u0432\u0435\u043b\u0438\u0447\u0438\u0442\u044c\/\u0443\u043c\u0435\u043d\u044c\u0448\u0438\u0442\u044c \u0435\u0451 \u043d\u0435\u043b\u044c\u0437\u044f):<\/p>\n<p><code>cat \/proc\/ax_proc\/npu\/clk<br \/>800M<\/code><\/p>\n<p>\u041d\u043e \u043f\u043e\u0442\u043e\u043c \u044f \u043d\u0430\u0448\u0451\u043b \u0440\u0435\u0444\u0435\u0440\u0435\u043d\u0441 \u043c\u0430\u043d\u0443\u0430\u043b \u043d\u0430 <strong>IVE <\/strong><em>(Intelligent Video Engine <\/em>\u2014 \u043e\u0442\u0434\u0435\u043b\u044c\u043d\u0430\u044f \u043f\u043e\u0434\u0441\u0438\u0441\u0442\u0435\u043c\u0430, <a href=\"https:\/\/dl.sipeed.com\/shareURL\/MaixIV\/M4N-Dock\/07_Software_Doc\/board\" rel=\"noopener noreferrer nofollow\">\u0434\u043e\u043a\u0443\u043c\u0435\u043d\u0442 42, AX IVE API<\/a><em>)<\/em>, \u0432 \u043a\u043e\u0442\u043e\u0440\u043e\u043c \u0431\u044b\u043b\u0430 \u043e\u043f\u0438\u0441\u0430\u043d\u0430 \u043f\u0440\u043e\u0438\u0437\u0432\u043e\u0434\u0438\u0442\u0435\u043b\u044c\u043d\u043e\u0441\u0442\u044c <strong>MAU<\/strong>:<\/p>\n<figure class=\"full-width \"><img decoding=\"async\" src=\"https:\/\/habrastorage.org\/r\/w1560\/getpro\/habr\/upload_files\/1a7\/005\/2ab\/1a70052abb21c8865441698b5e3793e1.jpeg\" width=\"878\" height=\"186\" sizes=\"auto, (max-width: 780px) 100vw, 50vw\" srcset=\"https:\/\/habrastorage.org\/r\/w780\/getpro\/habr\/upload_files\/1a7\/005\/2ab\/1a70052abb21c8865441698b5e3793e1.jpeg 780w,&#10;       https:\/\/habrastorage.org\/r\/w1560\/getpro\/habr\/upload_files\/1a7\/005\/2ab\/1a70052abb21c8865441698b5e3793e1.jpeg 781w\" loading=\"lazy\" decode=\"async\"\/><\/figure>\n<p><strong>25.6 GOPS<\/strong> \u2014 \u044d\u0442\u043e \u043f\u0440\u0438\u043c\u0435\u0440\u043d\u043e \u0432 30 \u0440\u0430\u0437 \u043c\u0435\u043d\u044c\u0448\u0435, \u043c\u043e\u0449\u043d\u043e\u0441\u0442\u0438 \u043e\u0434\u043d\u043e\u0433\u043e EU (~<strong>800 GOPS<\/strong>). \u041f\u043e\u043b\u0443\u0447\u0430\u0435\u0442\u0441\u044f, <strong>MAU<\/strong> \u2014 \u044d\u0442\u043e \u043d\u0435 \u043f\u043e\u043b\u043d\u043e\u0446\u0435\u043d\u043d\u044b\u0439 <strong>EU<\/strong>, \u0430 \u043e\u0442\u0434\u0435\u043b\u044c\u043d\u044b\u0439 \u043d\u0435\u0431\u043e\u043b\u044c\u0448\u043e\u0439 \u0431\u043b\u043e\u043a \u0434\u043b\u044f \u043c\u0430\u0442\u0440\u0438\u0447\u043d\u044b\u0445 \u043e\u043f\u0435\u0440\u0430\u0446\u0438\u0439. \u042f \u043f\u043e\u043a\u0430 \u043d\u0435 \u043d\u0430\u0448\u0451\u043b \u0441\u043f\u043e\u0441\u043e\u0431\u0430 \u0438\u0437\u043c\u0435\u0440\u0438\u0442\u044c \u043d\u0430\u0433\u0440\u0443\u0437\u043a\u0443 \u043d\u0430 \u043a\u0430\u0436\u0434\u044b\u0439 \u0438\u0437 <strong>EU<\/strong> \u043f\u043e \u043e\u0442\u0434\u0435\u043b\u044c\u043d\u043e\u0441\u0442\u0438 <em>(\u0434\u0440\u0430\u0439\u0432\u0435\u0440 \u0434\u0430\u0451\u0442 \u0442\u043e\u043b\u044c\u043a\u043e \u043e\u0446\u0435\u043d\u043a\u0443 \u0441\u0443\u043c\u043c\u0430\u0440\u043d\u043e\u0439 \u0443\u0442\u0438\u043b\u0438\u0437\u0430\u0446\u0438\u0438 \u0432\u0441\u0435\u0445 EU)<\/em>. \u041f\u043e\u044d\u0442\u043e\u043c\u0443 13-\u0439 EU, \u0441\u043a\u043e\u0440\u0435\u0435 \u0432\u0441\u0435\u0433\u043e, \u043b\u0438\u0431\u043e \u044f\u0432\u043b\u044f\u0435\u0442\u0441\u044f \u0447\u0430\u0441\u0442\u044c\u044e \u043e\u0434\u043d\u043e\u0433\u043e \u0438\u0437 <strong>Conv\/teng-\u0431\u043b\u043e\u043a\u043e\u0432,<\/strong> \u043b\u0438\u0431\u043e Pulsar 2 \u043f\u0440\u043e\u0441\u0442\u043e \u043d\u0435 \u0437\u0430\u0434\u0435\u0439\u0441\u0442\u0432\u043e\u0432\u0430\u043b \u0435\u0433\u043e \u0434\u043b\u044f YOLO.\u00a0<\/p>\n<p>\u0422\u0435\u043f\u0435\u0440\u044c \u043e\u0431\u0441\u0443\u0434\u0438\u043c \u0432\u043e\u0437\u043c\u043e\u0436\u043d\u043e\u0441\u0442\u0438 <strong>EU<\/strong> \u0431\u043b\u043e\u043a\u043e\u0432. \u0412\u043e-\u043f\u0435\u0440\u0432\u044b\u0445 \u044d\u0442\u043e \u043d\u0435 \u0443\u043d\u0438\u0432\u0435\u0440\u0441\u0430\u043b\u044c\u043d\u044b\u0435 \u0432\u044b\u0447\u0438\u0441\u043b\u0438\u0442\u0435\u043b\u044c\u043d\u044b\u0435 \u0431\u043b\u043e\u043a\u0438, \u043f\u043e\u0434\u043e\u0431\u043d\u043e <strong>CUDA<\/strong>. \u042d\u0442\u043e \u0441\u043f\u0435\u0446\u0438\u0444\u0438\u0447\u043d\u044b\u0435 \u0431\u043b\u043e\u043a\u0438, \u0440\u0435\u0430\u043b\u0438\u0437\u0443\u044e\u0449\u0438\u0435 \u043d\u0430\u0431\u043e\u0440 \u043a\u043e\u043d\u043a\u0440\u0435\u0442\u043d\u044b\u0445 \u043e\u043f\u0435\u0440\u0430\u0446\u0438\u0439. \u0421\u043f\u0438\u0441\u043e\u043a \u0438\u0437 \u0434\u0430\u0442\u0430\u0448\u0438\u0442\u0430:<\/p>\n<ul>\n<li>\n<p><strong>Convolution:<\/strong> general NxM, Depthwise\/Group Conv, Dilation, and ConvTranspose.<\/p>\n<\/li>\n<li>\n<p><strong>MatMul<\/strong> with dynamic weights.<\/p>\n<\/li>\n<li>\n<p><strong>Activations:<\/strong> ReLU\/PReLU\/Swish\/Tanh\/Sigmoid.<\/p>\n<\/li>\n<li>\n<p><strong>Pooling:<\/strong> MaxPool\/AvgPool\/ROIAlign.<\/p>\n<\/li>\n<li>\n<p><strong>Elementwise:<\/strong> Add\/Sub\/Mul\/Div\/Compare.<\/p>\n<\/li>\n<li>\n<p><strong>Reduction:<\/strong> Sum\/Mean\/Max\/ArgMax\/Softmax.<\/p>\n<\/li>\n<li>\n<p><strong>Normalization:<\/strong> BatchNorm\/InstanceNorm\/GroupNorm\/LayerNorm.<\/p>\n<\/li>\n<li>\n<p><strong>Tensor reshape:<\/strong> Reshape\/Transpose\/Concat\/Slice\/Reverse\/Pad\/Space2Depth\/Depth2Space.<\/p>\n<\/li>\n<li>\n<p><strong>CV:<\/strong> Resize\/Remap\/Warp.<\/p>\n<\/li>\n<\/ul>\n<p>\u0412\u0441\u0435 \u043c\u043e\u0434\u0435\u043b\u0438 \u043f\u043e\u0434 <strong>NPU<\/strong> \u043a\u043e\u043d\u0432\u0435\u0440\u0442\u0438\u0440\u0443\u044e\u0442\u0441\u044f \u0447\u0435\u0440\u0435\u0437 \u043f\u0440\u0435\u0434\u0432\u0430\u0440\u0438\u0442\u0435\u043b\u044c\u043d\u043e\u0435 \u043f\u0440\u0435\u0434\u0441\u0442\u0430\u0432\u043b\u0435\u043d\u0438\u0435 \u0432 \u0443\u043d\u0438\u0432\u0435\u0440\u0441\u0430\u043b\u044c\u043d\u043e\u043c \u0444\u043e\u0440\u043c\u0430\u0442\u0435 <strong>ONNX<\/strong>. \u0417\u0430\u044f\u0432\u043b\u044f\u0435\u0442\u0441\u044f \u043e\u0447\u0435\u043d\u044c \u0445\u043e\u0440\u043e\u0448\u0430\u044f \u043f\u043e\u0434\u0434\u0435\u0440\u0436\u043a\u0430 <strong>ONNX-\u0432\u0435\u0440\u0441\u0438\u0438 11<\/strong>. \u0422\u0430\u0431\u043b\u0438\u0446\u0430 \u0441 \u0443\u0440\u043e\u0432\u043d\u0435\u043c \u043f\u043e\u0434\u0434\u0435\u0440\u0436\u043a\u0438 \u043e\u0442\u0434\u0435\u043b\u044c\u043d\u044b\u0445 \u043e\u043f\u0435\u0446\u0430\u0438\u0439 \u0435\u0441\u0442\u044c \u0432 <a href=\"https:\/\/pulsar2-docs.readthedocs.io\/en\/latest\/appendix\/op_support_list_ax650.html\" rel=\"noopener noreferrer nofollow\">\u0434\u043e\u043a\u0443\u043c\u0435\u043d\u0442\u0430\u0446\u0438\u0438 Pulsar 2<\/a>.\u00a0<\/p>\n<h4>OCM \u0438 SDMA<\/h4>\n<p><strong>On-Chip Memory<\/strong> \u043e\u0431\u044a\u0451\u043c\u043e\u043c <strong>~11<\/strong> <strong>MB<\/strong> \u2014 \u044d\u0442\u043e \u0431\u0443\u0444\u0435\u0440 \u0434\u043b\u044f \u0432\u0435\u0441\u043e\u0432 \u0438 \u0430\u043a\u0442\u0438\u0432\u0430\u0446\u0438\u0439 \u0442\u0435\u043a\u0443\u0449\u0435\u0433\u043e \u0441\u043b\u043e\u044f \u043c\u043e\u0434\u0435\u043b\u0438 \u0432 \u043c\u043e\u043c\u0435\u043d\u0442 \u0438\u043d\u0444\u0435\u0440\u0435\u043d\u0441\u0430. \u0414\u043b\u044f \u043a\u0430\u0436\u0434\u043e\u0433\u043e \u0441\u043b\u043e\u044f \u043d\u0435\u0439\u0440\u043e\u0441\u0435\u0442\u0438 \u043f\u0430\u0439\u043f\u043b\u0430\u0439\u043d \u0432\u044b\u0433\u043b\u044f\u0434\u0438\u0442 \u0441\u043b\u0435\u0434\u0443\u044e\u0449\u0438\u043c \u043e\u0431\u0440\u0430\u0437\u043e\u043c:<\/p>\n<ol>\n<li>\n<p><strong>SDMA<\/strong> \u043a\u043e\u043f\u0438\u0440\u0443\u0435\u0442 \u0432\u0435\u0441\u0430 \u0441\u043b\u043e\u044f \u0438\u0437 <strong>CMM<\/strong> \u0432 <strong>OCM<\/strong><\/p>\n<\/li>\n<li>\n<p><strong>EU<\/strong> \u0438\u0441\u043f\u043e\u043b\u043d\u044f\u044e\u0442 \u043e\u043f\u0435\u0440\u0430\u0446\u0438\u044e \u043d\u0430 \u0434\u0430\u043d\u043d\u044b\u0445 <em>(\u0441\u0432\u0451\u0440\u0442\u043a\u0443, \u043d\u043e\u0440\u043c\u0430\u043b\u0438\u0437\u0430\u0446\u0438\u044e \u0438 \u0442.\u0434.)<\/em> \u0438\u0437 <strong>OCM<\/strong><\/p>\n<\/li>\n<li>\n<p><strong>SDMA<\/strong> \u043a\u043e\u043f\u0438\u0440\u0443\u0435\u0442 \u0440\u0435\u0437\u0443\u043b\u044c\u0442\u0430\u0442 \u043e\u0431\u0440\u0430\u0442\u043d\u043e \u0432 <strong>CMM<\/strong> \u0434\u043b\u044f \u0441\u043b\u0435\u0434\u0443\u044e\u0449\u0435\u0433\u043e \u0441\u043b\u043e\u044f<\/p>\n<\/li>\n<\/ol>\n<p>\u0422\u0440\u0438 \u043f\u0430\u0440\u0430\u043b\u043b\u0435\u043b\u044c\u043d\u044b\u0445 <strong>SDMA<\/strong> \u043f\u043e\u0437\u0432\u043e\u043b\u044f\u044e\u0442 \u043e\u0440\u0433\u0430\u043d\u0438\u0437\u043e\u0432\u0430\u0442\u044c \u0434\u0432\u043e\u0439\u043d\u0443\u044e \u0431\u0443\u0444\u0435\u0440\u0438\u0437\u0430\u0446\u0438\u044e: \u043f\u043e\u043a\u0430 <strong>EU<\/strong> \u0432\u044b\u043f\u043e\u043b\u043d\u044f\u044e\u0442 \u0442\u0435\u043a\u0443\u0449\u0438\u0439 \u0441\u043b\u043e\u0439, <strong>SDMA<\/strong> \u0443\u0436\u0435 \u0437\u0430\u0433\u0440\u0443\u0436\u0430\u044e\u0442 \u0432\u0435\u0441\u0430 \u0441\u043b\u0435\u0434\u0443\u044e\u0449\u0435\u0433\u043e. \u0414\u043b\u044f \u043d\u0435\u0431\u043e\u043b\u044c\u0448\u0438\u0445 \u043c\u043e\u0434\u0435\u043b\u0435\u0439 <em>(\u0441 \u0432\u0435\u0441\u0430\u043c\u0438, \u0443\u043c\u0435\u0449\u0430\u044e\u0449\u0438\u043c\u0438\u0441\u044f \u0432 OCM)<\/em> SDMA-\u043d\u0430\u043a\u043b\u0430\u0434\u043d\u044b\u0435 \u0440\u0430\u0441\u0445\u043e\u0434\u044b \u043c\u0438\u043d\u0438\u043c\u0430\u043b\u044c\u043d\u044b. \u0414\u043b\u044f \u043a\u0440\u0443\u043f\u043d\u044b\u0445 \u043c\u043e\u0434\u0435\u043b\u0435\u0439 \u0442\u0438\u043f\u0430 <strong>LLM<\/strong> \u0441 \u0432\u0435\u0441\u0430\u043c\u0438 \u0432 \u043d\u0435\u0441\u043a\u043e\u043b\u044c\u043a\u043e \u0433\u0438\u0433\u0430\u0431\u0430\u0439\u0442 \u044d\u0442\u043e\u0442 \u201cbandwidth bottleneck\u201d \u0441\u0442\u0430\u043d\u043e\u0432\u0438\u0442\u0441\u044f \u043e\u0441\u043d\u043e\u0432\u043d\u044b\u043c \u043e\u0433\u0440\u0430\u043d\u0438\u0447\u0438\u0442\u0435\u043b\u0435\u043c \u043f\u0440\u043e\u0438\u0437\u0432\u043e\u0434\u0438\u0442\u0435\u043b\u044c\u043d\u043e\u0441\u0442\u0438.<\/p>\n<p><strong>NPU<\/strong> \u043c\u043e\u0436\u043d\u043e \u0441\u043a\u043e\u043d\u0444\u0438\u0433\u0443\u0440\u0438\u0440\u043e\u0432\u0430\u0442\u044c \u043a\u0430\u043a <strong>\u0435\u0434\u0438\u043d\u044b\u0439 \u043c\u043e\u043d\u043e\u043b\u0438\u0442\u043d\u044b\u0439 NPU<\/strong> (\u0441 \u0434\u043e\u0441\u0442\u0443\u043f\u043e\u043c \u043a\u043e \u0432\u0441\u0435\u043c EU) \u0438\u043b\u0438 \u043a\u0430\u043a <strong>\u0442\u0440\u0438 \u043d\u0435\u0437\u0430\u0432\u0438\u0441\u0438\u043c\u044b\u0445 vNPU<\/strong>. \u0412 <strong>vNPU<\/strong> \u0440\u0435\u0436\u0438\u043c\u0435 \u0435\u0441\u0442\u044c \u0432\u043e\u0437\u043c\u043e\u0436\u043d\u043e\u0441\u0442\u044c \u0437\u0430\u043f\u0443\u0441\u043a\u0430\u0442\u044c \u043c\u043e\u0434\u0435\u043b\u044c \u043d\u0430 \u043e\u0434\u043d\u043e\u043c \u0438\u043b\u0438 \u0434\u0432\u0443\u0445 \u044f\u0434\u0440\u0430\u0445. \u0420\u0435\u0436\u0438\u043c <strong>vNPU<\/strong> \u0438\u0441\u043f\u043e\u043b\u044c\u0437\u0443\u0435\u0442\u0441\u044f, \u0435\u0441\u043b\u0438 \u0435\u0441\u0442\u044c \u043d\u0435\u043e\u0431\u0445\u043e\u0434\u0438\u043c\u043e\u0441\u0442\u044c \u043f\u0430\u0440\u0430\u043b\u043b\u0435\u043b\u044c\u043d\u043e \u0438\u043d\u0444\u0435\u0440\u0435\u043d\u0441\u0438\u0442\u044c \u043d\u0435\u0441\u043a\u043e\u043b\u044c\u043a\u043e \u0440\u0430\u0437\u043b\u0438\u0447\u043d\u044b\u0445 \u043c\u043e\u0434\u0435\u043b\u0435\u0439. \u0412 \u0441\u043b\u0443\u0447\u0430\u0435, \u0435\u0441\u043b\u0438 \u0432\u0430\u043c \u043d\u0443\u0436\u043d\u043e, \u0447\u0442\u043e\u0431\u044b \u043f\u043e\u043b\u043d\u043e\u0441\u0442\u044c\u044e \u043e\u0434\u0438\u043d\u0430\u043a\u043e\u0432\u0430\u044f \u043c\u043e\u0434\u0435\u043b\u044c \u043e\u0431\u0440\u0430\u0431\u0430\u0442\u044b\u0432\u0430\u043b\u0430 \u0438\u0437\u043e\u0431\u0440\u0430\u0436\u0435\u043d\u0438\u044f \u0441 \u0442\u0440\u0451\u0445 \u043a\u0430\u043c\u0435\u0440, <strong>vNPU<\/strong> \u043d\u0435 \u043d\u0443\u0436\u0435\u043d. \u0414\u043b\u044f \u044d\u0442\u043e\u0433\u043e \u043f\u0440\u0430\u0432\u0438\u043b\u044c\u043d\u0435\u0435 \u0431\u0443\u0434\u0435\u0442 \u0438\u0441\u043f\u043e\u043b\u044c\u0437\u043e\u0432\u0430\u0442\u044c <em>batched inference<\/em>. \u041d\u043e \u0435\u0441\u043b\u0438 \u043d\u0443\u0436\u043d\u0430 \u043f\u0430\u0440\u0430\u043b\u043b\u0435\u043b\u044c\u043d\u0430\u044f \u043e\u0431\u0440\u0430\u0431\u043e\u0442\u043a\u0430 \u0440\u0430\u0437\u043b\u0438\u0447\u043d\u044b\u043c\u0438 \u043c\u043e\u0434\u0435\u043b\u044f\u043c\u0438, \u0442\u043e <strong>vNPU<\/strong> \u0438\u043c\u0435\u0435\u0442 \u0441\u043c\u044b\u0441\u043b. \u041d\u0430\u043f\u0440\u0438\u043c\u0435\u0440: \u043e\u0434\u0438\u043d <strong>vNPU<\/strong> \u043c\u043e\u0436\u0435\u0442 \u0438\u0441\u043f\u043e\u043b\u043d\u044f\u0442\u044c <strong>YOLO<\/strong>, \u0432\u0442\u043e\u0440\u043e\u0439 \u2014 \u043c\u043e\u043d\u043e\u043a\u0443\u043b\u044f\u0440\u043d\u0443\u044e \u043e\u0446\u0435\u043d\u043a\u0443 \u0433\u043b\u0443\u0431\u0438\u043d\u044b (<strong>Depth Anything<\/strong>), \u0430 \u0442\u0440\u0435\u0442\u0438\u0439 \u2014 \u0434\u0435\u0442\u0435\u043a\u0442\u043e\u0440 \u043a\u043b\u044e\u0447\u0435\u0432\u044b\u0445 \u0442\u043e\u0447\u0435\u043a (<strong>SuperPoint<\/strong>), \u0438 \u0432\u0441\u0451 \u044d\u0442\u043e \u043f\u0430\u0440\u0430\u043b\u043b\u0435\u043b\u044c\u043d\u043e, \u0431\u0435\u0437 \u043a\u043e\u043d\u043a\u0443\u0440\u0435\u043d\u0446\u0438\u0438 \u0437\u0430 <strong>EU \u0440\u0435\u0441\u0443\u0440\u0441\u044b<\/strong>. \u041f\u0440\u0438 \u044d\u0442\u043e\u043c \u0432 \u0442\u0430\u043a\u043e\u043c \u0440\u0435\u0436\u0438\u043c\u0435 \u043f\u0440\u043e\u0431\u043b\u0435\u043c\u043e\u0439 \u043c\u043e\u0436\u0435\u0442 \u0441\u0442\u0430\u0442\u044c \u043a\u043e\u043d\u043a\u0443\u0440\u0435\u043d\u0446\u0438\u044f \u0437\u0430 <strong>OCM<\/strong> <em>(11 MB)<\/em> \u0438 <strong>DMA<\/strong> <em>(\u0442\u043e\u0447\u043d\u0435\u0435 \u043f\u0440\u043e\u043f\u0443\u0441\u043a\u043d\u0443\u044e \u0441\u043f\u043e\u0441\u043e\u0431\u043d\u043e\u0441\u0442\u044c DDR)<\/em>, \u043a\u043e\u0442\u043e\u0440\u044b\u0435 \u0431\u0443\u0434\u0443\u0442 \u0434\u0435\u043b\u0438\u0442\u044c\u0441\u044f \u043c\u0435\u0436\u0434\u0443 \u0442\u0440\u0435\u043c\u044f \u043c\u043e\u0434\u0435\u043b\u044f\u043c\u0438.<\/p>\n<h3>AI-ISP: Axera Proton (AxeraVision)<\/h3>\n<p>AI ISP \u0440\u0435\u0430\u043b\u0438\u0437\u0443\u0435\u0442\u0441\u044f \u043d\u0430 \u043e\u0441\u043d\u043e\u0432\u0435 \u043a\u043b\u0430\u0441\u0441\u0438\u0447\u0435\u0441\u043a\u043e\u0439 <strong>ISP<\/strong> \u043e\u0431\u0440\u0430\u0431\u043e\u0442\u043a\u0438, \u0441 \u0434\u043e\u0431\u0430\u0432\u043b\u0435\u043d\u0438\u0435\u043c \u043d\u0435\u0439\u0440\u043e\u043d\u043d\u044b\u0445 \u0441\u0435\u0442\u0435\u0439. \u0424\u0438\u0437\u0438\u0447\u0435\u0441\u043a\u0438 <strong>ISP<\/strong> \u043f\u043e\u0434\u0434\u0435\u0440\u0436\u0438\u0432\u0430\u0435\u0442 \u0434\u043e <strong>4 \u0441\u0435\u043d\u0441\u043e\u0440\u043e\u0432 \u043e\u0434\u043d\u043e\u0432\u0440\u0435\u043c\u0435\u043d\u043d\u043e<\/strong>, \u043c\u0430\u043a\u0441\u0438\u043c\u0430\u043b\u044c\u043d\u043e\u0435 \u0440\u0430\u0437\u0440\u0435\u0448\u0435\u043d\u0438\u0435 <strong>8192\u00d74320@30fps<\/strong> \u0432 \u0440\u0435\u0436\u0438\u043c\u0435 2DOL HDR. \u0412\u043e\u0437\u043c\u043e\u0436\u043d\u044b\u0435 \u043a\u043e\u043d\u0444\u0438\u0433\u0443\u0440\u0430\u0446\u0438\u0438: 1\u00d78-lane, 2\u00d74-lane \u0438\u043b\u0438 4\u00d72-lane.<\/p>\n<p>AI-\u043c\u043e\u0434\u0443\u043b\u0438, \u0432\u0441\u0442\u0440\u043e\u0435\u043d\u043d\u044b\u0435 \u0432 ISP-\u043f\u0430\u0439\u043f\u043b\u0430\u0439\u043d:<\/p>\n<div>\n<div class=\"table\">\n<table>\n<tbody>\n<tr>\n<td>\n<p align=\"left\"><strong>\u041c\u043e\u0434\u0443\u043b\u044c<\/strong><\/p>\n<\/td>\n<td>\n<p align=\"left\"><strong>\u041e\u043f\u0438\u0441\u0430\u043d\u0438\u0435<\/strong><\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p align=\"left\"><strong>AI-HDR<\/strong><\/p>\n<\/td>\n<td>\n<p align=\"left\">\u041c\u043d\u043e\u0433\u043e\u043a\u0430\u0434\u0440\u043e\u0432\u043e\u0435 HDR \u0441 \u043d\u0435\u0439\u0440\u043e\u0441\u0435\u0442\u0435\u0432\u043e\u0439 \u043a\u043e\u043c\u043f\u043e\u043d\u043e\u0432\u043a\u043e\u0439 2\/3\/4 \u043a\u0430\u0434\u0440\u043e\u0432<\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p align=\"left\"><strong>AI-3DNR<\/strong><\/p>\n<\/td>\n<td>\n<p align=\"left\">\u041d\u0435\u0439\u0440\u043e\u0441\u0435\u0442\u0435\u0432\u043e\u0435 3D-\u0448\u0443\u043c\u043e\u043f\u043e\u0434\u0430\u0432\u043b\u0435\u043d\u0438\u0435 (\u043f\u0440\u043e\u0441\u0442\u0440\u0430\u043d\u0441\u0442\u0432\u043e + \u0432\u0440\u0435\u043c\u044f)<\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p align=\"left\"><strong>AI-DIS\/EIS<\/strong><\/p>\n<\/td>\n<td>\n<p align=\"left\">\u042d\u043b\u0435\u043a\u0442\u0440\u043e\u043d\u043d\u0430\u044f \u0441\u0442\u0430\u0431\u0438\u043b\u0438\u0437\u0430\u0446\u0438\u044f \u0438\u0437\u043e\u0431\u0440\u0430\u0436\u0435\u043d\u0438\u044f \u043d\u0430 \u043d\u0435\u0439\u0440\u043e\u0441\u0435\u0442\u0438<\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p align=\"left\"><strong>AI-Demosaic<\/strong><\/p>\n<\/td>\n<td>\n<p align=\"left\">AI-\u0434\u0435\u043c\u043e\u0437\u0430\u0438\u043a\u0430<\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p align=\"left\"><strong>AI-Sharpen<\/strong><\/p>\n<\/td>\n<td>\n<p align=\"left\">\u041d\u0435\u0439\u0440\u043e\u0441\u0435\u0442\u0435\u0432\u043e\u0435 \u043f\u043e\u0432\u044b\u0448\u0435\u043d\u0438\u0435 \u0440\u0435\u0437\u043a\u043e\u0441\u0442\u0438<\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p align=\"left\"><strong>AI-Depurple<\/strong><\/p>\n<\/td>\n<td>\n<p align=\"left\">\u0423\u0441\u0442\u0440\u0430\u043d\u0435\u043d\u0438\u0435 \u0445\u0440\u043e\u043c\u0430\u0442\u0438\u0447\u0435\u0441\u043a\u0438\u0445 \u0430\u0431\u0435\u0440\u0440\u0430\u0446\u0438\u0439 \u0447\u0435\u0440\u0435\u0437 AI<\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p align=\"left\"><strong>AI Dual-Light Fusion<\/strong><\/p>\n<\/td>\n<td>\n<p align=\"left\">\u0421\u043b\u0438\u044f\u043d\u0438\u0435 RGB \u0438 IR-\u043a\u0430\u043d\u0430\u043b\u043e\u0432 (real-color night vision)<\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<\/div>\n<p>\u041a\u0440\u043e\u043c\u0435 <strong>AI \u043c\u043e\u0434\u0443\u043b\u0435\u0439<\/strong>, \u043f\u043e\u0434\u0434\u0435\u0440\u0436\u0438\u0432\u0430\u044e\u0442\u0441\u044f \u0438 \u0432\u0441\u0435 \u043a\u043b\u0430\u0441\u0441\u0438\u0447\u0435\u0441\u043a\u0438\u0435 <strong>ISP \u0430\u043b\u0433\u043e\u0440\u0438\u0442\u043c\u044b<\/strong>: 3A (AF\/AWB\/AE), PDAF, YNR\/CNR\/3DNR, DRC, LSC, Gamma, CAC, Fish-eye \u043a\u043e\u0440\u0440\u0435\u043a\u0446\u0438\u044f, dehaze.<\/p>\n<h4>\u041f\u043e\u0434\u0441\u0438\u0441\u0442\u0435\u043c\u0430 DSP: Tensilica Vision Q7<\/h4>\n<p><strong>AX650N<\/strong> \u0441\u043e\u0434\u0435\u0440\u0436\u0438\u0442 <strong>\u0434\u0432\u0430 \u043d\u0435\u0437\u0430\u0432\u0438\u0441\u0438\u043c\u044b\u0445 \u044f\u0434\u0440\u0430 Tensilica Vision Q7 VDSP<\/strong>, \u0440\u0430\u0431\u043e\u0442\u0430\u044e\u0449\u0438\u0445 \u043d\u0430 \u0447\u0430\u0441\u0442\u043e\u0442\u0435 \u0434\u043e <strong>800 MHz<\/strong>. \u042d\u0442\u043e IP \u043e\u0442 Cadence, \u0441\u043f\u0435\u0446\u0438\u0430\u043b\u0438\u0437\u0438\u0440\u043e\u0432\u0430\u043d\u043d\u043e\u0435 \u0438\u043c\u0435\u043d\u043d\u043e \u0434\u043b\u044f \u0437\u0430\u0434\u0430\u0447 \u043a\u043e\u043c\u043f\u044c\u044e\u0442\u0435\u0440\u043d\u043e\u0433\u043e \u0437\u0440\u0435\u043d\u0438\u044f. \u0425\u0430\u0440\u0430\u043a\u0442\u0435\u0440\u0438\u0441\u0442\u0438\u043a\u0438 \u043a\u0430\u0436\u0434\u043e\u0433\u043e \u044f\u0434\u0440\u0430:<\/p>\n<ul>\n<li>\n<p><strong>L1<\/strong>: 32 KB I-Cache + 16 KB D-Cache<\/p>\n<\/li>\n<li>\n<p><strong>TCM<\/strong>: 32 KB ITCM + 512 KB DTCM (Tightly Coupled Memory \u2014 \u0441\u0432\u0435\u0440\u0445\u0431\u044b\u0441\u0442\u0440\u044b\u0439 \u043b\u043e\u043a\u0430\u043b\u044c\u043d\u044b\u0439 \u041e\u0417\u0423, \u0441\u0440\u0430\u0432\u043d\u0438\u043c\u043e \u0441 OCM \u0432 NPU)<\/p>\n<\/li>\n<li>\n<p><strong>MAC<\/strong>: 256 MAC@INT8, 64 MAC@INT16<\/p>\n<\/li>\n<li>\n<p><strong>iDMA<\/strong>: 4 \u043a\u0430\u043d\u0430\u043b\u0430 \u0441 \u043f\u043e\u0434\u0434\u0435\u0440\u0436\u043a\u043e\u0439 3D-\u0442\u0440\u0430\u043d\u0441\u0444\u0435\u0440\u043e\u0432 (\u0434\u043b\u044f \u0442\u0435\u043d\u0437\u043e\u0440\u043d\u044b\u0445 \u043a\u043e\u043f\u0438\u0440\u043e\u0432\u0430\u043d\u0438\u0439)<\/p>\n<\/li>\n<li>\n<p><strong>AXI<\/strong>: 128-bit AMBA AXI main master + 128-bit AXI iDMA master<\/p>\n<\/li>\n<li>\n<p><strong>FBC\/FBDC<\/strong>: \u0438\u043d\u0442\u0435\u0433\u0440\u0430\u0446\u0438\u044f \u0441\u043e \u0441\u0445\u0435\u043c\u0430\u043c\u0438 \u0441\u0436\u0430\u0442\u0438\u044f \u0431\u0443\u0444\u0435\u0440\u043e\u0432 (\u0441\u043d\u0438\u0436\u0435\u043d\u0438\u0435 DDR bandwidth)<\/p>\n<\/li>\n<li>\n<p><strong>Address remap<\/strong>: \u043b\u043e\u0433\u0438\u043a\u0430 remapping \u0434\u043b\u044f \u043e\u0431\u0440\u0430\u0449\u0435\u043d\u0438\u044f \u043a \u043f\u043e\u043b\u043d\u043e\u043c\u0443 64-bit \u0430\u0434\u0440\u0435\u0441\u043d\u043e\u043c\u0443 \u043f\u0440\u043e\u0441\u0442\u0440\u0430\u043d\u0441\u0442\u0432\u0443 DDR \u0447\u0435\u0440\u0435\u0437 32-bit AXI-\u043c\u0430\u0441\u0442\u0435\u0440<\/p>\n<\/li>\n<\/ul>\n<p>\u0414\u0432\u0430 <strong>VDSP<\/strong> \u0440\u0430\u0431\u043e\u0442\u0430\u044e\u0442 <strong>\u043f\u043e\u043b\u043d\u043e\u0441\u0442\u044c\u044e \u043d\u0435\u0437\u0430\u0432\u0438\u0441\u0438\u043c\u043e<\/strong> \u0438 \u043c\u043e\u0433\u0443\u0442 \u0432\u044b\u043f\u043e\u043b\u043d\u044f\u0442\u044c \u0440\u0430\u0437\u043d\u044b\u0435 \u0437\u0430\u0434\u0430\u0447\u0438 \u043f\u0430\u0440\u0430\u043b\u043b\u0435\u043b\u044c\u043d\u043e. iDMA \u043f\u043e\u0434\u0434\u0435\u0440\u0436\u0438\u0432\u0430\u0435\u0442 \u0442\u0440\u0430\u043d\u0441\u0444\u0435\u0440\u044b TCM &lt;-&gt; TCM, TCM &lt;-&gt; DDR\/OCM \u0438 DDR\/OCM &lt;-&gt; DDR\/OCM. \u042d\u0442\u043e \u043e\u0437\u043d\u0430\u0447\u0430\u0435\u0442, \u0447\u0442\u043e <strong>DSP<\/strong> \u043c\u043e\u0436\u0435\u0442 \u043f\u0440\u0435\u0434\u043e\u0431\u0440\u0430\u0431\u043e\u0442\u0430\u0442\u044c \u0434\u0430\u043d\u043d\u044b\u0435 <em>(\u043d\u0430\u043f\u0440\u0438\u043c\u0435\u0440, ORB-\u0434\u0435\u0441\u043a\u0440\u0438\u043f\u0442\u043e\u0440\u044b \u0438\u043b\u0438 \u043e\u0431\u043b\u0430\u043a\u043e \u0442\u043e\u0447\u0435\u043a \u0441 \u043b\u0438\u0434\u0430\u0440\u0430)<\/em> \u0438 \u043e\u0442\u0434\u0430\u0442\u044c \u0440\u0435\u0437\u0443\u043b\u044c\u0442\u0430\u0442 \u0441\u0440\u0430\u0437\u0443 \u0432 \u0431\u0443\u0444\u0435\u0440 NPU \u2014 \u0431\u0435\u0437 \u043b\u0438\u0448\u043d\u0435\u0433\u043e \u043a\u043e\u043f\u0438\u0440\u043e\u0432\u0430\u043d\u0438\u044f \u0438 \u0437\u0430\u0434\u0435\u0439\u0441\u0442\u0432\u043e\u0432\u0430\u043d\u0438\u044f CPU.<\/p>\n<p>\u041d\u0430 \u0441\u0430\u0439\u0442\u0435 Tensilica \u0435\u0441\u0442\u044c <a href=\"https:\/\/tensilicatools.com\/download\/axera-AX650N-downloads\/\" rel=\"noopener noreferrer nofollow\">\u043e\u0442\u0434\u0435\u043b\u044c\u043d\u0430\u044f \u0441\u0442\u0440\u0430\u043d\u0438\u0446\u0430<\/a> \u0434\u043b\u044f <strong>AX650N<\/strong> \u0441\u043e \u0441\u043f\u0438\u0441\u043a\u043e\u043c \u0431\u0438\u0431\u043b\u0438\u043e\u0442\u0435\u043a:<\/p>\n<figure class=\"full-width \"><img decoding=\"async\" src=\"https:\/\/habrastorage.org\/r\/w1560\/getpro\/habr\/upload_files\/828\/56c\/130\/82856c130e96b81ccfe81e899348f1c9.png\" width=\"822\" height=\"1198\" sizes=\"auto, (max-width: 780px) 100vw, 50vw\" srcset=\"https:\/\/habrastorage.org\/r\/w780\/getpro\/habr\/upload_files\/828\/56c\/130\/82856c130e96b81ccfe81e899348f1c9.png 780w,&#10;       https:\/\/habrastorage.org\/r\/w1560\/getpro\/habr\/upload_files\/828\/56c\/130\/82856c130e96b81ccfe81e899348f1c9.png 781w\" loading=\"lazy\" decode=\"async\"\/><\/figure>\n<p>\u041f\u0440\u0430\u0432\u0434\u0430, \u043f\u0440\u043e\u0441\u0442\u043e \u0442\u0430\u043a \u0438\u0445 \u043d\u0435 \u0441\u043a\u0430\u0447\u0430\u0442\u044c, \u043d\u0443\u0436\u043d\u043e \u0437\u0430\u043f\u0440\u043e\u0441\u0438\u0442\u044c \u0434\u043e\u0441\u0442\u0443\u043f. \u041d\u043e \u043e\u043f\u0438\u0441\u0430\u043d\u0438\u0435 \u0431\u0438\u0431\u043b\u0438\u043e\u0442\u0435\u043a \u0432\u044b\u0433\u043b\u044f\u0434\u0438\u0442 \u0437\u0430\u043c\u0430\u043d\u0447\u0438\u0432\u043e: \u0430\u043f\u043f\u0430\u0440\u0430\u0442\u043d\u043e \u0443\u0441\u043a\u043e\u0440\u0435\u043d\u043d\u044b\u0439 OpenCV, SLAM \u0430\u043b\u0433\u043e\u0440\u0438\u0442\u043c\u044b \u0438 CNN. \u0412 \u0434\u0430\u043b\u044c\u043d\u0435\u0439\u0448\u0438\u0445 \u0442\u0435\u0441\u0442\u0430\u0445 \u044f \u043e\u0431\u044f\u0437\u0430\u0442\u0435\u043b\u044c\u043d\u043e \u043f\u0440\u043e\u0432\u0435\u0440\u044e \u043f\u0440\u043e\u0438\u0437\u0432\u043e\u0434\u0438\u0442\u0435\u043b\u044c\u043d\u043e\u0441\u0442\u044c DSP. \u041f\u043e\u0442\u0435\u043d\u0446\u0438\u0430\u043b\u044c\u043d\u043e DSP \u0434\u043e\u043b\u0436\u0435\u043d \u0440\u0435\u0448\u0430\u0442\u044c \u0437\u0430\u0434\u0430\u0447\u0438, \u043a\u043e\u0442\u043e\u0440\u044b\u0435 \u043f\u043b\u043e\u0445\u043e \u043b\u043e\u0436\u0430\u0442\u0441\u044f \u043d\u0430 CNN-\u043e\u0440\u0438\u0435\u043d\u0442\u0438\u0440\u043e\u0432\u0430\u043d\u043d\u044b\u0439 NPU (\u0445\u043e\u0442\u044f \u0438 DSP \u043c\u043e\u0436\u0435\u0442 \u0432\u044b\u043f\u043e\u043b\u043d\u044f\u0442\u044c \u0441\u0432\u0451\u0440\u0442\u043a\u0438, \u043d\u043e \u043d\u0435 \u0442\u0430\u043a\u0438\u0435 \u0431\u043e\u043b\u044c\u0448\u0438\u0435 \u0438 \u043d\u0435 \u0442\u0430\u043a \u0431\u044b\u0441\u0442\u0440\u043e), \u043d\u0430\u043f\u0440\u0438\u043c\u0435\u0440 \u043a\u043b\u0430\u0441\u0441\u0438\u0447\u0435\u0441\u043a\u0438\u0435 CV \u0430\u043b\u0433\u043e\u0440\u0438\u0442\u043c\u044b (Optical Flow, ORB, FAST). \u0422\u0430\u043a\u0436\u0435, \u043d\u0430\u043f\u0440\u0438\u043c\u0435\u0440, XPCL \u043f\u043e\u043c\u043e\u0436\u0435\u0442 \u044d\u0444\u0444\u0435\u043a\u0442\u0438\u0432\u043d\u043e \u043e\u0431\u0440\u0430\u0431\u0430\u0442\u044b\u0432\u0430\u0442\u044c \u0434\u0430\u043d\u043d\u044b\u0435 \u0441 3D \u043b\u0438\u0434\u0430\u0440\u0430.\u00a0<\/p>\n<h3>IVE: Intelligent Video Engine<\/h3>\n<p><strong>IVE <\/strong>(Intelligent Video Engine) \u2014 \u044d\u0442\u043e \u0435\u0434\u0438\u043d\u044b\u0439 API \u043f\u043e\u0432\u0435\u0440\u0445 \u0432\u0441\u0435\u0433\u043e vision-\u0441\u0442\u0435\u043a\u0430 <strong>AX650N<\/strong>. \u041e\u0434\u0438\u043d \u0445\u044d\u043d\u0434\u043b\u0435\u0440 \u0430\u0433\u0440\u0435\u0433\u0438\u0440\u0443\u0435\u0442 9 \u0438\u0441\u043f\u043e\u043b\u043d\u0438\u0442\u0435\u043b\u044c\u043d\u044b\u0445 \u0434\u0432\u0438\u0436\u043a\u043e\u0432: TDP, VPP, VGP, GDC, DSP, NPU, CPU, MAU \u0438 \u0441\u0430\u043c IVE-\u0434\u0432\u0438\u0436\u043e\u043a. CPU \u0441\u0442\u0440\u043e\u0438\u0442 \u0433\u0440\u0430\u0444 \u0437\u0430\u0434\u0430\u0447 \u0447\u0435\u0440\u0435\u0437 \u044d\u0442\u043e\u0442 API, \u043d\u0435 \u0443\u043f\u0440\u0430\u0432\u043b\u044f\u044f \u0430\u043f\u043f\u0430\u0440\u0430\u0442\u043d\u044b\u043c\u0438 \u043e\u0447\u0435\u0440\u0435\u0434\u044f\u043c\u0438 \u043d\u0430\u043f\u0440\u044f\u043c\u0443\u044e.<\/p>\n<figure class=\"full-width \"><img decoding=\"async\" src=\"https:\/\/habrastorage.org\/r\/w1560\/getpro\/habr\/upload_files\/a01\/285\/07a\/a0128507aad2503afe5b6ebeaf2577c8.png\" width=\"860\" height=\"472\" sizes=\"auto, (max-width: 780px) 100vw, 50vw\" srcset=\"https:\/\/habrastorage.org\/r\/w780\/getpro\/habr\/upload_files\/a01\/285\/07a\/a0128507aad2503afe5b6ebeaf2577c8.png 780w,&#10;       https:\/\/habrastorage.org\/r\/w1560\/getpro\/habr\/upload_files\/a01\/285\/07a\/a0128507aad2503afe5b6ebeaf2577c8.png 781w\" loading=\"lazy\" decode=\"async\"\/><\/figure>\n<p>\u042d\u0442\u0430 \u0434\u043e\u043f\u043e\u043b\u043d\u0438\u0442\u0435\u043b\u044c\u043d\u0430\u044f \u0430\u0431\u0441\u0442\u0440\u0430\u043a\u0446\u0438\u044f \u0441 \u043e\u0434\u043d\u043e\u0439 \u0441\u0442\u043e\u0440\u043e\u043d\u044b \u0443\u0434\u043e\u0431\u043d\u0430, \u043d\u043e \u0441 \u0434\u0440\u0443\u0433\u043e\u0439 \u043d\u0435 \u0440\u0435\u0430\u043b\u0438\u0437\u0443\u0435\u0442 \u043f\u043e\u0442\u0435\u043d\u0446\u0438\u0430\u043b \u043a\u0430\u0436\u0434\u043e\u0439 \u043e\u0442\u0434\u0435\u043b\u044c\u043d\u043e\u0439 \u043f\u043e\u0434\u0441\u0438\u0441\u0442\u0435\u043c\u044b \u043f\u043e\u043b\u043d\u043e\u0441\u0442\u044c\u044e. \u0412\u0437\u044f\u0442\u044c \u0442\u043e\u0442 \u0436\u0435 Q7 DSP, \u0447\u0435\u0440\u0435\u0437 <strong>IVE<\/strong> \u043e\u043d \u0438\u0441\u043f\u043e\u043b\u044c\u0437\u0443\u0435\u0442\u0441\u044f \u0434\u043b\u044f \u043d\u0435\u0431\u043e\u043b\u044c\u0448\u043e\u0439 \u0447\u0430\u0441\u0442\u0438 \u0437\u0430\u0434\u0430\u0447, \u043f\u0440\u0438 \u044d\u0442\u043e\u043c SDK \u043e\u0442 Tensilica \u043c\u043e\u0436\u0435\u0442 \u0438\u0441\u043f\u043e\u043b\u044c\u0437\u043e\u0432\u0430\u0442\u044c \u0435\u0433\u043e \u0431\u043e\u043b\u0435\u0435 \u0433\u0438\u0431\u043a\u043e \u0434\u043b\u044f \u0433\u043e\u0440\u0430\u0437\u0434\u043e \u0431\u043e\u043b\u0435\u0435 \u0448\u0438\u0440\u043e\u043a\u043e\u0433\u043e \u0441\u043f\u0435\u043a\u0442\u0440\u0430 \u0437\u0430\u0434\u0430\u0447.\u00a0<\/p>\n<p>\u0421\u0430\u043c <strong>IVE<\/strong>-\u0434\u0432\u0438\u0436\u043e\u043a \u0440\u0435\u0430\u043b\u0438\u0437\u0443\u0435\u0442 \u043a\u043b\u0430\u0441\u0441\u0438\u0447\u0435\u0441\u043a\u0438\u0435 \u0430\u043b\u0433\u043e\u0440\u0438\u0442\u043c\u044b:<\/p>\n<ul>\n<li>\n<p><strong>\u0410\u0440\u0438\u0444\u043c\u0435\u0442\u0438\u043a\u0430<\/strong>: Add (\u0441 \u0432\u0435\u0441\u043e\u0432\u044b\u043c\u0438 \u043a\u043e\u044d\u0444\u0444\u0438\u0446\u0438\u0435\u043d\u0442\u0430\u043c\u0438 Q1.7), Sub (abs\/shift), MSE (Mean Square Error)<\/p>\n<\/li>\n<li>\n<p><strong>\u0411\u0443\u043b\u0435\u0432\u0430 \u043b\u043e\u0433\u0438\u043a\u0430<\/strong>: And, Or, Xor (\u043d\u0430\u0434 U8C1-\u043c\u0430\u0441\u043a\u0430\u043c\u0438)<\/p>\n<\/li>\n<li>\n<p><strong>\u041f\u0440\u043e\u0441\u0442\u0440\u0430\u043d\u0441\u0442\u0432\u0435\u043d\u043d\u044b\u0435 \u0444\u0438\u043b\u044c\u0442\u0440\u044b<\/strong>: Filter 5\u00d75 (\u043f\u0440\u043e\u0438\u0437\u0432\u043e\u043b\u044c\u043d\u043e\u0435 \u044f\u0434\u0440\u043e), Sobel-like 5\u00d75 (\u0433\u0440\u0430\u0434\u0438\u0435\u043d\u0442 + \u0443\u0433\u043e\u043b), Canny (\u0434\u0432\u0443\u0445\u043f\u0440\u043e\u0445\u043e\u0434\u043d\u044b\u0439: CannHysEdge + CannyEdge)<\/p>\n<\/li>\n<li>\n<p><strong>\u041c\u043e\u0440\u0444\u043e\u043b\u043e\u0433\u0438\u044f<\/strong>: Erode 5\u00d75, Dilate 5\u00d75 (\u0441 \u043f\u0440\u043e\u0438\u0437\u0432\u043e\u043b\u044c\u043d\u043e\u0439 \u043c\u0430\u0441\u043a\u043e\u0439)<\/p>\n<\/li>\n<li>\n<p><strong>\u0421\u0442\u0430\u0442\u0438\u0441\u0442\u0438\u043a\u0430<\/strong>: Hist (\u0433\u0438\u0441\u0442\u043e\u0433\u0440\u0430\u043c\u043c\u0430), EqualizeHist, Integ (\u0438\u043d\u0442\u0435\u0433\u0440\u0430\u043b\u044c\u043d\u043e\u0435 \u0438\u0437\u043e\u0431\u0440\u0430\u0436\u0435\u043d\u0438\u0435, \u0432\u044b\u0432\u043e\u0434 U64C1)<\/p>\n<\/li>\n<li>\n<p><strong>\u0424\u043e\u043d\u043e\u0432\u043e\u0435 \u043c\u043e\u0434\u0435\u043b\u0438\u0440\u043e\u0432\u0430\u043d\u0438\u0435<\/strong>: GMM \u0438 GMM2 (Gaussian Mixture Model \u2014 \u0434\u043b\u044f \u0434\u0435\u0442\u0435\u043a\u0442\u043e\u0440\u0430 \u0434\u0432\u0438\u0436\u0435\u043d\u0438\u044f)<\/p>\n<\/li>\n<li>\n<p><strong>\u041f\u043e\u0440\u043e\u0433\u043e\u0432\u044b\u0435 \u043e\u043f\u0435\u0440\u0430\u0446\u0438\u0438<\/strong>: Thresh (8 \u0440\u0435\u0436\u0438\u043c\u043e\u0432: Binary, Binary-Inv, Trunc, ToZero \u0438 \u0434\u0440.)<\/p>\n<\/li>\n<li>\n<p><strong>\u0424\u043e\u0440\u043c\u0430\u0442\/\u0433\u0435\u043e\u043c\u0435\u0442\u0440\u0438\u044f<\/strong>: 16bit to 8bit \u043a\u043e\u043d\u0432\u0435\u0440\u0442\u0430\u0446\u0438\u044f, CropImage, CropResize, CropResize2, CSC (Color Space Conversion), CropResizeForSplitYUV<\/p>\n<\/li>\n<li>\n<p><strong>\u0421\u0432\u044f\u0437\u043d\u044b\u0435 \u043a\u043e\u043c\u043f\u043e\u043d\u0435\u043d\u0442\u044b<\/strong>: CCL (Connected Component Labeling), \u0434\u043e 2048 \u0440\u0435\u0433\u0438\u043e\u043d\u043e\u0432, \u0432\u044b\u0432\u043e\u0434 blob-\u0441\u0442\u0440\u0443\u043a\u0442\u0443\u0440 \u0441 bounding box<\/p>\n<\/li>\n<\/ul>\n<p>\u0412\u0441\u0435 \u043e\u043f\u0435\u0440\u0430\u0446\u0438\u0438 \u0440\u0430\u0431\u043e\u0442\u0430\u044e\u0442 \u043d\u0430\u043f\u0440\u044f\u043c\u0443\u044e \u0441 \u0444\u0438\u0437\u0438\u0447\u0435\u0441\u043a\u0438\u043c\u0438 \u0430\u0434\u0440\u0435\u0441\u0430\u043c\u0438 <strong>CMM<\/strong>. \u041a\u0430\u043a \u0431\u044b\u043b\u043e \u0441\u043a\u0430\u0437\u0430\u043d\u043e \u0432\u044b\u0448\u0435 <strong>IVE<\/strong> \u0443\u043f\u0440\u0430\u0432\u043b\u044f\u0435\u0442 \u0440\u0430\u0437\u043b\u0438\u0447\u043d\u044b\u043c\u0438 \u0434\u0440\u0443\u0433\u0438\u043c\u0438 Vision \u043f\u043e\u0434\u0441\u0438\u0441\u0442\u0435\u043c\u0430\u043c\u0438, \u043f\u043e\u044d\u0442\u043e\u043c\u0443 \u043a\u0440\u0430\u0442\u043a\u043e \u043e\u043f\u0438\u0448\u0443 \u0438\u0445:<\/p>\n<ul>\n<li>\n<p><strong>TDP<\/strong> (Two-Dimensional Processor) \u2014 2D-\u043e\u043f\u0435\u0440\u0430\u0446\u0438\u0438 \u043d\u0430\u0434 \u0431\u0443\u0444\u0435\u0440\u0430\u043c\u0438: \u043f\u043e\u0432\u043e\u0440\u043e\u0442\u044b, \u043e\u0442\u0440\u0430\u0436\u0435\u043d\u0438\u044f\/\u0444\u043b\u0438\u043f\u044b, alpha blending, \u043a\u043b\u0438\u043f\u043f\u0438\u043d\u0433 \u043c\u0430\u0441\u043a\u0438.\u00a0<\/p>\n<\/li>\n<li>\n<p><strong>VPP<\/strong> (Video Post Processor) \u0438 VGP (Video Graphics Processor) &#8212; \u041c\u0430\u0441\u0448\u0442\u0430\u0431\u0438\u0440\u043e\u0432\u0430\u043d\u0438\u0435 \u0438 OSD<\/p>\n<\/li>\n<li>\n<p><strong>GDC<\/strong> (Geometric Distortion Correction) \u2014 \u043a\u043e\u0440\u0440\u0435\u043a\u0446\u0438\u044f \u0434\u0438\u0441\u0442\u043e\u0440\u0441\u0438\u0438 \u043b\u0438\u043d\u0437\u044b: Fisheye dewarping, LGC (Local Gain Compensation), \u043f\u043e\u0434\u0434\u0435\u0440\u0436\u0438\u0432\u0430\u044e\u0442\u0441\u044f \u0441\u0435\u0442\u043a\u0438 \u0434\u043b\u044f \u043a\u0430\u0441\u0442\u043e\u043c\u043d\u044b\u0445 \u0442\u0438\u043f\u043e\u0432 \u0434\u0438\u0441\u0442\u043e\u0440\u0441\u0438\u0438<\/p>\n<\/li>\n<li>\n<p><strong>PyraLite<\/strong> \u2014 \u0430\u043f\u043f\u0430\u0440\u0430\u0442\u043d\u044b\u0439 \u0433\u0435\u043d\u0435\u0440\u0430\u0442\u043e\u0440 <strong>\u0433\u0430\u0443\u0441\u0441\u043e\u0432\u044b\u0445 \u043f\u0438\u0440\u0430\u043c\u0438\u0434<\/strong> \u0434\u043b\u044f \u0432\u044b\u0434\u0435\u043b\u0435\u043d\u0438\u044f \u0444\u0438\u0447. \u041f\u0438\u0440\u0430\u043c\u0438\u0434\u0430\u043b\u044c\u043d\u044b\u0435 \u043f\u0440\u0438\u0437\u043d\u0430\u043a\u0438 \u0438\u0441\u043f\u043e\u043b\u044c\u0437\u0443\u044e\u0442\u0441\u044f \u0432 \u0434\u0435\u0442\u0435\u043a\u0442\u043e\u0440\u0430\u0445 \u0444\u0438\u0447 SIFT, SURF, \u0430 \u0442\u0430\u043a\u0436\u0435 \u0430\u0440\u0445\u0438\u0442\u0435\u043a\u0442\u0443\u0440\u0430 FPN (feature pyramid networks)<\/p>\n<\/li>\n<\/ul>\n<h3>\u0412\u0438\u0434\u0435\u043e\u043a\u043e\u0434\u0435\u043a (VPU)<\/h3>\n<p>\u0418\u043c\u0435\u0435\u0442\u0441\u044f \u043f\u043e\u043b\u043d\u043e\u0446\u0435\u043d\u043d\u044b\u0439 \u0432\u0438\u0434\u0435\u043e\u043a\u043e\u0434\u0435\u043a H.264\/H.265.<\/p>\n<p><strong>VENC:<\/strong><\/p>\n<ul>\n<li>\n<p><strong>H.264 <\/strong>\u0438<strong> H.265<\/strong><\/p>\n<\/li>\n<li>\n<p>\u041c\u0430\u043a\u0441\u0438\u043c\u0443\u043c 7680\u00d74320@30fps<\/p>\n<\/li>\n<li>\n<p>\u041c\u043d\u043e\u0433\u043e\u043f\u043e\u0442\u043e\u0447\u043d\u043e\u0435 \u043a\u043e\u0434\u0438\u0440\u043e\u0432\u0430\u043d\u0438\u0435: 8K@30fps + 1080p@30fps \u043e\u0434\u043d\u043e\u0432\u0440\u0435\u043c\u0435\u043d\u043d\u043e<\/p>\n<\/li>\n<li>\n<p>\u041f\u043e\u0434\u0434\u0435\u0440\u0436\u043a\u0430 JPEG \u0434\u043e 3840\u00d72160@200fps<\/p>\n<\/li>\n<\/ul>\n<p><strong>VDEC:<\/strong><\/p>\n<ul>\n<li>\n<p>\u0414\u043e 32 \u043f\u043e\u0442\u043e\u043a\u043e\u0432 1080p@30fps \u043f\u0430\u0440\u0430\u043b\u043b\u0435\u043b\u044c\u043d\u043e<\/p>\n<\/li>\n<li>\n<p>\u0414\u043e 7680\u00d74320@60fps \u0432 \u043e\u0434\u043d\u043e\u043c \u043f\u043e\u0442\u043e\u043a\u0435<\/p>\n<\/li>\n<\/ul>\n<p>\u041f\u0430\u0440\u0430\u043b\u043b\u0435\u043b\u044c\u043d\u0430\u044f \u043e\u0431\u0440\u0430\u0431\u043e\u0442\u043a\u0430 <strong>32<\/strong> \u043f\u043e\u0442\u043e\u043a\u043e\u0432 \u043c\u043e\u0436\u0435\u0442 \u0431\u044b\u0442\u044c \u043f\u043e\u043b\u0435\u0437\u043d\u0430 \u0432 \u0440\u043e\u0431\u043e\u0442\u043e\u0442\u0435\u0445\u043d\u0438\u043a\u0435, \u043a\u043e\u0433\u0434\u0430 \u043e\u0434\u0438\u043d \u0430\u0433\u0435\u043d\u0442 \u043e\u0431\u0440\u0430\u0431\u0430\u0442\u044b\u0432\u0430\u0435\u0442 \u0432\u0438\u0434\u0435\u043e \u043e\u0442 \u0431\u043e\u043b\u044c\u0448\u043e\u0433\u043e \u043a\u043e\u043b\u0438\u0447\u0435\u0441\u0442\u0432\u0430 \u043e\u0441\u0442\u0430\u043b\u044c\u043d\u044b\u0445 \u0430\u0433\u0435\u043d\u0442\u043e\u0432 \u0432 \u0440\u043e\u0435.<\/p>\n<h3>\u041f\u0440\u0438\u043c\u0435\u0440 vision \u043f\u0430\u0439\u043f\u043b\u0430\u0439\u043d\u0430<\/h3>\n<figure class=\"full-width \"><img decoding=\"async\" src=\"https:\/\/habrastorage.org\/r\/w1560\/getpro\/habr\/upload_files\/584\/dbb\/71f\/584dbb71f5befddfece8fde1234cbd6c.png\" width=\"951\" height=\"808\" sizes=\"auto, (max-width: 780px) 100vw, 50vw\" srcset=\"https:\/\/habrastorage.org\/r\/w780\/getpro\/habr\/upload_files\/584\/dbb\/71f\/584dbb71f5befddfece8fde1234cbd6c.png 780w,&#10;       https:\/\/habrastorage.org\/r\/w1560\/getpro\/habr\/upload_files\/584\/dbb\/71f\/584dbb71f5befddfece8fde1234cbd6c.png 781w\" loading=\"lazy\" decode=\"async\"\/><\/figure>\n<ul>\n<li>\n<p>\u0421\u0435\u043d\u0441\u043e\u0440 \u0432\u044b\u0434\u0430\u0451\u0442 RAW-\u0434\u0430\u043d\u043d\u044b\u0435 \u043f\u043e MIPI<\/p>\n<\/li>\n<li>\n<p>Sensor System (CSI RX \u043a\u043e\u043d\u0442\u0440\u043e\u043b\u043b\u0435\u0440\u044b) \u043f\u0440\u0438\u043d\u0438\u043c\u0430\u0435\u0442 \u043f\u0430\u043a\u0435\u0442\u044b \u0438 \u0434\u0435\u0441\u0435\u0440\u0438\u0430\u043b\u0438\u0437\u0443\u0435\u0442 \u043f\u043e\u0442\u043e\u043a<\/p>\n<\/li>\n<li>\n<p>ISP (IFE -&gt; ITP) \u0432\u044b\u043f\u043e\u043b\u043d\u044f\u0435\u0442 \u043f\u043e\u043b\u043d\u044b\u0439 \u043f\u0430\u0439\u043f\u043b\u0430\u0439\u043d, \u0432\u043a\u043b\u044e\u0447\u0430\u044f AI-3DNR, AI-HDR, AI-Demosaic<\/p>\n<\/li>\n<li>\n<p>ISP \u0432\u044b\u0434\u0430\u0451\u0442 3 YUV-\u0432\u044b\u0445\u043e\u0434\u0430 \u0441 \u0440\u0430\u0437\u043d\u044b\u043c \u0440\u0430\u0437\u0440\u0435\u0448\u0435\u043d\u0438\u0435\u043c; \u043e\u0434\u0438\u043d \u2014 \u043d\u0430 VENC \u0434\u043b\u044f \u0437\u0430\u043f\u0438\u0441\u0438\/\u0441\u0442\u0440\u0438\u043c\u0438\u043d\u0433\u0430 \u0447\u0435\u0440\u0435\u0437 RTSP, \u0434\u0440\u0443\u0433\u043e\u0439 \u2014 \u0432 VPP\/GDC<\/p>\n<\/li>\n<li>\n<p>GDC \u043a\u043e\u0440\u0440\u0435\u043a\u0442\u0438\u0440\u0443\u0435\u0442 \u0434\u0438\u0441\u0442\u043e\u0440\u0441\u0438\u044e \u043b\u0438\u043d\u0437\u044b<\/p>\n<\/li>\n<li>\n<p>VPP \u043c\u0430\u0441\u0448\u0442\u0430\u0431\u0438\u0440\u0443\u0435\u0442 \u0434\u043e \u043d\u0443\u0436\u043d\u043e\u0433\u043e \u0440\u0430\u0437\u043c\u0435\u0440\u0430 \u0432\u0445\u043e\u0434\u043d\u043e\u0433\u043e \u0442\u0435\u043d\u0437\u043e\u0440\u0430 NPU \u0438 \u0437\u0430\u043f\u0438\u0441\u044b\u0432\u0430\u0435\u0442 \u0432 CMM<\/p>\n<\/li>\n<li>\n<p>NPU (\u0447\u0435\u0440\u0435\u0437 SDMA) \u0437\u0430\u0433\u0440\u0443\u0436\u0430\u0435\u0442 \u0432\u0435\u0441\u0430 \u0438\u0437 CMM \u0432 OCM, Convolution EU \u0432\u044b\u043f\u043e\u043b\u043d\u044f\u044e\u0442 \u0441\u0432\u0451\u0440\u0442\u043a\u0438<\/p>\n<\/li>\n<li>\n<p>\u0420\u0435\u0437\u0443\u043b\u044c\u0442\u0430\u0442\u044b \u0437\u0430\u043f\u0438\u0441\u044b\u0432\u0430\u044e\u0442\u0441\u044f \u0432 CMM<\/p>\n<\/li>\n<li>\n<p>CPU (\u0438\u043b\u0438 DSP) \u043e\u043f\u0446\u0438\u043e\u043d\u0430\u043b\u044c\u043d\u043e \u0432\u044b\u043f\u043e\u043b\u043d\u044f\u0435\u0442 \u043f\u043e\u0441\u0442\u043f\u0440\u043e\u0446\u0435\u0441\u0441\u0438\u043d\u0433 \u0438 \u043f\u0440\u0438\u043d\u0438\u043c\u0430\u0435\u0442 \u0440\u0435\u0448\u0435\u043d\u0438\u0435\u00a0\u00a0<\/p>\n<\/li>\n<\/ul>\n<p>\u042d\u0442\u043e \u043e\u0447\u0435\u043d\u044c \u0443\u043f\u0440\u043e\u0449\u0451\u043d\u043d\u044b\u0439 \u043f\u0430\u0439\u043f\u043b\u0430\u0439\u043d, \u043a\u043e\u0442\u043e\u0440\u044b\u0439 \u0434\u0430\u0451\u0442 \u0431\u0430\u0437\u043e\u0432\u043e\u0435 \u043f\u043e\u043d\u0438\u043c\u0430\u043d\u0438\u0435, \u043a\u0430\u043a \u0432\u0441\u0451 \u0444\u0443\u043d\u043a\u0446\u0438\u043e\u043d\u0438\u0440\u0443\u0435\u0442: NPU \u0441\u043f\u0435\u0446\u0438\u0430\u043b\u044c\u043d\u043e \u0443\u043f\u0440\u043e\u0449\u0435\u043d, \u0431\u0435\u0437 \u0443\u0447\u0451\u0442\u0430 <strong>vNPU<\/strong>, \u043f\u0440\u0435\u0434\u043e\u0431\u0440\u0430\u0431\u043e\u0442\u043a\u0430 \u0432\u0445\u043e\u0434\u043d\u043e\u0433\u043e \u043a\u0430\u0434\u0440\u0430 \u043c\u043e\u0436\u0435\u0442 \u043d\u0435 \u043e\u0433\u0440\u0430\u043d\u0438\u0447\u0438\u0432\u0430\u0442\u044c\u0441\u044f VPP+GDC. \u0412\u0430\u0440\u0438\u0430\u0446\u0438\u0439 \u043c\u043d\u043e\u0433\u043e, vision \u043f\u043e\u0434\u0441\u0438\u0441\u0442\u0435\u043c\u0430 <strong>AX650N<\/strong> \u043c\u043e\u0436\u0435\u0442 \u0434\u043e\u0441\u0442\u0430\u0442\u043e\u0447\u043d\u043e \u0433\u0438\u0431\u043a\u043e \u043c\u0430\u0441\u0448\u0442\u0430\u0431\u0438\u0440\u043e\u0432\u0430\u0442\u044c\u0441\u044f.<\/p>\n<h3>\u0422\u0440\u0430\u043d\u0441\u0444\u043e\u0440\u043c\u0435\u0440\u044b: \u043a\u0430\u043a vision \u2014 SoC \u0437\u0430\u043f\u0443\u0441\u043a\u0430\u0435\u0442 LLM<\/h3>\n<p>\u0418\u0437 \u043f\u0440\u0435\u0434\u044b\u0434\u0443\u0449\u0435\u0433\u043e \u043e\u043f\u0438\u0441\u0430\u043d\u0438\u044f \u0430\u043f\u043f\u0430\u0440\u0430\u0442\u043d\u044b\u0445 \u0431\u043b\u043e\u043a\u043e\u0432 \u043c\u043e\u0436\u0435\u0442 \u0441\u043b\u043e\u0436\u0438\u0442\u044c\u0441\u044f \u0441\u043f\u0440\u0430\u0432\u0435\u0434\u043b\u0438\u0432\u043e\u0435 \u043c\u043d\u0435\u043d\u0438\u0435, \u0447\u0442\u043e <strong>AX650N<\/strong> \u2014 \u044d\u0442\u043e Vision SoC \u0441 <strong>CNN<\/strong> \u043e\u0440\u0438\u0435\u043d\u0442\u0438\u0440\u043e\u0432\u0430\u043d\u043d\u044b\u043c <strong>NPU<\/strong>. \u042d\u0442\u043e \u0434\u0435\u0439\u0441\u0442\u0432\u0438\u0442\u0435\u043b\u044c\u043d\u043e \u0442\u0430\u043a, \u0442\u0435\u043c \u043d\u0435 \u043c\u0435\u043d\u0435\u0435 Axera \u0437\u0430\u044f\u0432\u043b\u044f\u0435\u0442 \u043e <a href=\"https:\/\/www.axera-tech.com\/en\/tech\/npu\" rel=\"noopener noreferrer nofollow\"><em>\u201cnative support for Transformer intelligent processing platform\u201d<\/em><\/a> \u0438 \u0432 \u0438\u0445 Model Zoo \u043d\u0430 Hugging Face \u043c\u043e\u0436\u043d\u043e \u043d\u0430\u0439\u0442\u0438 \u043e\u0447\u0435\u043d\u044c \u043c\u043d\u043e\u0433\u043e \u0433\u043e\u0442\u043e\u0432\u044b\u0445 <strong>LLM<\/strong> \u0438 <strong>VLM<\/strong> \u043c\u043e\u0434\u0435\u043b\u0435\u0439.<\/p>\n<p>\u041e\u0441\u043d\u043e\u0432\u043d\u0430\u044f \u0441\u0442\u0440\u0430\u0442\u0435\u0433\u0438\u044f \u0440\u0430\u0437\u0432\u0451\u0440\u0442\u044b\u0432\u0430\u043d\u0438\u044f <strong>LLM<\/strong> \u043d\u0430 <strong>AX650N<\/strong> \u2014 <strong>W8A16<\/strong> <em>(Weights 8-bit, Activations 16-bit)<\/em> \u043a\u0432\u0430\u043d\u0442\u043e\u0432\u0430\u043d\u0438\u0435. \u042d\u0442\u043e \u043e\u0437\u043d\u0430\u0447\u0430\u0435\u0442:<\/p>\n<ul>\n<li>\n<p>\u0412\u0435\u0441\u0430 \u043f\u0440\u0435\u0434\u0441\u0442\u0430\u0432\u043b\u044f\u044e\u0442\u0441\u044f \u0432 <strong>INT8<\/strong><\/p>\n<\/li>\n<li>\n<p>\u0410\u043a\u0442\u0438\u0432\u0430\u0446\u0438\u0438 \u043f\u0440\u0435\u0434\u0441\u0442\u0430\u0432\u043b\u044f\u044e\u0442\u0441\u044f <strong>FP16<\/strong> (\u0441\u043e\u0445\u0440\u0430\u043d\u044f\u0435\u0442 \u0442\u043e\u0447\u043d\u043e\u0441\u0442\u044c \u0434\u0438\u043d\u0430\u043c\u0438\u0447\u0435\u0441\u043a\u043e\u0433\u043e \u0434\u0438\u0430\u043f\u0430\u0437\u043e\u043d\u0430)<\/p>\n<\/li>\n<\/ul>\n<p>\u041f\u043e\u0434\u0434\u0435\u0440\u0436\u043a\u0430 <strong>INT4<\/strong> \u043a\u0432\u0430\u043d\u0442\u043e\u0432\u0430\u043d\u0438\u044f \u043f\u043e\u0437\u0432\u043e\u043b\u044f\u0435\u0442 \u0438\u0441\u043f\u043e\u043b\u044c\u0437\u043e\u0432\u0430\u0442\u044c \u0438 <strong>W4A16<\/strong> \u0440\u0435\u0436\u0438\u043c, \u043d\u043e \u0442\u0430\u043a\u0430\u044f \u0430\u0433\u0440\u0435\u0441\u0441\u0438\u0432\u043d\u043e\u0435 \u043a\u0432\u0430\u043d\u0442\u043e\u0432\u0430\u043d\u0438\u0435 \u0443\u0445\u0443\u0434\u0448\u0430\u0435\u0442 \u043a\u0430\u0447\u0435\u0441\u0442\u0432\u043e \u0440\u0430\u0431\u043e\u0442\u044b \u043c\u043e\u0434\u0435\u043b\u0438, \u043d\u0435\u0441\u043c\u043e\u0442\u0440\u044f \u043d\u0430 \u0437\u043d\u0430\u0447\u0438\u0442\u0435\u043b\u044c\u043d\u043e\u0435 \u0443\u0432\u0435\u043b\u0438\u0447\u0435\u043d\u0438\u0435 \u0441\u043a\u043e\u0440\u043e\u0441\u0442\u0438 \u0433\u0435\u043d\u0435\u0440\u0430\u0446\u0438\u0438 \u043e\u0442\u0432\u0435\u0442\u043e\u0432.\u00a0<\/p>\n<p>\u0421\u0432\u0451\u0440\u0442\u043e\u0447\u043d\u044b\u0435 \u0431\u043b\u043e\u043a\u0438 \u0432\u043d\u0443\u0442\u0440\u0438 <strong>NPU<\/strong> \u0444\u0430\u043a\u0442\u0438\u0447\u0435\u0441\u043a\u0438 \u0432\u044b\u043f\u043e\u043b\u043d\u044f\u044e\u0442 <strong>GEMM<\/strong> \u043e\u043f\u0435\u0440\u0430\u0446\u0438\u0438:<\/p>\n<p>GEMM C = A * B<\/p>\n<p>\u041c\u0430\u0442\u0440\u0438\u0446\u0443 \u0430\u043a\u0442\u0438\u0432\u0430\u0446\u0438\u0439 <strong>A<\/strong> <em>(\u0440\u0430\u0437\u043c\u0435\u0440\u043d\u043e\u0441\u0442\u044c [batch, seq_len, d_model])<\/em> \u043c\u043e\u0436\u043d\u043e \u0440\u0430\u0441\u0441\u043c\u0430\u0442\u0440\u0438\u0432\u0430\u0442\u044c \u043a\u0430\u043a \u0438\u0437\u043e\u0431\u0440\u0430\u0436\u0435\u043d\u0438\u0435 \u0441 \u043f\u0440\u043e\u0441\u0442\u0440\u0430\u043d\u0441\u0442\u0432\u0435\u043d\u043d\u044b\u043c\u0438 \u0440\u0430\u0437\u043c\u0435\u0440\u0430\u043c\u0438 <em>[seq_len, 1]<\/em> \u0438 d_model \u043a\u0430\u043d\u0430\u043b\u0430\u043c\u0438. \u0410 \u043c\u0430\u0442\u0440\u0438\u0446\u0443 \u0432\u0435\u0441\u043e\u0432 <strong>B<\/strong> \u2014 \u043a\u0430\u043a \u044f\u0434\u0440\u0430 \u0441\u0432\u0451\u0440\u0442\u043a\u0438 <em>[1&#215;1 \u0441 d_model \u0432\u0445\u043e\u0434\u043d\u044b\u043c\u0438 \u0438 d_out \u0432\u044b\u0445\u043e\u0434\u043d\u044b\u043c\u0438 \u043a\u0430\u043d\u0430\u043b\u0430\u043c\u0438]<\/em>.\u00a0<\/p>\n<p>\u0422\u0430\u043a\u0438\u043c \u043e\u0431\u0440\u0430\u0437\u043e\u043c <strong>Conv EU<\/strong> \u0432 <strong>NPU<\/strong> \u043c\u043e\u0433\u0443\u0442 \u0432\u044b\u043f\u043e\u043b\u043d\u044f\u0442\u044c \u043b\u0438\u043d\u0435\u0439\u043d\u044b\u0435 \u043f\u0440\u043e\u0435\u043a\u0446\u0438\u0438 \u0442\u0440\u0430\u043d\u0441\u0444\u043e\u0440\u043c\u0435\u0440\u0430, \u043f\u0440\u0435\u0434\u0441\u0442\u0430\u0432\u043b\u044f\u044f \u0438\u0445 \u0432 \u0432\u0438\u0434\u0435 1&#215;1 \u0441\u0432\u0451\u0440\u0442\u043e\u043a. \u041d\u043e <strong>Conv EU<\/strong> \u0430\u0434\u0430\u043f\u0442\u0438\u0440\u043e\u0432\u0430\u043d\u044b \u043f\u043e\u0434 \u0440\u0430\u0431\u043e\u0442\u0443 \u0441 \u043c\u0430\u0442\u0440\u0438\u0446\u0430\u043c\u0438 \u0441\u0442\u0430\u0442\u0438\u0447\u043d\u043e\u0433\u043e \u0440\u0430\u0437\u043c\u0435\u0440\u0430, \u0430 \u0432 \u0441\u043b\u0443\u0447\u0430\u0435 \u0441 <strong>LLM<\/strong>, \u0434\u043b\u0438\u043d\u0430 \u043f\u043e\u0441\u043b\u0435\u0434\u043e\u0432\u0430\u0442\u0435\u043b\u044c\u043d\u043e\u0441\u0442\u0438 \u0442\u043e\u043a\u0435\u043d\u043e\u0432 (seq_len) \u043e\u0431\u044b\u0447\u043d\u043e \u043f\u0435\u0440\u0435\u043c\u0435\u043d\u043d\u0430\u044f, \u043f\u043e\u044d\u0442\u043e\u043c\u0443 Pulsar 2 \u043c\u043e\u0436\u0435\u0442 \u0433\u0435\u043d\u0435\u0440\u0438\u0440\u043e\u0432\u0430\u0442\u044c \u0440\u0430\u0437\u043d\u044b\u0435 \u043c\u043e\u0434\u0435\u043b\u0438 \u0434\u043b\u044f \u0440\u0430\u0437\u043d\u043e\u0439 \u0434\u043b\u0438\u043d\u044b \u043d\u0430\u0447\u0430\u043b\u044c\u043d\u043e\u0433\u043e \u043a\u043e\u043d\u0442\u0435\u043a\u0441\u0442\u0430 <em>(prefill_len)<\/em>.<\/p>\n<p>\u041f\u0440\u0438 \u0438\u043d\u0444\u0435\u0440\u0435\u043d\u0441\u0435 <strong>ax-llm<\/strong> \u0434\u0435\u043b\u0430\u0435\u0442 \u0441\u043b\u0435\u0434\u0443\u044e\u0449\u0435\u0435:<\/p>\n<ul>\n<li>\n<p><strong>Prefill<\/strong>: \u043f\u0440\u043e\u043c\u043f\u0442 \u043d\u0430\u0440\u0435\u0437\u0430\u0435\u0442\u0441\u044f \u0447\u0430\u043d\u043a\u0430\u043c\u0438 \u043f\u043e prefill_len \u0442\u043e\u043a\u0435\u043d\u043e\u0432, \u043a\u0430\u0436\u0434\u044b\u0439 \u0447\u0430\u043d\u043a \u043f\u0440\u043e\u0433\u043e\u043d\u044f\u0435\u0442\u0441\u044f \u0447\u0435\u0440\u0435\u0437 p128_l{N} axmodel&#8217;\u0438. <strong>KV-\u043a\u044d\u0448<\/strong> \u043d\u0430\u043a\u0430\u043f\u043b\u0438\u0432\u0430\u0435\u0442\u0441\u044f \u0432 <strong>CMM<\/strong>.<\/p>\n<\/li>\n<li>\n<p><strong>Decode<\/strong>: \u043a\u0430\u0436\u0434\u044b\u0439 \u043d\u043e\u0432\u044b\u0439 \u0442\u043e\u043a\u0435\u043d \u043f\u0440\u043e\u0433\u043e\u043d\u044f\u0435\u0442\u0441\u044f \u0447\u0435\u0440\u0435\u0437 decode-\u0432\u0435\u0440\u0441\u0438\u044e \u0442\u0435\u0445 \u0436\u0435 \u0441\u043b\u043e\u0451\u0432 (\u0441\u043a\u043e\u043d\u0432\u0435\u0440\u0442\u0438\u0440\u043e\u0432\u0430\u043d\u043d\u0443\u044e \u043f\u043e\u0434 seq_len=1). <strong>KV-\u043a\u044d\u0448<\/strong> \u0447\u0438\u0442\u0430\u0435\u0442\u0441\u044f \u0438\u0437 <strong>CMM<\/strong>.<\/p>\n<\/li>\n<\/ul>\n<p>\u0422\u0430\u043a\u043e\u0439 \u043f\u043e\u0434\u0445\u043e\u0434 \u043e\u043f\u0440\u0430\u0432\u0434\u0430\u043d \u0434\u043b\u044f \u0442\u0435\u043a\u0443\u0449\u0435\u0439 \u0430\u0440\u0445\u0438\u0442\u0435\u043a\u0442\u0443\u0440\u044b \u0443\u0441\u043a\u043e\u0440\u0438\u0442\u0435\u043b\u044f, \u043d\u043e \u0433\u043b\u0430\u0432\u043d\u0430\u044f \u043f\u0440\u043e\u0431\u043b\u0435\u043c\u0430 \u0437\u0430\u043a\u043b\u044e\u0447\u0430\u0435\u0442\u0441\u044f \u0432 \u0442\u043e\u043c, \u0447\u0442\u043e \u0440\u0435\u0430\u043b\u0438\u0437\u0430\u0446\u0438\u044f \u043f\u0430\u043c\u044f\u0442\u0438 <em>(CMM &lt;-&gt; OCM)<\/em> \u043d\u0435 \u043e\u043f\u0442\u0438\u043c\u0438\u0437\u0438\u0440\u043e\u0432\u0430\u043d\u0430 \u043f\u043e\u0434 \u0442\u0430\u043a\u0443\u044e \u0432\u044b\u0441\u043e\u043a\u0443\u044e \u043d\u0430\u0433\u0440\u0443\u0437\u043a\u0443. \u0411\u043e\u043b\u0435\u0435 \u0434\u0435\u0442\u0430\u043b\u044c\u043d\u044b\u0439 \u0430\u043d\u0430\u043b\u0438\u0437 \u0431\u0443\u0434\u0435\u0442 \u0434\u0430\u043b\u0435\u0435.<\/p>\n<p>\u0422\u043e\u043a\u0435\u043d\u0438\u0437\u0430\u0442\u043e\u0440 \u0438\u0441\u043f\u043e\u043b\u043d\u044f\u0435\u0442\u0441\u044f \u043d\u0430 <strong>CPU<\/strong> \u0438 \u0440\u0430\u0431\u043e\u0442\u0430\u0435\u0442 \u043f\u0440\u0438\u043c\u0435\u0440\u043d\u043e \u0437\u0430 <strong>O(N)<\/strong>, \u043f\u043e\u044d\u0442\u043e\u043c\u0443 \u043e\u0441\u043e\u0431\u043e\u0433\u043e \u0441\u043c\u044b\u0441\u043b\u0430 \u0432 \u0435\u0433\u043e \u043f\u0435\u0440\u0435\u043d\u043e\u0441\u0435 \u043d\u0430 <strong>NPU<\/strong> \u043d\u0435\u0442. <strong>A55<\/strong> \u0441\u043f\u0440\u0430\u0432\u043b\u044f\u0435\u0442\u0441\u044f \u0434\u043e\u0441\u0442\u0430\u0442\u043e\u0447\u043d\u043e \u0431\u044b\u0441\u0442\u0440\u043e.<\/p>\n<h4>Vision Encoder \u0432 VLM<\/h4>\n<p><strong>VLM<\/strong> \u043f\u043e \u0441\u0443\u0442\u0438 \u0441\u043e\u0441\u0442\u043e\u0438\u0442 \u0438\u0437 \u0434\u0432\u0443\u0445 \u043d\u0435\u0437\u0430\u0432\u0438\u0441\u0438\u043c\u044b\u0445 \u043f\u0430\u0439\u043f\u043b\u0430\u0439\u043d\u043e\u0432 \u2014 <strong>Vision Encoder<\/strong> \u0438 <strong>LLM<\/strong>. Vision \u0447\u0430\u0441\u0442\u044c \u043c\u043e\u0434\u0435\u043b\u0438 \u0441\u043e\u0431\u0438\u0440\u0430\u0435\u0442\u0441\u044f \u0432 \u043e\u0442\u0434\u0435\u043b\u044c\u043d\u0443\u044e axera \u043c\u043e\u0434\u0435\u043b\u044c \u0441 \u0444\u0438\u043a\u0441\u0438\u0440\u043e\u0432\u0430\u043d\u043d\u044b\u043c \u0440\u0430\u0437\u043c\u0435\u0440\u043e\u043c \u0432\u0445\u043e\u0434\u0430 (\u0440\u0430\u0437\u0440\u0435\u0448\u0435\u043d\u0438\u0435\u043c \u0438\u0437\u043e\u0431\u0440\u0430\u0436\u0435\u043d\u0438\u044f). \u0418\u043d\u0444\u0435\u0440\u0435\u043d\u0441 ViT \u0441\u043e\u0441\u0442\u043e\u0438\u0442 \u0438\u0437 \u0441\u043b\u0435\u0434\u0443\u044e\u0449\u0438\u0445 \u044d\u0442\u0430\u043f\u043e\u0432:<\/p>\n<ul>\n<li>\n<p>Patch Embedding \u2014 2D-\u0441\u0432\u0451\u0440\u0442\u043a\u0430 \u0441 \u044f\u0434\u0440\u043e\u043c \u0440\u0430\u0437\u043c\u0435\u0440\u0430 patch_size x patch_size. \u041e\u0431\u0440\u0430\u0431\u0430\u0442\u044b\u0432\u0430\u0435\u0442\u0441\u044f \u043d\u0430 <strong>Conv EU<\/strong><\/p>\n<\/li>\n<li>\n<p>Positional Embedding \u2014 <strong>Tensor EU<\/strong><\/p>\n<\/li>\n<li>\n<p>\u0411\u043b\u043e\u043a\u0438 \u0442\u0440\u0430\u043d\u0441\u0444\u043e\u0440\u043c\u0435\u0440\u043e\u0432, \u0442\u0430\u043a \u0436\u0435 \u043a\u0430\u043a \u0438 LLM:\u00a0<\/p>\n<ul>\n<li>\n<p>LayerNorm -&gt; QKV-\u043f\u0440\u043e\u0435\u043a\u0446\u0438\u0438 (<strong>Conv EU<\/strong>) -&gt; Attention (<strong>Conv EU + Tensor EU<\/strong>) -&gt; FFN (<strong>Conv EU<\/strong>)<\/p>\n<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<p>\u0422\u0435\u043f\u0435\u0440\u044c \u043f\u0435\u0440\u0435\u0439\u0434\u0451\u043c \u043a \u0442\u0435\u0441\u0442\u0443 \u043f\u0440\u043e\u0438\u0437\u0432\u043e\u0434\u0438\u0442\u0435\u043b\u044c\u043d\u043e\u0441\u0442\u0438 \u043d\u0435\u043a\u043e\u0442\u043e\u0440\u044b\u0445 \u043c\u043e\u0434\u0435\u043b\u0435\u0439.<\/p>\n<h3>\u041f\u0435\u0440\u0432\u044b\u0435 \u0442\u0435\u0441\u0442\u044b<\/h3>\n<figure class=\"full-width \"><img decoding=\"async\" src=\"https:\/\/habrastorage.org\/r\/w1560\/getpro\/habr\/upload_files\/5db\/e61\/fc9\/5dbe61fc9cf7b8e700a3d6d355075ee4.png\" width=\"1588\" height=\"781\" sizes=\"auto, (max-width: 780px) 100vw, 50vw\" srcset=\"https:\/\/habrastorage.org\/r\/w780\/getpro\/habr\/upload_files\/5db\/e61\/fc9\/5dbe61fc9cf7b8e700a3d6d355075ee4.png 780w,&#10;       https:\/\/habrastorage.org\/r\/w1560\/getpro\/habr\/upload_files\/5db\/e61\/fc9\/5dbe61fc9cf7b8e700a3d6d355075ee4.png 781w\" loading=\"lazy\" decode=\"async\"\/><\/figure>\n<p>\u0413\u0440\u0430\u0444\u0438\u043a \u043d\u0430\u0437\u044b\u0432\u0430\u0435\u0442\u0441\u044f <em>\u201cModel Performance Comparison (FPS)\u201d<\/em>, <strong>FPS<\/strong> \u043e\u0442\u043a\u043b\u0430\u0434\u044b\u0432\u0430\u0435\u0442\u0441\u044f \u043f\u043e \u0432\u0435\u0440\u0442\u0438\u043a\u0430\u043b\u0438. \u041f\u0440\u0438 \u044d\u0442\u043e\u043c \u043d\u0430 \u0441\u0430\u043c\u043e\u043c \u0434\u0435\u043b\u0435 \u044d\u0442\u043e \u0433\u0440\u0430\u0444\u0438\u043a \u043f\u0440\u043e\u0438\u0437\u0432\u043e\u0434\u0438\u0442\u0435\u043b\u044c\u043d\u043e\u0441\u0442\u0438 \u0442\u043e\u043b\u044c\u043a\u043e \u0438\u043d\u0444\u0435\u0440\u0435\u043d\u0441\u0430 \u043c\u043e\u0434\u0435\u043b\u0435\u0439. \u041d\u0435 \u0434\u043b\u044f \u0432\u0441\u0435\u0445 \u043c\u043e\u0434\u0435\u043b\u0435\u0439 \u0438\u043d\u0444\u0435\u0440\u0435\u043d\u0441 \u0434\u0430\u0451\u0442 \u0441\u0440\u0430\u0437\u0443 \u0433\u043e\u0442\u043e\u0432\u044b\u0439 \u0434\u043b\u044f \u0438\u0441\u043f\u043e\u043b\u044c\u0437\u043e\u0432\u0430\u043d\u0438\u044f \u0440\u0435\u0437\u0443\u043b\u044c\u0442\u0430\u0442. \u0427\u0430\u0441\u0442\u043e \u0442\u0440\u0435\u0431\u0443\u0435\u0442\u0441\u044f \u0432\u044b\u043f\u043e\u043b\u043d\u0438\u0442\u044c \u043d\u0430 <strong>CPU<\/strong> \u043d\u0435\u043a\u043e\u0442\u043e\u0440\u044b\u0439 \u043f\u043e\u0441\u0442\u043f\u0440\u043e\u0446\u0435\u0441\u0441\u0438\u043d\u0433, \u043a\u043e\u0442\u043e\u0440\u044b\u0439 \u0442\u043e\u0436\u0435 \u0437\u0430\u043d\u0438\u043c\u0430\u0435\u0442 \u0432\u0440\u0435\u043c\u044f. \u041f\u0440\u0438\u043c\u0435\u0440\u044b \u043a\u043e\u0434\u0430 \u0434\u043b\u044f \u0438\u043d\u0444\u0435\u0440\u0435\u043d\u0441\u0430 \u0435\u0441\u0442\u044c \u0432 <a href=\"https:\/\/github.com\/AXERA-TECH\/ax-samples\" rel=\"noopener noreferrer nofollow\">\u0440\u0435\u043f\u043e\u0437\u0438\u0442\u043e\u0440\u0438\u0438 \u043e\u0442 Axera<\/a>, \u0430 Model Zoo \u043d\u0430 <a href=\"https:\/\/huggingface.co\/AXERA-TECH\/models\" rel=\"noopener noreferrer nofollow\">Hugging Face<\/a>.<\/p>\n<p>\u0412\u0441\u0435 \u043c\u043e\u0434\u0435\u043b\u0438 \u0442\u0435\u0441\u0442\u0438\u0440\u043e\u0432\u0430\u043b\u0438\u0441\u044c \u0432 <strong>INT8 \u0432\u0435\u0441\u0430\u0445<\/strong>. \u0420\u0435\u0430\u043b\u0438\u0437\u0430\u0446\u0438\u044f <strong>YOLO<\/strong> \u043e\u0442 Axera \u0434\u043b\u044f \u043c\u043e\u0434\u0435\u043b\u0438 <strong>YOLO11s<\/strong> \u0437\u0430\u043d\u0438\u043c\u0430\u0435\u0442 <strong>3.2 \u043c\u0441<\/strong> \u043d\u0430 \u0438\u043d\u0444\u0435\u0440\u0435\u043d\u0441 \u0438 <strong>4.45 \u043c\u0441<\/strong> \u043d\u0430 \u043f\u043e\u0441\u0442\u0440\u043e\u0446\u0435\u0441\u0441\u0438\u043d\u0433 <em>(\u043f\u0440\u0438 7 \u0434\u0435\u0442\u0435\u043a\u0446\u0438\u044f\u0445, \u0447\u0435\u043c \u0431\u043e\u043b\u044c\u0448\u0435 \u0434\u0435\u0442\u0435\u043a\u0446\u0438\u0439 \u2014 \u0442\u0435\u043c \u0434\u043e\u043b\u044c\u0448\u0435 \u043f\u043e\u0441\u0442\u043f\u0440\u043e\u0446\u0435\u0441\u0441\u0438\u043d\u0433)<\/em>. \u0412 \u0438\u0442\u043e\u0433\u0435, \u043c\u043e\u0436\u043d\u043e \u0440\u0430\u0441\u0441\u0447\u0438\u0442\u044b\u0432\u0430\u0442\u044c \u043d\u0430 \u043b\u0430\u0442\u0435\u043d\u0442\u043d\u043e\u0441\u0442\u044c <strong>7.65 \u043c\u0441<\/strong> (130.7 FPS) \u043e\u0442 \u043f\u043e\u043b\u0443\u0447\u0435\u043d\u0438\u044f \u043a\u0430\u0434\u0440\u0430 \u0434\u043e \u0433\u043e\u0442\u043e\u0432\u044b\u0445 \u0440\u0435\u0437\u0443\u043b\u044c\u0442\u0430\u0442\u043e\u0432 \u0434\u0435\u0442\u0435\u043a\u0446\u0438\u0438. \u041b\u0430\u0442\u0435\u043d\u0442\u043d\u043e\u0441\u0442\u044c \u0438\u043d\u0444\u0435\u0440\u0435\u043d\u0441\u0430 YOLO11n \u2014 1.675 \u043c\u0441 + \u0442\u0435 \u0436\u0435 4.45 \u043c\u0441 \u043f\u043e\u0441\u0442\u043f\u0440\u043e\u0446\u0435\u0441\u0441\u0438\u043d\u0433\u0430 \u0434\u0430\u044e\u0442 \u043b\u0430\u0442\u0435\u043d\u0442\u043d\u043e\u0441\u0442\u044c \u043d\u0435 \u0441\u0438\u043b\u044c\u043d\u043e \u043c\u0435\u043d\u044c\u0448\u0435 \u2014 <strong>6.125 \u043c\u0441<\/strong> (163.2 FPS). \u0422\u0443\u0442 \u043c\u043e\u0436\u043d\u043e \u0441\u0440\u0430\u0437\u0443 \u0441\u0434\u0435\u043b\u0430\u0442\u044c \u0432\u044b\u0432\u043e\u0434, \u0447\u0442\u043e bottleneck \u0432\u043e\u0437\u043d\u0438\u043a\u0430\u0435\u0442 \u043d\u0435 \u0432 \u0441\u043a\u043e\u0440\u043e\u0441\u0442\u0438 <strong>NPU<\/strong>, \u0430 \u0432 <strong>CPU<\/strong>, \u043a\u043e\u0442\u043e\u0440\u044b\u0439 \u0441\u043b\u0438\u0448\u043a\u043e\u043c \u0434\u043e\u043b\u0433\u043e \u043f\u0440\u043e\u0438\u0437\u0432\u043e\u0434\u0438\u0442 \u043f\u043e\u0441\u0442\u043f\u0440\u043e\u0446\u0435\u0441\u0441\u0438\u043d\u0433 \u0440\u0435\u0437\u0443\u043b\u044c\u0442\u0430\u0442\u043e\u0432 \u0432\u044b\u0445\u043e\u0434\u043d\u044b\u0445 \u0441\u043b\u043e\u0451\u0432 \u043c\u043e\u0434\u0435\u043b\u0438. \u041f\u0440\u0438 \u044d\u0442\u043e\u043c, \u0435\u0441\u043b\u0438 \u0433\u043e\u0432\u043e\u0440\u0438\u0442\u044c \u043f\u0440\u043e realtime \u0434\u0435\u0442\u0435\u043a\u0446\u0438\u044e \u0441 \u043a\u0430\u043c\u0435\u0440\u044b, \u0442\u043e \u043f\u0440\u0438 \u0447\u0430\u0441\u0442\u043e\u0442\u0435 \u043a\u0430\u0434\u0440\u043e\u0432 \u0432 60 <strong>FPS<\/strong>, \u0437\u0430\u0434\u0435\u0440\u0436\u043a\u0430 \u043c\u0435\u0436\u0434\u0443 \u0447\u0442\u0435\u043d\u0438\u0435\u043c \u043a\u0430\u0434\u0440\u043e\u0432 \u0441\u043e\u0441\u0442\u0430\u0432\u0438\u0442 16, \u0432 \u043a\u043e\u0442\u043e\u0440\u044b\u0435 \u0442\u0435\u043a\u0443\u0449\u0430\u044f \u0437\u0430\u0434\u0435\u0440\u0436\u043a\u0430 \u0440\u0430\u0431\u043e\u0442\u044b <strong>YOLO<\/strong> \u0443\u043a\u043b\u0430\u0434\u044b\u0432\u0430\u0435\u0442\u0441\u044f, \u0442.\u0435. \u043a\u0430\u0434\u0440\u044b \u043f\u0440\u043e\u043f\u0443\u0441\u043a\u0430\u0442\u044c\u0441\u044f \u043d\u0435 \u0431\u0443\u0434\u0443\u0442. \u0414\u043b\u044f \u043a\u0430\u043c\u0435\u0440 \u0441 \u0431\u043e\u043b\u044c\u0448\u0435\u0439 \u0447\u0430\u0441\u0442\u043e\u0442\u043e\u0439, \u043d\u0430\u043f\u0440\u0438\u043c\u0435\u0440,<strong> IMX477<\/strong>, \u043a\u043e\u0442\u043e\u0440\u0430\u044f \u043c\u043e\u0436\u0435\u0442 \u0432\u044b\u0434\u0430\u0442\u044c <strong>240 FPS<\/strong>, \u0432\u043e\u0437\u043c\u043e\u0436\u043d\u043e, \u0441\u0430\u043c\u044b\u0439 \u043f\u0440\u043e\u0441\u0442\u043e\u0439 \u0432\u0430\u0440\u0438\u0430\u043d\u0442 \u043e\u043f\u0442\u0438\u043c\u0438\u0437\u0430\u0446\u0438\u0438 \u0431\u0443\u0434\u0435\u0442 \u0438\u0441\u043f\u043e\u043b\u044c\u0437\u043e\u0432\u0430\u043d\u0438\u0435 double buffering \u043f\u043e\u0434\u0445\u043e\u0434\u0430 <em>(NPU \u0441\u0440\u0430\u0437\u0443 \u043f\u043e\u0441\u043b\u0435 \u0438\u043d\u0444\u0435\u0440\u0435\u043d\u0441\u0430 \u043f\u0435\u0440\u0432\u043e\u0433\u043e \u043a\u0430\u0434\u0440\u0430 \u043d\u0430\u0447\u0438\u043d\u0430\u0435\u0442 \u0440\u0430\u0431\u043e\u0442\u0443 \u0441\u043e \u0441\u043b\u0435\u0434\u0443\u044e\u0449\u0438\u043c \u043a\u0430\u0434\u0440\u043e\u043c, \u043f\u043e\u043a\u0430 CPU \u0437\u0430\u043d\u0438\u043c\u0430\u0435\u0442\u0441\u044f \u043f\u043e\u0441\u0442\u043f\u0440\u043e\u0446\u0435\u0441\u0441\u0438\u043d\u0433\u043e\u043c \u043f\u0440\u0435\u0434\u044b\u0434\u0443\u0449\u0438\u0445 \u0440\u0435\u0437\u0443\u043b\u044c\u0442\u0430\u0442\u043e\u0432)<\/em>. \u0422\u0430\u043a\u0436\u0435 \u0440\u0435\u0448\u0435\u043d\u0438\u0435\u043c \u043c\u043e\u0436\u0435\u0442 \u0431\u044b\u0442\u044c \u043f\u0435\u0440\u0435\u043d\u043e\u0441 \u043f\u043e\u0441\u0442\u043f\u0440\u043e\u0446\u0435\u0441\u0441\u0438\u043d\u0433\u0430 \u043d\u0430 NPU\/DSP \u0438\u043b\u0438 \u0438\u0441\u043f\u043e\u043b\u044c\u0437\u043e\u0432\u0430\u043d\u0438\u0435 \u0430\u0440\u0445\u0438\u0442\u0435\u043a\u0442\u0443\u0440 \u0441 \u0431\u043e\u043b\u0435\u0435 \u043f\u0440\u043e\u0441\u0442\u044b\u043c \u043f\u043e\u0441\u0442\u043f\u0440\u043e\u0446\u0435\u0441\u0441\u0438\u043d\u0433\u043e\u043c, \u043d\u0430\u043f\u0440\u0438\u043c\u0435\u0440, <strong>YOLOv10<\/strong>, \u0441 \u043e\u0431\u0443\u0447\u0430\u0435\u043c\u044b\u043c <em>(learnable)<\/em> NMS-\u0444\u0438\u043b\u044c\u0442\u0440\u043e\u043c, \u043a\u043e\u0442\u043e\u0440\u044b\u0439 \u043d\u0435 \u0432\u044b\u043d\u043e\u0441\u0438\u0442\u0441\u044f \u0437\u0430 \u043f\u0440\u0435\u0434\u0435\u043b\u044b \u043c\u043e\u0434\u0435\u043b\u0438 \u0438 \u0431\u0443\u0434\u0435\u0442 \u0441\u0440\u0430\u0437\u0443 \u0432\u044b\u043f\u043e\u043b\u043d\u044f\u0442\u044c\u0441\u044f \u043d\u0430 NPU.\u00a0<\/p>\n<p><strong>YOLO<\/strong>:\u00a0<\/p>\n<ul>\n<li>\n<p>\u041a\u0432\u0430\u043d\u0442\u043e\u0432\u0430\u043d\u0438\u0435 \u0432 <strong>INT8<\/strong>\u00a0<\/p>\n<\/li>\n<li>\n<p>\u0420\u0430\u0437\u0440\u0435\u0448\u0435\u043d\u0438\u0435 <strong>640&#215;640<\/strong><\/p>\n<\/li>\n<li>\n<p>\u0418\u043d\u0444\u0435\u0440\u0435\u043d\u0441 \u0432 \u0440\u0435\u0436\u0438\u043c\u0435 <strong>NPU3<\/strong> (\u043c\u0430\u043a\u0441\u0438\u043c\u0430\u043b\u044c\u043d\u0430\u044f \u0443\u0442\u0438\u043b\u0438\u0437\u0430\u0446\u0438\u044f NPU)<\/p>\n<\/li>\n<\/ul>\n<p>\u041a\u043e\u043d\u0444\u0438\u0433\u0443\u0440\u0430\u0446\u0438\u044f \u044d\u043a\u0441\u043f\u043e\u0440\u0442\u0430 \u0447\u0435\u0440\u0435\u0437 Pulsar 2:<\/p>\n<pre><code class=\"json\">{  \"model_type\": \"ONNX\",  \"npu_mode\": \"NPU3\",  \"quant\": {    \"input_configs\": [      {        \"tensor_name\": \"images\",        \"calibration_dataset\": \".\/dataset\/coco_1000.tar\",        \"calibration_size\": 32,        \"calibration_mean\": [0, 0, 0],        \"calibration_std\": [255.0, 255.0, 255.0]      }    ],    \"calibration_method\": \"MinMax\",    \"precision_analysis\": true,    \"precision_analysis_method\":\"EndToEnd\"  },  \"input_processors\": [    {      \"tensor_name\": \"images\",      \"tensor_format\": \"BGR\",      \"src_format\": \"BGR\",      \"src_dtype\": \"U8\",      \"src_layout\": \"NHWC\"    }  ],  \"output_processors\": [    {      \"tensor_name\": \"\/model.23\/Concat_output_0\",      \"dst_perm\": [0, 2, 3, 1]    },    {      \"tensor_name\": \"\/model.23\/Concat_1_output_0\",      \"dst_perm\": [0, 2, 3, 1]    },    {      \"tensor_name\": \"\/model.23\/Concat_2_output_0\",      \"dst_perm\": [0, 2, 3, 1]    }  ],  \"compiler\": {    \"check\": 2  }}<\/code><div class=\"code-explainer\"><a href=\"https:\/\/sourcecraft.dev\/\" class=\"tm-button code-explainer__link\" style=\"visibility: hidden;\"><img style=\"width:87px;height:14px;object-fit:cover;object-position:left;\"\/><\/a><\/div><\/pre>\n<p>\u0421\u0440\u0430\u0432\u043d\u0435\u043d\u0438\u0435 <strong>YOLOv8<\/strong> \u0438 <strong>YOLO11<\/strong> \u0432 N \u0438 S \u0440\u0430\u0437\u043c\u0435\u0440\u0430\u0445:<\/p>\n<div>\n<div class=\"table\">\n<table>\n<tbody>\n<tr>\n<td>\n<p align=\"left\"><strong>\u041c\u043e\u0434\u0435\u043b\u044c<\/strong><\/p>\n<\/td>\n<td>\n<p align=\"left\"><strong>\u0418\u043d\u0444\u0435\u0440\u0435\u043d\u0441 \u043c\u043e\u0434\u0435\u043b\u0438<\/strong><\/p>\n<\/td>\n<td>\n<p align=\"left\"><strong>\u042d\u0444\u0444\u0435\u043a\u0442\u0438\u0432\u043d\u044b\u0439 FPS<\/strong><\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p align=\"left\">YOLOv8 N<\/p>\n<\/td>\n<td>\n<p align=\"left\">1.7<\/p>\n<\/td>\n<td>\n<p align=\"left\">162.6<\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p align=\"left\">YOLOv8 S<\/p>\n<\/td>\n<td>\n<p align=\"left\">3.6<\/p>\n<\/td>\n<td>\n<p align=\"left\">124.2<\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p align=\"left\">YOLO11 N<\/p>\n<\/td>\n<td>\n<p align=\"left\">1.7<\/p>\n<\/td>\n<td>\n<p align=\"left\">162.6<\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p align=\"left\">YOLO11 S<\/p>\n<\/td>\n<td>\n<p align=\"left\">3.2<\/p>\n<\/td>\n<td>\n<p align=\"left\">130.7<\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<\/div>\n<h4>Depth Anything V2<\/h4>\n<figure class=\"full-width \"><img decoding=\"async\" src=\"https:\/\/habrastorage.org\/getpro\/habr\/upload_files\/8c5\/670\/f95\/8c5670f955ea6fb0d4573e89e1d79840.gif\" width=\"640\" height=\"97\" sizes=\"auto, (max-width: 780px) 100vw, 50vw\" srcset=\"https:\/\/habrastorage.org\/getpro\/habr\/upload_files\/8c5\/670\/f95\/8c5670f955ea6fb0d4573e89e1d79840.gif 780w,&#10;       https:\/\/habrastorage.org\/getpro\/habr\/upload_files\/8c5\/670\/f95\/8c5670f955ea6fb0d4573e89e1d79840.gif 781w\" loading=\"lazy\" decode=\"async\"\/><\/figure>\n<p><strong>Depth Anything<\/strong> \u2014 \u044d\u0442\u043e \u0432\u0438\u0437\u0443\u0430\u043b\u044c\u043d\u044b\u0439 \u0442\u0440\u0430\u043d\u0441\u0444\u043e\u0440\u043c\u0435\u0440 \u0434\u043b\u044f \u0442\u043e\u0447\u043d\u043e\u0439 \u043c\u043e\u043d\u043e\u043a\u0443\u043b\u044f\u0440\u043d\u043e\u0439 \u043e\u0446\u0435\u043d\u043a\u0438 \u0433\u043b\u0443\u0431\u0438\u043d\u044b. \u0415\u0433\u043e \u043c\u043e\u0436\u043d\u043e \u0438\u0441\u043f\u043e\u043b\u044c\u0437\u043e\u0432\u0430\u0442\u044c \u0434\u043b\u044f \u043d\u0430\u0432\u0438\u0433\u0430\u0446\u0438\u0438\/SLAM, \u043f\u043b\u0430\u043d\u0438\u0440\u043e\u0432\u0430\u043d\u0438\u044f \u043f\u0443\u0442\u0438 \u0438 \u0442\u0430\u043a \u0434\u0430\u043b\u0435\u0435. \u0422\u0430\u043a \u0436\u0435 \u0440\u0435\u0437\u0443\u043b\u044c\u0442\u0430\u0442\u044b \u043c\u043e\u0434\u0435\u043b\u0438 \u043c\u043e\u0436\u043d\u043e \u0432\u0437\u0430\u0438\u043c\u043e\u0434\u043e\u043f\u043e\u043b\u043d\u044f\u0442\u044c \u043b\u0438\u0434\u0430\u0440\u043d\u044b\u043c\u0438 \u0438 \u0441\u0442\u0435\u0440\u0435\u043e \u0434\u0430\u043d\u043d\u044b\u043c\u0438.<\/p>\n<p>\u041f\u0440\u0438 \u0432\u0445\u043e\u0434\u043d\u043e\u043c \u0440\u0430\u0437\u0440\u0435\u0448\u0435\u043d\u0438\u0438 <strong>518&#215;518<\/strong> \u043b\u0430\u0442\u0435\u043d\u0442\u043d\u043e\u0441\u0442\u044c \u0438\u043d\u0444\u0435\u0440\u0435\u043d\u0441\u0430 \u043e\u043a\u043e\u043b\u043e <strong>33-34<\/strong> \u043c\u0441, \u0430 \u0432\u0440\u0435\u043c\u044f \u043f\u043e\u0441\u0442\u043f\u0440\u043e\u0446\u0435\u0441\u0441\u0438\u043d\u0433\u0430 <strong>10-13<\/strong> \u043c\u0441. \u042d\u0444\u0444\u0435\u043a\u0442\u0438\u0432\u043d\u044b\u0439 FPS \u043f\u043e\u043b\u0443\u0447\u0430\u0435\u0442\u0441\u044f \u043e\u043a\u043e\u043b\u043e <strong>20-23 FPS<\/strong>. \u041c\u043e\u0434\u0435\u043b\u044c \u0440\u0430\u0431\u043e\u0442\u0430\u0435\u0442 \u0432 \u0440\u0435\u0436\u0438\u043c\u0435 <strong>NPU3<\/strong>. \u0427\u0430\u0441\u0442\u044c \u043f\u043e\u0441\u043b\u0435\u0434\u043d\u0438\u0445 \u0441\u043b\u043e\u0451\u0432, \u044d\u043a\u0441\u043f\u043e\u0440\u0442\u0438\u0440\u043e\u0432\u0430\u043d\u044b \u0432 float16.<\/p>\n<h4>Super Point Mono VO<\/h4>\n<p><strong>SuperPoint<\/strong> \u2014 \u044d\u0442\u043e \u043f\u043e\u043b\u043d\u043e\u0441\u0442\u044c\u044e \u0441\u0432\u0451\u0440\u0442\u043e\u0447\u043d\u0430\u044f \u043d\u0435\u0439\u0440\u043e\u043d\u043d\u0430\u044f \u0441\u0435\u0442\u044c, \u043a\u043e\u0442\u043e\u0440\u0430\u044f \u0434\u0435\u0442\u0435\u043a\u0442\u0438\u0440\u0443\u0435\u0442 \u043a\u043b\u044e\u0447\u0435\u0432\u044b\u0435 \u0442\u043e\u0447\u043a\u0438 \u0438 \u0438\u0445 \u0434\u0435\u0441\u043a\u0440\u0438\u043f\u0442\u043e\u0440\u044b \u043d\u0430 \u0438\u0437\u043e\u0431\u0440\u0430\u0436\u0435\u043d\u0438\u0438 \u0437\u0430 \u043e\u0434\u0438\u043d \u043f\u0440\u043e\u0445\u043e\u0434. \u0412 \u043e\u0442\u043b\u0438\u0447\u0438\u0435 \u043e\u0442 \u043a\u043b\u0430\u0441\u0441\u0438\u0447\u0435\u0441\u043a\u0438\u0445 \u043c\u0435\u0442\u043e\u0434\u043e\u0432 (SIFT, ORB), \u043c\u043e\u0434\u0435\u043b\u044c \u0440\u0430\u0431\u043e\u0442\u0430\u0435\u0442 \u043d\u0430 \u0438\u0437\u043e\u0431\u0440\u0430\u0436\u0435\u043d\u0438\u0438 \u043f\u043e\u043b\u043d\u043e\u0433\u043e \u0440\u0430\u0437\u043c\u0435\u0440\u0430 \u0438 \u0441\u043f\u043e\u0441\u043e\u0431\u043d\u0430 \u043d\u0430\u0445\u043e\u0434\u0438\u0442\u044c \u0431\u043e\u043b\u0435\u0435 \u0448\u0438\u0440\u043e\u043a\u0438\u0439 \u043d\u0430\u0431\u043e\u0440 \u043f\u043e\u0432\u0442\u043e\u0440\u044f\u0435\u043c\u044b\u0445 \u0442\u043e\u0447\u0435\u043a \u0438\u043d\u0442\u0435\u0440\u0435\u0441\u0430. \u041f\u043e\u0438\u0441\u043a \u043a\u043b\u044e\u0447\u0435\u0432\u044b\u0445 \u0442\u043e\u0447\u0435\u043a \u043d\u0430 \u0438\u0437\u043e\u0431\u0440\u0430\u0436\u0435\u043d\u0438\u0438 \u0440\u0430\u0437\u0440\u0435\u0448\u0435\u043d\u0438\u044f <strong>640&#215;480<\/strong> \u0437\u0430\u043d\u0438\u043c\u0430\u0435\u0442 \u043e\u043a\u043e\u043b\u043e <strong>28 \u043c\u0441<\/strong> <em>(35 FPS)<\/em>. \u0412\u0445\u043e\u0434\u043d\u044b\u0435 \u0438 \u0432\u044b\u0445\u043e\u0434\u043d\u044b\u0435 \u0441\u043b\u043e\u0438 \u043d\u0435 \u043a\u0432\u0430\u043d\u0442\u043e\u0432\u0430\u043d\u044b, \u0442.\u0435. \u0440\u0435\u0437\u0443\u043b\u044c\u0442\u0438\u0440\u0443\u044e\u0449\u0438\u0435 \u0434\u0435\u0441\u043a\u0440\u0438\u043f\u0442\u043e\u0440\u044b \u043a\u043b\u044e\u0447\u0435\u0432\u044b\u0445 \u0442\u043e\u0447\u0435\u043a \u0432\u043e\u0437\u0432\u0440\u0430\u0449\u0430\u044e\u0442\u0441\u044f \u0432 float32.<\/p>\n<p>\u0414\u0430\u043b\u0435\u0435, \u0435\u0441\u043b\u0438 \u0434\u043e\u0431\u0430\u0432\u0438\u0442\u044c \u043c\u0430\u0442\u0447\u0438\u043d\u0433 \u0442\u043e\u0447\u0435\u043a <em>(\u043d\u0430\u043f\u0440\u0438\u043c\u0435\u0440, \u0447\u0435\u0440\u0435\u0437 BFMatcher)<\/em>, \u0442\u043e \u043c\u043e\u0436\u043d\u043e \u0441\u043e\u043f\u043e\u0441\u0442\u0430\u0432\u043b\u044f\u0442\u044c \u043a\u043b\u044e\u0447\u0435\u0432\u044b\u0435 \u0442\u043e\u0447\u043a\u0438 \u043c\u0435\u0436\u0434\u0443 \u0434\u0432\u0443\u043c\u044f \u043f\u043e\u0441\u043b\u0435\u0434\u043e\u0432\u0430\u0442\u0435\u043b\u044c\u043d\u044b\u043c\u0438 \u043a\u0430\u0434\u0440\u0430\u043c\u0438 \u0438 \u0437\u0430\u0442\u0435\u043c \u0432\u043e\u0441\u0441\u0442\u0430\u043d\u0430\u0432\u043b\u0438\u0432\u0430\u0442\u044c \u0444\u0438\u0437\u0438\u0447\u0435\u0441\u043a\u043e\u0435 \u043f\u0435\u0440\u0435\u043c\u0435\u0449\u0435\u043d\u0438\u0435 \u043a\u0430\u043c\u0435\u0440\u044b \u043c\u0435\u0436\u0434\u0443 \u043d\u0438\u043c\u0438 (\u043d\u043e \u0431\u0435\u0437 \u043f\u0440\u0438\u0432\u044f\u0437\u043a\u0438 \u043a \u0440\u0435\u0430\u043b\u044c\u043d\u044b\u043c \u043a\u043e\u043e\u0440\u0434\u0438\u043d\u0430\u0442\u0430\u043c). \u0422\u0430\u043a \u043c\u043e\u0436\u043d\u043e \u0441\u0434\u0435\u043b\u0430\u0442\u044c \u043f\u0440\u043e\u0441\u0442\u0443\u044e \u0432\u0438\u0437\u0443\u0430\u043b\u044c\u043d\u0443\u044e \u043e\u0434\u043e\u043c\u0435\u0442\u0440\u0438\u044e \u0441 \u043f\u043e\u0442\u0435\u043d\u0446\u0438\u0430\u043b\u044c\u043d\u044b\u043c \u0440\u0430\u0441\u0448\u0438\u0440\u0435\u043d\u0438\u0435\u043c \u0434\u043e \u043f\u043e\u043b\u043d\u043e\u0446\u0435\u043d\u043d\u043e\u0433\u043e vSLAM. \u0411\u043b\u0430\u0433\u043e\u0434\u0430\u0440\u044f \u043d\u0435\u0431\u043e\u043b\u044c\u0448\u043e\u043c\u0443 \u043a\u0432\u0430\u043d\u0442\u043e\u0432\u0430\u043d\u0438\u044e, \u043a\u043e\u0442\u043e\u0440\u043e\u0435 \u043f\u0440\u0438\u043c\u0435\u043d\u044f\u0435\u0442\u0441\u044f \u0432 \u043e\u0441\u043d\u043e\u0432\u043d\u043e\u043c \u043a \u0441\u0432\u0451\u0440\u0442\u043e\u0447\u043d\u044b\u043c \u0441\u043b\u043e\u044f\u043c \u0432\u043d\u0443\u0442\u0440\u0438 \u043c\u043e\u0434\u0435\u043b\u0438, \u043a\u0430\u0447\u0435\u0441\u0442\u0432\u043e \u043f\u043e\u0438\u0441\u043a\u0430 \u043a\u043b\u044e\u0447\u0435\u0432\u044b\u0445 \u0442\u043e\u0447\u0435\u043a \u043f\u0430\u0434\u0430\u0435\u0442 \u043d\u0435\u0437\u043d\u0430\u0447\u0438\u0442\u0435\u043b\u044c\u043d\u043e. \u041d\u043e \u043c\u0430\u0442\u0447\u0438\u043d\u0433 \u0447\u0435\u0440\u0435\u0437 <strong>BFMatcher<\/strong> \u043f\u0440\u043e\u0438\u0437\u0432\u043e\u0434\u0438\u0442\u0441\u044f \u043d\u0430 CPU \u0438 \u0437\u0430\u043d\u0438\u043c\u0430\u0435\u0442 <strong>25 \u043c\u0441<\/strong>, \u0447\u0442\u043e \u0441\u043b\u0438\u0448\u043a\u043e\u043c \u043c\u043d\u043e\u0433\u043e \u0434\u043b\u044f \u0440\u0430\u0431\u043e\u0442\u044b \u0432 \u0440\u0435\u0430\u043b\u0442\u0430\u0439\u043c\u0435. \u041f\u043e\u044d\u0442\u043e\u043c\u0443 \u0434\u043b\u044f \u043c\u0430\u0442\u0447\u0438\u043d\u0433\u0430 \u043c\u043e\u0436\u043d\u043e \u0438\u0441\u043f\u043e\u043b\u044c\u0437\u043e\u0432\u0430\u0442\u044c <strong>LightGlue<\/strong>, \u043f\u0435\u0440\u0435\u043d\u0435\u0441\u0451\u043d\u043d\u044b\u0439 \u043d\u0430 NPU. \u042d\u0442\u043e \u0443\u0432\u0435\u043b\u0438\u0447\u0438\u0442 \u0438 \u043a\u0430\u0447\u0435\u0441\u0442\u0432\u043e \u043c\u0430\u0442\u0447\u0438\u043d\u0433\u0430, \u0438 \u043f\u0440\u043e\u0438\u0437\u0432\u043e\u0434\u0438\u0442\u0435\u043b\u044c\u043d\u043e\u0441\u0442\u044c. \u0418\u043b\u0438 \u043a\u0430\u043a \u043f\u0440\u043e\u0441\u0442\u043e\u0439 \u0432\u0430\u0440\u0438\u0430\u043d\u0442, \u043c\u043e\u0436\u043d\u043e \u0438\u0441\u043f\u043e\u043b\u044c\u0437\u043e\u0432\u0430\u0442\u044c L2-\u043d\u043e\u0440\u043c\u0443.<\/p>\n<p>\u041d\u0438\u0436\u0435 \u043f\u0440\u0438\u0432\u0435\u0434\u0451\u043d \u043f\u0440\u0438\u043c\u0435\u0440 \u0432\u044b\u0447\u0438\u0441\u043b\u0435\u043d\u043d\u043e\u0439 \u043e\u0434\u043e\u043c\u0435\u0442\u0440\u0438\u0438 \u043d\u0430 \u043e\u0441\u043d\u043e\u0432\u0435 \u0434\u0435\u0442\u0435\u043a\u0446\u0438\u0438 \u043a\u043b\u044e\u0447\u0435\u0432\u044b\u0445 \u0442\u043e\u0447\u0435\u043a \u0447\u0435\u0440\u0435\u0437 <strong>SuperPoint + BFMatcher + recoverPose<\/strong> \u043d\u0430 \u0434\u0430\u0442\u0430\u0441\u0435\u0442\u0435 KITTI. \u0414\u0430\u043b\u044c\u043d\u0435\u0439\u0448\u0435\u0439 \u043e\u043f\u0442\u0438\u043c\u0438\u0437\u0430\u0446\u0438\u0435\u0439 \u043c\u043e\u0436\u0435\u0442 \u0431\u044b\u0442\u044c \u043f\u0435\u0440\u0435\u043d\u043e\u0441 \u043c\u0430\u0442\u0447\u0435\u0440\u0430 \u043d\u0430 <strong>NPU<\/strong> \u0438 <strong>findEssentialMat + recoverPose<\/strong> \u043d\u0430 <strong>DSP<\/strong>. \u0422\u0430\u043a \u043a\u0430\u043a \u0432 \u0440\u0435\u0430\u043b\u044c\u043d\u043e\u0441\u0442\u0438 \u0442\u0430\u043a\u043e\u0439 \u043f\u0430\u0439\u043f\u043b\u0430\u0439\u043d \u0440\u0430\u0431\u043e\u0442\u0430\u0435\u0442 \u043d\u0435 \u0432 \u0440\u0435\u0430\u043b\u044c\u043d\u044b\u0445 \u0440\u0430\u0437\u043c\u0435\u0440\u0430\u0445, \u0442\u043e \u0434\u043b\u044f \u0441\u043e\u043f\u043e\u0441\u0442\u0430\u0432\u043b\u0435\u043d\u0438\u044f \u0441 <em>ground truth<\/em> \u0440\u0435\u0437\u0443\u043b\u044c\u0442\u0430\u0442 \u043e\u0434\u043e\u043c\u0435\u0442\u0440\u0438\u0438 \u0441\u043a\u0435\u0439\u043b\u0438\u0442\u0441\u044f \u043d\u0430 \u043e\u0441\u043d\u043e\u0432\u0435 <em>ground truth<\/em>. \u0412 \u0440\u0435\u0430\u043b\u044c\u043d\u043e\u0441\u0442\u0438 <em>scale<\/em> \u043c\u043e\u0436\u043d\u043e \u0432\u043e\u0441\u0441\u0442\u0430\u043d\u0430\u0432\u043b\u0438\u0432\u0430\u0442\u044c \u043d\u0430 \u043e\u0441\u043d\u043e\u0432\u0435, \u043d\u0430\u043f\u0440\u0438\u043c\u0435\u0440, <strong>IMU \u0434\u0430\u043d\u043d\u044b\u0445<\/strong>.\u00a0<\/p>\n<figure class=\"full-width \"><img decoding=\"async\" src=\"https:\/\/habrastorage.org\/r\/w1560\/getpro\/habr\/upload_files\/43c\/2af\/ea0\/43c2afea0c015f95b640eeeafb6054e3.png\" alt=\"\u0412\u0438\u0434 \u0441\u0432\u0435\u0440\u0445\u0443, 2D-\u0442\u0440\u0430\u0435\u043a\u0442\u043e\u0440\u0438\u044f\" title=\"\u0412\u0438\u0434 \u0441\u0432\u0435\u0440\u0445\u0443, 2D-\u0442\u0440\u0430\u0435\u043a\u0442\u043e\u0440\u0438\u044f\" width=\"902\" height=\"790\" sizes=\"auto, (max-width: 780px) 100vw, 50vw\" srcset=\"https:\/\/habrastorage.org\/r\/w780\/getpro\/habr\/upload_files\/43c\/2af\/ea0\/43c2afea0c015f95b640eeeafb6054e3.png 780w,&#10;       https:\/\/habrastorage.org\/r\/w1560\/getpro\/habr\/upload_files\/43c\/2af\/ea0\/43c2afea0c015f95b640eeeafb6054e3.png 781w\" loading=\"lazy\" decode=\"async\"\/><\/p>\n<div><figcaption>\u0412\u0438\u0434 \u0441\u0432\u0435\u0440\u0445\u0443, 2D-\u0442\u0440\u0430\u0435\u043a\u0442\u043e\u0440\u0438\u044f<\/figcaption><\/div>\n<\/figure>\n<figure class=\"full-width \"><img decoding=\"async\" src=\"https:\/\/habrastorage.org\/r\/w1560\/getpro\/habr\/upload_files\/0b8\/2a9\/751\/0b82a97513abedb4c537a33f84313e83.png\" alt=\"\u0418\u043d\u0442\u0435\u0440\u0435\u0441\u0443\u044e\u0442 \u0442\u043e\u043b\u044c\u043a\u043e X \u0438 Z\" title=\"\u0418\u043d\u0442\u0435\u0440\u0435\u0441\u0443\u044e\u0442 \u0442\u043e\u043b\u044c\u043a\u043e X \u0438 Z\" width=\"1441\" height=\"791\" sizes=\"auto, (max-width: 780px) 100vw, 50vw\" srcset=\"https:\/\/habrastorage.org\/r\/w780\/getpro\/habr\/upload_files\/0b8\/2a9\/751\/0b82a97513abedb4c537a33f84313e83.png 780w,&#10;       https:\/\/habrastorage.org\/r\/w1560\/getpro\/habr\/upload_files\/0b8\/2a9\/751\/0b82a97513abedb4c537a33f84313e83.png 781w\" loading=\"lazy\" decode=\"async\"\/><\/p>\n<div><figcaption>\u0418\u043d\u0442\u0435\u0440\u0435\u0441\u0443\u044e\u0442 \u0442\u043e\u043b\u044c\u043a\u043e X \u0438 Z<\/figcaption><\/div>\n<\/figure>\n<h4>Qwen LLM<\/h4>\n<p>\u0414\u043b\u044f \u0442\u0435\u0441\u0442\u0430 \u044f \u0437\u0430\u043f\u0443\u0441\u0442\u0438\u043b \u043c\u0443\u043b\u044c\u0442\u0438\u043c\u043e\u0434\u0430\u043b\u044c\u043d\u0443\u044e <strong>Qwen3-VL-2B-Instruct<\/strong> \u0432 <strong>w8a16 <\/strong>\u043a\u0432\u0430\u043d\u0442\u043e\u0432\u0430\u043d\u0438\u0438. \u0412 \u0442\u043e\u043b\u044c\u043a\u043e \u0442\u0435\u043a\u0441\u0442\u043e\u0432\u043e\u043c \u0440\u0435\u0436\u0438\u043c\u0435 \u043f\u0440\u0438 \u0432\u0445\u043e\u0434\u043d\u043e\u0439 \u0434\u043b\u0438\u043d\u0435 \u0432 31 \u0442\u043e\u043a\u0435\u043d <em>(\u0441\u0438\u0441\u0442\u0435\u043c\u043d\u044b\u0439 \u043f\u0440\u043e\u043c\u043f\u0442 + \u043f\u043e\u043b\u044c\u0437\u043e\u0432\u0430\u0442\u0435\u043b\u044c\u0441\u043a\u0438\u0439 \u0437\u0430\u043f\u0440\u043e\u0441)<\/em>, ttft <em>(time to first token <\/em>\u2014<em> \u0432\u0440\u0435\u043c\u044f, \u0447\u0435\u0440\u0435\u0437 \u043a\u043e\u0442\u043e\u0440\u043e\u0435 \u043c\u043e\u0434\u0435\u043b\u044c \u043d\u0430\u0447\u043d\u0451\u0442 \u0434\u0435\u043a\u043e\u0434\u0438\u0440\u043e\u0432\u0430\u0442\u044c \u0442\u043e\u043a\u0435\u043d\u044b)<\/em> \u0441\u043e\u0441\u0442\u0430\u0432\u043b\u044f\u0435\u0442 <strong>157.91 \u043c\u0441<\/strong>. \u0421\u043a\u043e\u0440\u043e\u0441\u0442\u044c \u0433\u0435\u043d\u0435\u0440\u0430\u0446\u0438\u0438 <strong>~12.3<\/strong> \u0442\u043e\u043a\u0435\u043d\u0430 \u0432 \u0441\u0435\u043a\u0443\u043d\u0434\u0443. \u041e\u0442\u0432\u0435\u0442\u044b \u0432\u044b\u0433\u043b\u044f\u0434\u044f\u0442 \u0430\u0434\u0435\u043a\u0432\u0430\u0442\u043d\u043e (\u0432 \u0441\u0438\u0441\u0442\u0435\u043c\u043d\u043e\u043c \u043f\u0440\u043e\u043c\u043f\u0442\u0435 \u0431\u044b\u043b\u043e \u0443\u043a\u0430\u0437\u0430\u043d\u043e:<em> \u201cyou are helpful assistant specialized in computer science and robotics\u201d)<\/em>:<\/p>\n<pre><code>prompt &gt;&gt; Hello! What is NPU?media &gt;&gt;16:29:54.103 INF SetKVCache:1437 | decode_grpid:0 prefill_grpid:1 history_cap:0 total_cap:128 symbolic_cap:1 precompute_len:0 input_num_token:31 prefer_symbolic_group:016:29:54.103 INF SetKVCache:1458 | current prefill_max_token_num:128016:29:54.221 INF SetKVCache:1462 | first run16:29:54.268 INF Run:1553 | input token num : 31, prefill_split_num : 116:29:54.268 INF Run:1640 | prefill chunk p=0 history_len=0 grpid=1 kv_cache_num=0 input_tokens=3116:29:54.269 INF Run:1665 | prefill indices shape: p=0 idx_elems=384 idx_rows=3 pos_rows=016:29:54.426 INF Run:1837 | ttft: 157.91 msHello! NPU stands for **Neural Processing Unit**. It is a specialized hardware component designed to accelerate the processing of artificial intelligence (AI) and machine learning (ML) workloads, particularly those involving matrix operations and neural network computations.In the context of **AI and robotics**, NPU is commonly used in **edge computing** devices (like smartphones, IoT devices, or robotics systems) to perform real-time AI tasks more efficiently than traditional CPUs or GPUs.### Key Features of NPU:- **Specialized for AI workloads**: NPU is optimized for matrix multiplications, which are common in deep learning models.- **Lower power consumption**: Designed to run AI models efficiently on low-power devices.- **Faster for neural network operations**: Enables faster inference (prediction) of machine learning models.### Example Use Case:In robotics, an NPU can be used to process real-time video data from a camera, run object recognition algorithms, or enable autonomous decision-making in robotic systems.### In Summary:&gt; **NPU (Neural Processing Unit)** is a hardware chip designed to accelerate AI and machine learning computations, especially for tasks like image recognition, speech recognition, and real-time pattern recognition in robotics and AI systems.Let me know if you'd like to know more about how NPU compares to GPUs or CPUs!16:30:16.181 NTC Run:2102 | hit eos,decode avg 12.37 token\/s16:30:16.181 INF GetKVCache:1408 | precompute_len:301, remaining:979<\/code><div class=\"code-explainer\"><a href=\"https:\/\/sourcecraft.dev\/\" class=\"tm-button code-explainer__link\" style=\"visibility: hidden;\"><img style=\"width:14px;height:14px;object-fit:cover;object-position:left;\"\/><\/a><\/div><\/pre>\n<p>\u041c\u043e\u0436\u043d\u043e \u043f\u0440\u0435\u0434\u043f\u043e\u043b\u043e\u0436\u0438\u0442\u044c \u0438\u0441\u043f\u043e\u043b\u044c\u0437\u043e\u0432\u0430\u043d\u0438\u0435 \u043c\u043e\u0434\u0435\u043b\u0438 \u0434\u043b\u044f \u0432\u044b\u0441\u043e\u043a\u043e\u0443\u0440\u043e\u0432\u043d\u0435\u0432\u043e\u0433\u043e \u0443\u043f\u0440\u0430\u0432\u043b\u0435\u043d\u0438\u044f \u0432 \u0440\u043e\u0431\u043e\u0442\u043e\u0442\u0435\u0445\u043d\u0438\u043a\u0435 (\u043f\u043b\u0430\u043d\u0438\u0440\u043e\u0432\u0430\u043d\u0438\u0435 \u0438 \u043f\u0435\u0440\u0435\u0441\u0442\u0440\u043e\u0435\u043d\u0438\u0435 \u043c\u0438\u0441\u0441\u0438\u0438 \u043d\u0430 \u043e\u0441\u043d\u043e\u0432\u0435 \u043d\u0435\u043f\u0440\u0435\u0434\u0441\u043a\u0430\u0437\u0443\u0435\u043c\u044b\u0445 \u0441\u0446\u0435\u043d\u0430\u0440\u0438\u0435\u0432), \u0432 \u0442\u043e\u043c \u0447\u0438\u0441\u043b\u0435 \u0432\u043d\u0443\u0442\u0440\u0438 LLM \u0430\u0433\u0435\u043d\u0442\u0430, \u0442\u043e \u0435\u0441\u043b\u0438 \u0440\u0435\u0437\u0443\u043b\u044c\u0442\u0430\u0442 \u043c\u043e\u0434\u0435\u043b\u0438 \u0443\u043c\u0435\u0441\u0442\u0438\u0442\u044c \u0432 80 \u0442\u043e\u043a\u0435\u043d\u043e\u0432 (\u043f\u0440\u0438\u043c\u0435\u0440\u043d\u043e 40-50 \u0430\u043d\u0433\u043b\u0438\u0439\u0441\u043a\u0438\u0445 \u0441\u043b\u043e\u0432, \u0441\u0442\u0430\u043d\u0434\u0430\u0440\u0442\u043d\u044b\u0439 \u0442\u043e\u043a\u0435\u043d\u0438\u0437\u0430\u0442\u043e\u0440 Qwen \u0431\u0435\u0437 \u0430\u0434\u0430\u043f\u0442\u0430\u0446\u0438\u0438 \u043f\u043e\u0434 \u0437\u0430\u0434\u0430\u0447\u0443) \u0440\u0435\u0448\u0435\u043d\u0438\u0435 \u043f\u043e\u043b\u0443\u0447\u0438\u0442\u0441\u044f \u0433\u0435\u043d\u0435\u0440\u0438\u0440\u043e\u0432\u0430\u0442\u044c \u0437\u0430 ~6.6 \u0441\u0435\u043a\u0443\u043d\u0434\u044b (0.15 Hz). \u041f\u0440\u0438 \u044d\u0442\u043e\u043c, \u0435\u0441\u043b\u0438 \u0443\u043c\u0435\u043d\u044c\u0448\u0430\u0442\u044c \u0442\u0440\u0435\u0431\u043e\u0432\u0430\u043d\u0438\u0435 \u043a \u043e\u0431\u044a\u0451\u043c\u0443 \u0432\u044b\u0445\u043e\u0434\u043d\u044b\u0445 \u0442\u043e\u043a\u0435\u043d\u043e\u0432, \u043d\u0430\u043f\u0440\u0438\u043c\u0435\u0440 \u0432 2 \u0440\u0430\u0437\u0430 (40 \u0442\u043e\u043a\u0435\u043d\u043e\u0432), \u0442\u043e \u0432\u0440\u0435\u043c\u044f \u043d\u0430 \u0433\u0435\u043d\u0435\u0440\u0430\u0446\u0438\u044e \u0441\u043e\u0441\u0442\u0430\u0432\u0438\u0442 3.4 \u0441\u0435\u043a\u0443\u043d\u0434\u044b + TTFT 157 \u043c\u0441. \u041f\u0440\u043e\u0434\u043e\u043b\u0436\u0430\u044f \u0443\u043c\u0435\u043d\u044c\u0448\u0430\u0442\u044c \u043a\u043e\u043b\u0438\u0447\u0435\u0441\u0442\u0432\u043e \u0442\u043e\u043a\u0435\u043d\u043e\u0432 TTFT \u043d\u0430\u0447\u0438\u043d\u0430\u0435\u0442 \u0434\u0430\u0432\u0430\u0442\u044c \u0432\u0441\u0451 \u0431\u043e\u043b\u044c\u0448\u0443\u044e \u0437\u0430\u0434\u0435\u0440\u0436\u043a\u0443 (\u043f\u0440\u0438 80 \u0442\u043e\u043a\u0435\u043d\u0430\u0445 \u043e\u043d \u0437\u0430\u043d\u0438\u043c\u0430\u0435\u0442 2% \u0432\u0440\u0435\u043c\u0435\u043d\u0438, \u043f\u0440\u0438 40 \u0443\u0436\u0435 4.6%). \u041f\u0440\u0438 \u0434\u0430\u043b\u044c\u043d\u0435\u0439\u0448\u0435\u043c \u0441\u043e\u043a\u0440\u0430\u0449\u0435\u043d\u0438\u0438 \u0442\u043e\u043a\u0435\u043d\u043e\u0432 TTFT \u0441\u0442\u0430\u043d\u0435\u0442 \u0434\u043e\u043c\u0438\u043d\u0438\u0440\u0443\u044e\u0449\u0438\u043c \u043e\u0433\u0440\u0430\u043d\u0438\u0447\u0435\u043d\u0438\u0435\u043c, \u0438 \u0441\u043d\u0438\u0436\u0435\u043d\u0438\u0435 \u043a\u043e\u043b\u0438\u0447\u0435\u0441\u0442\u0432\u0430 \u0442\u043e\u043a\u0435\u043d\u043e\u0432 \u043f\u0435\u0440\u0435\u0441\u0442\u0430\u043d\u0435\u0442 \u0434\u0430\u0432\u0430\u0442\u044c \u043b\u0438\u043d\u0435\u0439\u043d\u044b\u0439 \u043f\u0440\u0438\u0440\u043e\u0441\u0442 \u0447\u0430\u0441\u0442\u043e\u0442\u044b. \u041a\u0430\u0447\u0435\u0441\u0442\u0432\u043e \u043e\u0442\u0432\u0435\u0442\u043e\u0432 \u043c\u043e\u0434\u0435\u043b\u0438 \u044f \u043f\u043b\u0430\u043d\u0438\u0440\u0443\u044e \u043f\u0440\u043e\u0432\u0435\u0440\u0438\u0442\u044c \u0447\u0435\u0440\u0435\u0437 \u0440\u0430\u0437\u043b\u0438\u0447\u043d\u044b\u0435 \u043e\u0442\u043a\u0440\u044b\u0442\u044b\u0435 \u0431\u0435\u043d\u0447\u043c\u0430\u0440\u043a\u0438.<\/p>\n<p>\u0422\u0435\u043f\u0435\u0440\u044c \u043c\u043e\u0436\u043d\u043e \u043f\u043e\u043f\u0440\u043e\u0431\u043e\u0432\u0430\u0442\u044c \u043c\u0443\u043b\u044c\u0442\u0438\u043c\u043e\u0434\u0430\u043b\u044c\u043d\u043e\u0441\u0442\u044c. Vision \u044d\u043d\u043a\u043e\u0434\u0435\u0440 \u0443 \u043c\u043e\u0434\u0435\u043b\u0438 \u0440\u0430\u0431\u043e\u0442\u0430\u0435\u0442 \u0441 \u0438\u0437\u043e\u0431\u0440\u0430\u0436\u0435\u043d\u0438\u044f \u0440\u0430\u0437\u0440\u0435\u0448\u0435\u043d\u0438\u044f <strong>384&#215;384<\/strong>. \u0421\u043e \u0441\u043b\u0435\u0434\u0443\u044e\u0449\u0438\u043c \u0441\u0438\u0441\u0442\u0435\u043c\u043d\u044b\u043c \u043f\u0440\u043e\u043c\u043f\u0442\u043e\u043c \u043c\u043e\u0434\u0435\u043b\u044c \u0431\u0443\u0434\u0435\u0442 \u0440\u0430\u0431\u043e\u0442\u0430\u0442\u044c \u043a\u0430\u043a \u0434\u0435\u0442\u0435\u043a\u0442\u043e\u0440 \u043e\u0431\u044a\u0435\u043a\u0442\u043e\u0432:<\/p>\n<blockquote>\n<p>You are a helpful assistant for object detection in images. When detecting objects, return a valid minified JSON array. Each entry must contain just only 2 entries: &#8212; label: the object name (string) &#8212; bbox_2d: [x_min, y_min, x_max, y_max] in normalized 0-1000 coordinates Return ONLY a minified JSON array, no explanation text.<\/p>\n<\/blockquote>\n<p>\u041c\u043e\u0434\u0435\u043b\u044c \u043d\u0430\u0442\u0438\u0432\u043d\u043e \u043e\u043f\u0435\u0440\u0438\u0440\u0443\u0435\u0442 \u043d\u043e\u0440\u043c\u0430\u043b\u0438\u0437\u043e\u0432\u0430\u043d\u043d\u044b\u043c\u0438 \u043a\u043e\u043e\u0440\u0434\u0438\u043d\u0430\u0442\u0430\u043c\u0438 \u043e\u0442 0 \u0434\u043e 1000, \u043f\u043e\u044d\u0442\u043e\u043c\u0443\u00a0<\/p>\n<p>\u0434\u043b\u044f \u043f\u043e\u0441\u043b\u0435\u0434\u0443\u044e\u0449\u0435\u0439 \u043e\u0431\u0440\u0430\u0431\u043e\u0442\u043a\u0438 \u0438\u0445 \u043d\u0443\u0436\u043d\u043e \u043f\u0435\u0440\u0435\u0441\u0447\u0438\u0442\u044b\u0432\u0430\u0442\u044c. \u0412 \u0441\u0430\u043c\u043e\u043c \u043f\u0440\u043e\u043c\u043f\u0442\u0435 \u043c\u043e\u0436\u043d\u043e \u0443\u0436\u0435 \u043e\u0433\u0440\u0430\u043d\u0438\u0447\u0438\u0432\u0430\u0442\u044c \u043a\u043e\u043b\u0438\u0447\u0435\u0441\u0442\u0432\u043e \u0434\u0435\u0442\u0435\u043a\u0446\u0438\u0439 \u0438\u043b\u0438 \u0432\u043d\u043e\u0441\u0438\u0442\u044c \u0434\u043e\u043f\u043e\u043b\u043d\u0438\u0442\u0435\u043b\u044c\u043d\u0443\u044e \u0438\u043d\u0444\u043e\u0440\u043c\u0430\u0446\u0438\u044e. <strong>KV-\u043a\u044d\u0448<\/strong> \u0441\u0438\u0441\u0442\u0435\u043c\u043d\u043e\u0433\u043e \u043f\u0440\u043e\u043c\u043f\u0442\u0430 \u0432\u044b\u0447\u0438\u0441\u043b\u044f\u0435\u0442\u0441\u044f \u043e\u0434\u0438\u043d \u0440\u0430\u0437 \u043f\u0440\u0438 \u0441\u0442\u0430\u0440\u0442\u0435, \u043f\u043e\u044d\u0442\u043e\u043c\u0443 \u043f\u0440\u0438 \u043f\u043e\u0441\u043b\u0435\u0434\u0443\u044e\u0449\u0438\u0445 \u0437\u0430\u043f\u0440\u043e\u0441\u0430\u0445 <strong>ttft<\/strong> \u0431\u0443\u0434\u0435\u0442 \u043d\u0435\u043c\u043d\u043e\u0433\u043e \u043c\u0435\u043d\u044c\u0448\u0435, \u0447\u0435\u043c \u0435\u0441\u043b\u0438 \u0431\u044b \u0432 \u043a\u0430\u0436\u0434\u043e\u043c \u043f\u0440\u043e\u043c\u043f\u0442\u0435 \u043e\u043f\u0438\u0441\u044b\u0432\u0430\u0442\u044c <strong>LLM<\/strong> \u043a\u0442\u043e \u043e\u043d\u0430 \u0442\u0430\u043a\u0430\u044f.\u00a0<\/p>\n<p>\u0422\u0430\u043a\u0436\u0435 \u0441\u0442\u043e\u0438\u0442 \u043e\u0431\u0440\u0430\u0442\u0438\u0442\u044c \u0432\u043d\u0438\u043c\u0430\u043d\u0438\u0435, \u0447\u0442\u043e json \u043d\u0435 \u0441\u0430\u043c\u044b\u0439 \u0442\u043e\u043a\u0435\u043d\u043d\u043e-\u044d\u0444\u0444\u0435\u043a\u0442\u0438\u0432\u043d\u044b\u0439 \u0444\u043e\u0440\u043c\u0430\u0442 \u0434\u043b\u044f \u043f\u0440\u0435\u0434\u0441\u0442\u0430\u0432\u043b\u0435\u043d\u0438\u044f \u0440\u0435\u0437\u0443\u043b\u044c\u0442\u0430\u0442\u043e\u0432, \u043d\u043e \u0437\u0430\u0442\u043e \u043b\u044e\u0431\u0430\u044f \u043c\u043e\u0434\u0435\u043b\u044c \u043e\u0447\u0435\u043d\u044c \u0445\u043e\u0440\u043e\u0448\u043e \u0435\u0433\u043e \u0437\u043d\u0430\u0435\u0442 \u0438 \u0447\u0451\u0442\u043a\u043e \u043f\u0440\u0438\u0434\u0435\u0440\u0436\u0438\u0432\u0430\u0435\u0442\u0441\u044f \u0444\u043e\u0440\u043c\u0430\u0442\u0430. \u0411\u043e\u043b\u0435\u0435 \u044d\u0444\u0444\u0435\u043a\u0442\u0438\u0432\u043d\u044b\u0439 \u0444\u043e\u0440\u043c\u0430\u0442 \u043c\u043e\u0436\u0435\u0442 \u0442\u0440\u0435\u0431\u043e\u0432\u0430\u0442\u044c \u043f\u043e\u043b\u043d\u043e\u0433\u043e \u0438\u043b\u0438 \u043d\u0438\u0437\u043a\u043e\u0440\u0430\u043d\u0433\u043e\u0432\u043e\u0433\u043e <em>(LoRA)<\/em> \u0434\u043e\u043e\u0431\u0443\u0447\u0435\u043d\u0438\u044f.\u00a0<\/p>\n<p>\u041f\u0440\u0438\u043c\u0435\u0440 \u0440\u0435\u0437\u0443\u043b\u044c\u0442\u0430\u0442\u0430 \u0440\u0430\u0431\u043e\u0442\u044b:<\/p>\n<pre><code class=\"json\">[{\"label\": \"car\", \"bbox_2d\": [200, 370, 371, 734]},{\"label\": \"car\", \"bbox_2d\": [500, 488, 998, 998]},{\"label\": \"car\", \"bbox_2d\": [744, 770, 1000, 1000]}]<\/code><div class=\"code-explainer\"><a href=\"https:\/\/sourcecraft.dev\/\" class=\"tm-button code-explainer__link\" style=\"visibility: hidden;\"><img style=\"width:14px;height:14px;object-fit:cover;object-position:left;\"\/><\/a><\/div><\/pre>\n<figure class=\"\"><img decoding=\"async\" src=\"https:\/\/habrastorage.org\/r\/w1560\/getpro\/habr\/upload_files\/a8f\/a24\/536\/a8fa24536fdce43d569ae04e03754eca.jpeg\" width=\"384\" height=\"384\" sizes=\"auto, (max-width: 780px) 100vw, 50vw\" srcset=\"https:\/\/habrastorage.org\/r\/w780\/getpro\/habr\/upload_files\/a8f\/a24\/536\/a8fa24536fdce43d569ae04e03754eca.jpeg 780w,&#10;       https:\/\/habrastorage.org\/r\/w1560\/getpro\/habr\/upload_files\/a8f\/a24\/536\/a8fa24536fdce43d569ae04e03754eca.jpeg 781w\" loading=\"lazy\" decode=\"async\"\/><\/figure>\n<p>\u0412\u0440\u0435\u043c\u044f \u0440\u0430\u0431\u043e\u0442\u044b \u0432\u0438\u0437\u0443\u0430\u043b\u044c\u043d\u043e\u0433\u043e \u044d\u043d\u043a\u043e\u0434\u0435\u0440\u0430 \u043f\u0440\u0438 \u0440\u0430\u0437\u0440\u0435\u0448\u0435\u043d\u0438\u0438 <strong>384&#215;384<\/strong> \u043f\u0440\u0438\u043c\u0435\u0440\u043d\u043e <strong>200 \u043c\u0441<\/strong>, \u043d\u0430 \u0440\u0430\u0437\u0440\u0435\u0448\u0435\u043d\u0438\u0438 <strong>640&#215;640<\/strong> \u2014 <strong>1015 \u043c\u0441<\/strong>.<\/p>\n<p>\u041d\u0435\u043b\u044c\u0437\u044f \u0441\u043a\u0430\u0437\u0430\u0442\u044c, \u0447\u0442\u043e \u043c\u043e\u0434\u0435\u043b\u044c \u0441\u043f\u0440\u0430\u0432\u0438\u043b\u0430\u0441\u044c \u0438\u0434\u0435\u0430\u043b\u044c\u043d\u043e, \u043d\u043e \u0438 \u0440\u0430\u043d\u0434\u043e\u043c\u043d\u044b\u043c \u0440\u0435\u0437\u0443\u043b\u044c\u0442\u0430\u0442 \u0435\u0451 \u0440\u0430\u0431\u043e\u0442\u044b \u043d\u0430\u0437\u0432\u0430\u0442\u044c \u043d\u0435\u043b\u044c\u0437\u044f. \u0412\u043e\u043e\u0431\u0449\u0435, \u043f\u0440\u0438\u043c\u0435\u043d\u0435\u043d\u0438\u0435 <strong>VLM<\/strong> \u0434\u043b\u044f \u043f\u0440\u043e\u0441\u0442\u043e\u0439 \u0434\u0435\u0442\u0435\u043a\u0446\u0438\u0438 \u043e\u0431\u044a\u0435\u043a\u0442\u043e\u0432 \u043a\u0430\u0436\u0435\u0442\u0441\u044f \u043d\u0435 \u043e\u0447\u0435\u043d\u044c \u043b\u043e\u0433\u0438\u0447\u043d\u044b\u043c \u0438 \u0434\u0430\u0436\u0435 \u0440\u0430\u0441\u0442\u043e\u0447\u0438\u0442\u0435\u043b\u044c\u043d\u044b\u043c, \u043f\u043e\u044d\u0442\u043e\u043c\u0443 \u043c\u043e\u0436\u043d\u043e \u043f\u043e\u043f\u0440\u043e\u0431\u043e\u0432\u0430\u0442\u044c \u0437\u0430\u0441\u0442\u0430\u0432\u0438\u0442\u044c \u043c\u043e\u0434\u0435\u043b\u044c \u0430\u043d\u0430\u043b\u0438\u0437\u0438\u0440\u043e\u0432\u0430\u0442\u044c \u043f\u0440\u043e\u0438\u0441\u0445\u043e\u0434\u044f\u0449\u0435\u0435 \u0438 \u043f\u043b\u0430\u043d\u0438\u0440\u043e\u0432\u0430\u0442\u044c \u0440\u0435\u0448\u0435\u043d\u0438\u0435 \u0437\u0430\u0434\u0430\u0447\u0438.<\/p>\n<p>\u041d\u0430\u043f\u0440\u0438\u043c\u0435\u0440, \u043f\u043b\u0430\u043d\u0438\u0440\u043e\u0432\u0430\u043d\u0438\u0435 \u0440\u0430\u0431\u043e\u0442\u044b \u0434\u043b\u044f \u0443\u043d\u0438\u0432\u0435\u0440\u0441\u0430\u043b\u044c\u043d\u043e\u0433\u043e \u0437\u0430\u0445\u0432\u0430\u0442\u0430\/\u043c\u0430\u043d\u0438\u043f\u0443\u043b\u044f\u0442\u043e\u0440\u0430. \u0426\u0435\u043b\u044c \u0441\u043b\u0435\u0434\u0443\u044e\u0449\u0435\u0433\u043e \u0442\u0435\u0441\u0442\u0430 \u043f\u043e\u043a\u0430\u0437\u0430\u0442\u044c, \u0447\u0442\u043e \u043c\u043e\u0434\u0435\u043b\u044c \u0441 \u0443\u0447\u0451\u0442\u043e\u043c \u043d\u0435\u0431\u043e\u043b\u044c\u0448\u043e\u0433\u043e \u043a\u043e\u043b\u0438\u0447\u0435\u0441\u0442\u0432\u0430 \u0432\u0435\u0441\u043e\u0432 \u0438 \u043a\u0432\u0430\u043d\u0442\u043e\u0432\u0430\u043d\u0438\u0438 \u043f\u0440\u043e\u0434\u043e\u043b\u0436\u0430\u0435\u0442 \u043f\u043e\u043d\u0438\u043c\u0430\u0442\u044c \u0438\u0437\u043e\u0431\u0440\u0430\u0436\u0435\u043d\u0438\u0435 \u0438 \u0433\u0435\u043d\u0435\u0440\u0438\u0440\u043e\u0432\u0430\u0442\u044c \u0430\u0434\u0435\u043a\u0432\u0430\u0442\u043d\u044b\u0439 \u0442\u0435\u043a\u0441\u0442 \u0431\u0435\u0437 \u0433\u0430\u043b\u043b\u044e\u0446\u0438\u043d\u0430\u0446\u0438\u0439. \u041a\u043e\u043d\u0435\u0447\u043d\u043e, \u0434\u043b\u044f \u0440\u0435\u0430\u043b\u044c\u043d\u043e\u0433\u043e \u0438\u0441\u043f\u043e\u043b\u044c\u0437\u043e\u0432\u0430\u043d\u0438\u044f \u0440\u0435\u0437\u0443\u043b\u044c\u0442\u0430\u0442\u043e\u0432 \u043e\u0442 \u043c\u043e\u0434\u0435\u043b\u0438, \u043d\u0443\u0436\u043d\u043e \u0431\u043e\u043b\u0435\u0435 \u0447\u0451\u0442\u043a\u043e \u043d\u0430\u0441\u0442\u0440\u0430\u0438\u0432\u0430\u0442\u044c \u0441\u0438\u0441\u0442\u0435\u043c\u043d\u044b\u0439 \u043f\u0440\u043e\u043c\u043f\u0442 \u0438 \u0443\u0436\u0438\u043c\u0430\u0442\u044c \u043e\u0442\u0432\u0435\u0442 \u0432 \u043c\u0430\u043a\u0441\u0438\u043c\u0443\u043c \u043f\u0430\u0440\u0443 \u0441\u043e\u0442\u0435\u043d \u0442\u043e\u043a\u0435\u043d\u043e\u0432.<\/p>\n<p>\u0421\u0438\u0441\u0442\u0435\u043c\u043d\u044b\u0439 \u043f\u0440\u043e\u043c\u043f\u0442:<\/p>\n<blockquote>\n<p>You are a high-level planner for a robotic grasping arm. You receive an image of fruits or vegetables on a surface. Your job is to analyze the scene and output a structured grasp plan. Generate grasp strategy with object specific methods.<\/p>\n<\/blockquote>\n<p>\u0412\u0445\u043e\u0434\u043d\u043e\u0435 \u0438\u0437\u043e\u0431\u0440\u0430\u0436\u0435\u043d\u0438\u0435:<\/p>\n<figure class=\"\"><img decoding=\"async\" src=\"https:\/\/habrastorage.org\/r\/w1560\/getpro\/habr\/upload_files\/0aa\/5c5\/4ad\/0aa5c54ad4d23750b77bb544a4fe3670.png\" width=\"384\" height=\"384\" sizes=\"auto, (max-width: 780px) 100vw, 50vw\" srcset=\"https:\/\/habrastorage.org\/r\/w780\/getpro\/habr\/upload_files\/0aa\/5c5\/4ad\/0aa5c54ad4d23750b77bb544a4fe3670.png 780w,&#10;       https:\/\/habrastorage.org\/r\/w1560\/getpro\/habr\/upload_files\/0aa\/5c5\/4ad\/0aa5c54ad4d23750b77bb544a4fe3670.png 781w\" loading=\"lazy\" decode=\"async\"\/><\/figure>\n<p>\u0420\u0435\u0437\u0443\u043b\u044c\u0442\u0430\u0442 \u0440\u0430\u0431\u043e\u0442\u044b \u043c\u043e\u0434\u0435\u043b\u0438:<\/p>\n<pre><code class=\"markdown\">### Grasp Strategy for the Image**Objective**: Safely and efficiently grasp a variety of strawberries and other fruits from a hanging basket.---### **Step-by-Step Grasp Plan**#### **1. Scene Analysis**- **Object Identification**: The image shows a hanging basket of strawberries, some green and some red, along with other fruits like grapes and possibly some green vegetables.- **Spatial Layout**: The fruits are arranged in a vertical, hanging basket. The basket is suspended by a metal frame, and the fruits are arranged in a way that allows for a vertical grasp.- **Constraints**: The basket is suspended, so the grasp must be stable and avoid falling. The fruits are not on a flat surface, so the grasp must be stable and secure.---### **Grasp Strategy**#### **Step 1: Identify the Target Object**- **Target Object**: The primary target is the strawberries in the basket.- **Object Characteristics**: The strawberries are hanging from the basket, and they are not on a flat surface. They are suspended and may be slightly tilted.#### **Step 2: Choose the Grasp Method**- **Grasp Method**: Use a **vertical grasp**.- **Why Vertical Grasp?**: The vertical grasp is ideal for hanging objects. It allows the robotic arm to grasp the object from the side, which is more stable and secure than a horizontal grasp.#### **Step 3: Position the Grasp**- **Grasp Point**: The grasp should be placed on the **side of the fruit** that is closest to the basket.- **Grasp Point Location**: The grasp should be placed on the **side of the fruit** that is closest to the basket. This ensures that the grasp is stable and secure.#### **Step 4: Apply the Grasp**- **Grasp Action**: Use the robotic arm to grasp the fruit from the side.- **Grasp Force**: Apply a moderate force to ensure the fruit is securely held without slipping.---### **Final Grasp Plan**- **Grasp Method**: Vertical grasp- **Grasp Point**: Side of the fruit, closest to the basket- **Grasp Action**: Use the robotic arm to grasp the fruit from the side- **Grasp Force**: Moderate---### **Final Answer**The robotic arm should use a vertical grasp on the side of the fruit, closest to the basket, to securely hold the fruit. The grasp should be applied with a moderate force to ensure stability and safety.<\/code><div class=\"code-explainer\"><a href=\"https:\/\/sourcecraft.dev\/\" class=\"tm-button code-explainer__link\" style=\"visibility: hidden;\"><img style=\"width:14px;height:14px;object-fit:cover;object-position:left;\"\/><\/a><\/div><\/pre>\n<p>\u041e\u0442\u0432\u0435\u0442 \u0441\u043e\u0434\u0435\u0440\u0436\u0438\u0442 \u0438\u0437\u0431\u044b\u0442\u043e\u0447\u043d\u044b\u0439 \u0442\u0435\u043a\u0441\u0442 \u0438 \u043f\u043e\u0432\u0442\u043e\u0440\u0435\u043d\u0438\u044f, \u043d\u043e \u0441\u0435\u043c\u0430\u043d\u0442\u0438\u043a\u0430 \u0441\u0446\u0435\u043d\u044b \u0440\u0430\u0441\u043f\u043e\u0437\u043d\u0430\u043d\u0430 \u043a\u043e\u0440\u0440\u0435\u043a\u0442\u043d\u043e.<\/p>\n<h3>\u0410\u0440\u0445\u0438\u0442\u0435\u043a\u0442\u0443\u0440\u043d\u0430\u044f \u0444\u0438\u043b\u043e\u0441\u043e\u0444\u0438\u044f AX650N: \u0433\u0435\u0442\u0435\u0440\u043e\u0433\u0435\u043d\u043d\u044b\u0439 \u043a\u043e\u043d\u0432\u0435\u0439\u0435\u0440<\/h3>\n<p><strong>AX650N<\/strong> \u043c\u043e\u0436\u043d\u043e \u0441\u0440\u0430\u0432\u043d\u0438\u0432\u0430\u0442\u044c \u0441 \u0434\u0440\u0443\u0433\u0438\u043c\u0438 \u043a\u0438\u0442\u0430\u0439\u0441\u043a\u0438\u043c\u0438 SoC <em>(RK3588, Kendryte K230, Sophgo BM1688)<\/em>, \u043d\u043e \u043d\u0430\u0438\u0431\u043e\u043b\u0435\u0435 \u043f\u043e\u043a\u0430\u0437\u0430\u0442\u0435\u043b\u044c\u043d\u043e \u0441\u0440\u0430\u0432\u043d\u0435\u043d\u0438\u0435 \u0441 \u043b\u0438\u0434\u0435\u0440\u043e\u043c Edge ML \u2014 <strong>NVIDIA Jetson<\/strong>. <strong>Jetson Orin<\/strong> \u0438 <strong>AX650N<\/strong> \u0440\u0435\u0448\u0430\u044e\u0442 \u043e\u0434\u043d\u0443 \u0437\u0430\u0434\u0430\u0447\u0443 \u2014 Edge ML \u0434\u043b\u044f Vision \u2014 \u043f\u0440\u0438\u043d\u0446\u0438\u043f\u0438\u0430\u043b\u044c\u043d\u043e \u0440\u0430\u0437\u043d\u044b\u043c\u0438 \u0441\u043f\u043e\u0441\u043e\u0431\u0430\u043c\u0438.<\/p>\n<p><strong>Jetson Orin<\/strong> \u2014 \u044d\u0442\u043e <strong>\u0432\u0435\u0440\u0442\u0438\u043a\u0430\u043b\u044c\u043d\u043e \u043e\u0434\u043d\u043e\u0440\u043e\u0434\u043d\u0430\u044f \u0430\u0440\u0445\u0438\u0442\u0435\u043a\u0442\u0443\u0440\u0430<\/strong>. CUDA-\u044f\u0434\u0440\u0430 \u0443\u043c\u0435\u044e\u0442 \u0432\u0441\u0451: \u0441\u0432\u0451\u0440\u0442\u043a\u0438, attention, resize, \u043c\u043e\u0440\u0444\u043e\u043b\u043e\u0433\u0438\u044e, softmax \u0438 \u0442\u0430\u043a \u0434\u0430\u043b\u0435\u0435. \u041f\u0440\u0438 \u0440\u0430\u0437\u0440\u0430\u0431\u043e\u0442\u043a\u0435 \u0440\u0435\u0430\u043b\u0438\u0437\u0443\u0435\u0442\u0441\u044f \u043e\u0434\u0438\u043d CUDA-\u0433\u0440\u0430\u0444, \u043a\u043e\u0442\u043e\u0440\u044b\u0439 \u0438\u0441\u043f\u043e\u043b\u043d\u044f\u0435\u0442\u0441\u044f \u043d\u0430 \u043e\u0434\u043d\u043e\u043c \u043f\u0443\u043b\u0435 \u0432\u044b\u0447\u0438\u0441\u043b\u0438\u0442\u0435\u043b\u044c\u043d\u044b\u0445 \u0440\u0435\u0441\u0443\u0440\u0441\u043e\u0432. \u0421\u0438\u0441\u0442\u0435\u043c\u0430 \u043f\u0440\u0435\u0434\u0441\u043a\u0430\u0437\u0443\u0435\u043c\u0430 \u0438 \u0433\u0438\u0431\u043a\u0430 \u2014 \u043b\u044e\u0431\u0430\u044f \u043e\u043f\u0435\u0440\u0430\u0446\u0438\u044f \u0432\u044b\u0440\u0430\u0436\u0430\u0435\u0442\u0441\u044f \u0447\u0435\u0440\u0435\u0437 CUDA-\u043f\u0440\u0438\u043c\u0438\u0442\u0438\u0432\u044b.<\/p>\n<p><strong>AX650N<\/strong> \u2014 \u044d\u0442\u043e <strong>\u0433\u043e\u0440\u0438\u0437\u043e\u043d\u0442\u0430\u043b\u044c\u043d\u043e \u0433\u0435\u0442\u0435\u0440\u043e\u0433\u0435\u043d\u043d\u0430\u044f \u0430\u0440\u0445\u0438\u0442\u0435\u043a\u0442\u0443\u0440\u0430<\/strong>. \u041a\u0430\u0436\u0434\u044b\u0439 \u043a\u043b\u0430\u0441\u0441 \u043e\u043f\u0435\u0440\u0430\u0446\u0438\u0439 \u0438\u043c\u0435\u0435\u0442 \u0441\u043e\u0431\u0441\u0442\u0432\u0435\u043d\u043d\u044b\u0439 \u0430\u043f\u043f\u0430\u0440\u0430\u0442\u043d\u044b\u0439 IP-\u0431\u043b\u043e\u043a. \u041d\u0430 \u0441\u0430\u043c\u043e\u043c \u0434\u0435\u043b\u0435, \u0431\u043e\u043b\u044c\u0448\u0438\u043d\u0441\u0442\u0432\u043e \u043a\u0438\u0442\u0430\u0439\u0441\u043a\u0438\u0445 AI-capable SoC \u0441\u0442\u0440\u043e\u044f\u0442\u0441\u044f \u0438\u043c\u0435\u043d\u043d\u043e \u043f\u043e \u0442\u0430\u043a\u043e\u043c\u0443 \u043f\u0440\u0438\u043d\u0446\u0438\u043f\u0443. \u041d\u0435\u0431\u043e\u043b\u044c\u0448\u043e\u0435 \u0441\u0440\u0430\u0432\u043d\u0435\u043d\u0438\u0435:<\/p>\n<div>\n<div class=\"table\">\n<table>\n<tbody>\n<tr>\n<td>\n<p align=\"left\"><strong>\u041a\u043b\u0430\u0441\u0441 \u043e\u043f\u0435\u0440\u0430\u0446\u0438\u0439<\/strong><\/p>\n<\/td>\n<td>\n<p align=\"left\"><strong>\u0411\u043b\u043e\u043a \u043d\u0430 Jetson Orin<\/strong><\/p>\n<\/td>\n<td>\n<p align=\"left\"><strong>\u0411\u043b\u043e\u043a \u043d\u0430 AX650N<\/strong><\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p align=\"left\">\u0421\u0432\u0451\u0440\u0442\u043a\u0438 CNN<\/p>\n<\/td>\n<td>\n<p align=\"left\">CUDA SM + TensorCore<\/p>\n<\/td>\n<td>\n<p align=\"left\">Conv EU (\u00d73)<\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p align=\"left\">Elementwise, Norm, Pool<\/p>\n<\/td>\n<td>\n<p align=\"left\">CUDA SM<\/p>\n<\/td>\n<td>\n<p align=\"left\">Tensor EU \/ teng (\u00d73)<\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p align=\"left\">Spatial CV ops<\/p>\n<\/td>\n<td>\n<p align=\"left\">CUDA SM (CUDA)<\/p>\n<\/td>\n<td>\n<p align=\"left\">CV EU \/ IVE<\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p align=\"left\">GEMM (attention, lm_head)<\/p>\n<\/td>\n<td>\n<p align=\"left\">TensorCore<\/p>\n<\/td>\n<td>\n<p align=\"left\">Conv EU<\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p align=\"left\">TopK, Softmax<\/p>\n<\/td>\n<td>\n<p align=\"left\">CUDA SM<\/p>\n<\/td>\n<td>\n<p align=\"left\">Tensor EU + CPU argmax<\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p align=\"left\">Feature matching, ReID<\/p>\n<\/td>\n<td>\n<p align=\"left\">cuBLAS GEMM \u043d\u0430 CUDA<\/p>\n<\/td>\n<td>\n<p align=\"left\">MAU (0.8 TOPS)<\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p align=\"left\">Fisheye \u043a\u043e\u0440\u0440\u0435\u043a\u0446\u0438\u044f<\/p>\n<\/td>\n<td>\n<p align=\"left\">CUDA kernel<\/p>\n<\/td>\n<td>\n<p align=\"left\">GDC (fixed-function)<\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p align=\"left\">Pyramid generation<\/p>\n<\/td>\n<td>\n<p align=\"left\">CUDA kernel<\/p>\n<\/td>\n<td>\n<p align=\"left\">PyraLite (fixed-function)<\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p align=\"left\">Background subtraction<\/p>\n<\/td>\n<td>\n<p align=\"left\">CUDA kernel<\/p>\n<\/td>\n<td>\n<p align=\"left\">IVE \/ GMM<\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p align=\"left\">Video encode\/decode<\/p>\n<\/td>\n<td>\n<p align=\"left\">NVENC\/NVDEC<\/p>\n<\/td>\n<td>\n<p align=\"left\">VENC\/VDEC<\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p align=\"left\">ISP + denoising<\/p>\n<\/td>\n<td>\n<p align=\"left\">CUDA + ISP<\/p>\n<\/td>\n<td>\n<p align=\"left\">AI-ISP (IFE + ITP, \u043d\u0435\u0439\u0440\u043e\u0441\u0435\u0442\u0438 \u0432\u043d\u0443\u0442\u0440\u0438)<\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p align=\"left\">\u0422\u043e\u043a\u0435\u043d embedding lookup<\/p>\n<\/td>\n<td>\n<p align=\"left\">CUDA memcpy \/ GEMM<\/p>\n<\/td>\n<td>\n<p align=\"left\">CPU mmap lookup<\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p align=\"left\">\u0422\u043e\u043a\u0435\u043d\u0438\u0437\u0430\u0446\u0438\u044f<\/p>\n<\/td>\n<td>\n<p align=\"left\">CPU<\/p>\n<\/td>\n<td>\n<p align=\"left\">CPU<\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<\/div>\n<p>\u041d\u0430 <strong>Jetson Orin<\/strong>, \u0435\u0441\u043b\u0438 \u043d\u0443\u0436\u043d\u043e \u043e\u0434\u043d\u043e\u0432\u0440\u0435\u043c\u0435\u043d\u043d\u043e \u0437\u0430\u043f\u0443\u0441\u0442\u0438\u0442\u044c \u0434\u0435\u0442\u0435\u043a\u0442\u043e\u0440 \u043e\u0431\u044a\u0435\u043a\u0442\u043e\u0432 \u0438 \u0438\u043d\u0442\u0435\u043b\u043b\u0435\u043a\u0442\u0443\u0430\u043b\u044c\u043d\u043e\u0435 ISP-\u0448\u0443\u043c\u043e\u043f\u043e\u0434\u0430\u0432\u043b\u0435\u043d\u0438\u0435, \u043e\u0431\u0430 \u0437\u0430\u043d\u0438\u043c\u0430\u044e\u0442 SM. \u041a\u043e\u043d\u043a\u0443\u0440\u0435\u043d\u0446\u0438\u044f \u0437\u0430 \u043e\u0434\u0438\u043d \u043f\u0443\u043b \u0432\u044b\u0447\u0438\u0441\u043b\u0438\u0442\u0435\u043b\u044c\u043d\u044b\u0445 \u0440\u0435\u0441\u0443\u0440\u0441\u043e\u0432 \u043d\u0435\u0438\u0437\u0431\u0435\u0436\u043d\u0430 \u2014 \u043f\u043b\u0430\u043d\u0438\u0440\u043e\u0432\u0449\u0438\u043a CUDA \u0432\u044b\u043d\u0443\u0436\u0434\u0435\u043d \u043c\u0443\u043b\u044c\u0442\u0438\u043f\u043b\u0435\u043a\u0441\u0438\u0440\u043e\u0432\u0430\u0442\u044c \u0447\u0435\u0440\u0435\u0437 CUDA streams.<\/p>\n<p>\u041d\u0430 <strong>AX650N<\/strong> ISP \u0441 GDC \u0438 \u0438\u043d\u0444\u0435\u0440\u0435\u043d\u0441 NPU \u0432\u044b\u043f\u043e\u043b\u043d\u044f\u044e\u0442\u0441\u044f \u043d\u0430 <strong>\u0444\u0438\u0437\u0438\u0447\u0435\u0441\u043a\u0438 \u043e\u0442\u0434\u0435\u043b\u044c\u043d\u044b\u0445 \u0430\u043f\u043f\u0430\u0440\u0430\u0442\u043d\u044b\u0445 IP-\u0431\u043b\u043e\u043a\u0430\u0445 \u043d\u0430 \u043e\u0434\u043d\u043e\u043c \u043a\u0440\u0438\u0441\u0442\u0430\u043b\u043b\u0435 SoC<\/strong>. Conv EU \u0438 GDC \u0440\u0430\u0431\u043e\u0442\u0430\u044e\u0442 \u043f\u0430\u0440\u0430\u043b\u043b\u0435\u043b\u044c\u043d\u043e \u0431\u0435\u0437 \u043a\u043e\u043d\u043a\u0443\u0440\u0435\u043d\u0446\u0438\u0438 \u0437\u0430 \u0432\u044b\u0447\u0438\u0441\u043b\u0438\u0442\u0435\u043b\u044c\u043d\u044b\u0435 \u0440\u0435\u0441\u0443\u0440\u0441\u044b \u2014 \u044d\u0442\u043e \u0430\u043f\u043f\u0430\u0440\u0430\u0442\u043d\u044b\u0439 \u043f\u0430\u0440\u0430\u043b\u043b\u0435\u043b\u0438\u0437\u043c \u0441 \u043d\u0443\u043b\u0435\u0432\u044b\u043c\u0438 \u043d\u0430\u043a\u043b\u0430\u0434\u043d\u044b\u043c\u0438 \u0440\u0430\u0441\u0445\u043e\u0434\u0430\u043c\u0438 \u043f\u043b\u0430\u043d\u0438\u0440\u043e\u0432\u0449\u0438\u043a\u0430. \u041d\u043e \u0432\u0441\u0451 \u043d\u0435 \u0442\u0430\u043a \u043f\u0440\u043e\u0441\u0442\u043e.\u00a0<\/p>\n<h4>\u0413\u043b\u0430\u0432\u043d\u0430\u044f \u043f\u0440\u043e\u0431\u043b\u0435\u043c\u0430 AX650N: CMM \u0438 \u043a\u043e\u043d\u043a\u0443\u0440\u0435\u043d\u0446\u0438\u044f \u0437\u0430 DDR-bandwidth<\/h4>\n<p>\u041e\u0431\u043e\u0440\u043e\u0442\u043d\u0430\u044f \u0441\u0442\u043e\u0440\u043e\u043d\u0430 \u0433\u0435\u0442\u0435\u0440\u043e\u0433\u0435\u043d\u043d\u043e\u0441\u0442\u0438 \u2014 \u0432\u0441\u0435 IP-\u0431\u043b\u043e\u043a\u0438 \u0438\u0441\u043f\u043e\u043b\u044c\u0437\u0443\u044e\u0442 \u043e\u0434\u043d\u0443 \u0444\u0438\u0437\u0438\u0447\u0435\u0441\u043a\u0443\u044e RAM. \u042d\u0442\u043e \u043d\u0435 \u0443\u043d\u0438\u043a\u0430\u043b\u044c\u043d\u0430\u044f \u043e\u0441\u043e\u0431\u0435\u043d\u043d\u043e\u0441\u0442\u044c <strong>AX650N<\/strong> \u2014 \u043d\u0430 <strong>Jetson Orin<\/strong> \u0442\u0430\u043a\u0436\u0435 \u0440\u0435\u0430\u043b\u0438\u0437\u043e\u0432\u0430\u043d\u0430 UMA <em>(unified memory architecture)<\/em>, \u0433\u0434\u0435 CUDA-\u044f\u0434\u0440\u0430, CPU, NVENC \u0438 ISP \u0434\u0435\u043b\u044f\u0442 \u0435\u0434\u0438\u043d\u0443\u044e LPDDR5. \u041f\u0440\u043e\u0431\u043b\u0435\u043c\u0430 \u043a\u043e\u043d\u043a\u0443\u0440\u0435\u043d\u0446\u0438\u0438 \u0437\u0430 DDR-bandwidth \u0441\u0443\u0449\u0435\u0441\u0442\u0432\u0443\u0435\u0442 \u043d\u0430 \u043e\u0431\u0435\u0438\u0445 \u043f\u043b\u0430\u0442\u0444\u043e\u0440\u043c\u0430\u0445. \u0420\u0430\u0437\u043d\u0438\u0446\u0430 \u2014 \u0432 \u0442\u043e\u043c, <strong>\u043d\u0430\u0441\u043a\u043e\u043b\u044c\u043a\u043e \u043e\u0441\u0442\u0440\u043e \u043e\u043d\u0430 \u043f\u0440\u043e\u044f\u0432\u043b\u044f\u0435\u0442\u0441\u044f<\/strong>.<\/p>\n<p>\u0421\u0443\u043c\u043c\u0430\u0440\u043d\u0430\u044f \u0442\u0435\u043e\u0440\u0435\u0442\u0438\u0447\u0435\u0441\u043a\u0430\u044f \u043f\u0440\u043e\u043f\u0443\u0441\u043a\u043d\u0430\u044f \u0441\u043f\u043e\u0441\u043e\u0431\u043d\u043e\u0441\u0442\u044c <strong>AX650N<\/strong> \u2014 \u043e\u043a\u043e\u043b\u043e <strong>34 GB\/s<\/strong> (\u0434\u0432\u0430 32-bit LPDDR4x @ 4266 Mbps). \u0414\u043b\u044f \u0441\u0440\u0430\u0432\u043d\u0435\u043d\u0438\u044f, Jetson Orin NX \u0438\u043c\u0435\u0435\u0442 128-bit LPDDR5 \u0441 \u043f\u0440\u043e\u043f\u0443\u0441\u043a\u043d\u043e\u0439 \u0441\u043f\u043e\u0441\u043e\u0431\u043d\u043e\u0441\u0442\u044c\u044e \u043e\u043a\u043e\u043b\u043e <strong>102 GB\/s<\/strong> \u2014 \u0440\u0430\u0437\u043d\u0438\u0446\u0430 \u0432 \u0442\u0440\u0438 \u0440\u0430\u0437\u0430. \u041d\u043e \u044d\u0442\u043e \u043d\u0435 \u0430\u0440\u0445\u0438\u0442\u0435\u043a\u0442\u0443\u0440\u043d\u0430\u044f \u043e\u0441\u043e\u0431\u0435\u043d\u043d\u043e\u0441\u0442\u044c, \u0440\u0430\u0437\u043d\u0438\u0446\u0430 \u0432 \u0441\u043a\u043e\u0440\u043e\u0441\u0442\u0438 \u043f\u0430\u043c\u044f\u0442\u0438 \u043d\u0435 \u0438\u0433\u0440\u0430\u0435\u0442 \u043a\u043b\u044e\u0447\u0435\u0432\u0443\u044e \u0440\u043e\u043b\u044c.<\/p>\n<h4>\u041f\u043e\u0447\u0435\u043c\u0443 \u043d\u0430 AX650N \u043a\u043e\u043d\u043a\u0443\u0440\u0435\u043d\u0446\u0438\u044f \u043e\u0441\u0442\u0440\u0435\u0435<\/h4>\n<p>\u041d\u0430 <strong>Jetson<\/strong> CUDA-\u044f\u0434\u0440\u0430 \u043e\u0431\u0440\u0430\u0449\u0430\u044e\u0442\u0441\u044f \u043a RAM \u0447\u0435\u0440\u0435\u0437 <strong>\u0438\u0435\u0440\u0430\u0440\u0445\u0438\u044e \u043a\u044d\u0448\u0435\u0439<\/strong> (L1 -&gt; L2 -&gt; DRAM). L2-\u043a\u044d\u0448 \u043d\u0430 Orin \u0441\u043e\u0441\u0442\u0430\u0432\u043b\u044f\u0435\u0442 \u0434\u043e 4 MB \u043d\u0430 \u043a\u043b\u0430\u0441\u0442\u0435\u0440 CUDA SM \u0438 \u0441\u043f\u043e\u0441\u043e\u0431\u0435\u043d \u0443\u0434\u0435\u0440\u0436\u0438\u0432\u0430\u0442\u044c \u0433\u043e\u0440\u044f\u0447\u0438\u0435 \u0434\u0430\u043d\u043d\u044b\u0435 \u043c\u0435\u0436\u0434\u0443 \u0432\u044b\u0437\u043e\u0432\u0430\u043c\u0438 \u043e\u0434\u043d\u043e\u0433\u043e \u0441\u043b\u043e\u044f \u0434\u043b\u044f \u0440\u0430\u0437\u043d\u044b\u0445 \u044d\u043b\u0435\u043c\u0435\u043d\u0442\u043e\u0432 \u0431\u0430\u0442\u0447\u0430, \u0441\u043d\u0438\u0436\u0430\u044f \u0444\u0430\u043a\u0442\u0438\u0447\u0435\u0441\u043a\u0443\u044e DDR-\u043d\u0430\u0433\u0440\u0443\u0437\u043a\u0443. \u041a\u043e\u043c\u043f\u0438\u043b\u044f\u0442\u043e\u0440 CUDA \u0443\u043c\u0435\u0435\u0442 \u0432 coalescing \u2014 \u043e\u0431\u044a\u0435\u0434\u0438\u043d\u0435\u043d\u0438\u0435 \u0442\u0440\u0430\u043d\u0437\u0430\u043a\u0446\u0438\u0439 32 \u043f\u043e\u0442\u043e\u043a\u043e\u0432 warp&#8217;\u0430 \u0432 \u043e\u0434\u043d\u0443 \u0448\u0438\u0440\u043e\u043a\u0443\u044e burst-\u043e\u043f\u0435\u0440\u0430\u0446\u0438\u044e \u0441 \u0432\u044b\u0441\u043e\u043a\u0438\u043c \u041a\u041f\u0414 \u0448\u0438\u043d\u044b.<\/p>\n<p>\u041d\u0430 <strong>AX650N<\/strong> \u043a\u0430\u0436\u0434\u044b\u0439 IP-\u0431\u043b\u043e\u043a \u2014 \u044d\u0442\u043e <strong>\u043a\u043e\u043d\u0435\u0447\u043d\u044b\u0439 \u0430\u0432\u0442\u043e\u043c\u0430\u0442 \u0441 \u0444\u0438\u043a\u0441\u0438\u0440\u043e\u0432\u0430\u043d\u043d\u044b\u043c \u043f\u0430\u0442\u0442\u0435\u0440\u043d\u043e\u043c \u0434\u043e\u0441\u0442\u0443\u043f\u0430<\/strong>: SDMA \u0434\u0435\u043b\u0430\u0435\u0442 burst \u043e\u043f\u0440\u0435\u0434\u0435\u043b\u0451\u043d\u043d\u043e\u0433\u043e \u0440\u0430\u0437\u043c\u0435\u0440\u0430, ISP \u043f\u0438\u0448\u0435\u0442 \u043f\u043e\u0441\u0442\u0440\u043e\u0447\u043d\u043e, MAU \u0447\u0438\u0442\u0430\u0435\u0442 \u043c\u0430\u0442\u0440\u0438\u0446\u0443 \u0446\u0435\u043b\u0438\u043a\u043e\u043c. \u041d\u0435\u0442 \u0443\u043d\u0438\u0432\u0435\u0440\u0441\u0430\u043b\u044c\u043d\u043e\u0433\u043e \u043a\u044d\u0448\u0430 \u0432\u0442\u043e\u0440\u043e\u0433\u043e \u0443\u0440\u043e\u0432\u043d\u044f, \u043a\u043e\u0442\u043e\u0440\u044b\u0439 \u0441\u0433\u043b\u0430\u0436\u0438\u0432\u0430\u043b \u0431\u044b \u0442\u0440\u0430\u043d\u0437\u0430\u043a\u0446\u0438\u0438 \u043c\u0435\u0436\u0434\u0443 \u0440\u0430\u0437\u043d\u044b\u043c\u0438 \u043c\u0430\u0441\u0442\u0435\u0440\u0430\u043c\u0438. AXI-\u0430\u0440\u0431\u0438\u0442\u0440 \u043f\u0440\u043e\u0441\u0442\u043e \u0440\u0435\u0448\u0430\u0435\u0442, \u043a\u0442\u043e \u0438\u0437 \u043c\u0430\u0441\u0442\u0435\u0440\u043e\u0432 \u043f\u043e\u043b\u0443\u0447\u0430\u0435\u0442 \u0434\u043e\u0441\u0442\u0443\u043f \u043a \u0448\u0438\u043d\u0435 \u0432 \u0434\u0430\u043d\u043d\u044b\u0439 \u0442\u0430\u043a\u0442 \u2014 \u0438 \u0435\u0441\u043b\u0438 \u043e\u0434\u043d\u043e\u0432\u0440\u0435\u043c\u0435\u043d\u043d\u043e \u0430\u043a\u0442\u0438\u0432\u043d\u044b <strong>SDMA<\/strong> (NPU), ISP, VENC \u0438 MAU, \u0430\u0440\u0431\u0438\u0442\u0440 \u0441\u0435\u0440\u0438\u0430\u043b\u0438\u0437\u0443\u0435\u0442 \u0447\u0430\u0441\u0442\u044c \u0437\u0430\u043f\u0440\u043e\u0441\u043e\u0432, \u0441\u043e\u0437\u0434\u0430\u0432\u0430\u044f stall&#8217;\u044b.<\/p>\n<h4>\u041a\u043e\u043d\u043a\u0440\u0435\u0442\u043d\u044b\u0435 bottleneck-\u0441\u0446\u0435\u043d\u0430\u0440\u0438\u0438<\/h4>\n<p><strong>1. LLM-\u0438\u043d\u0444\u0435\u0440\u0435\u043d\u0441 \u043a\u0430\u043a bandwidth-bound \u0437\u0430\u0434\u0430\u0447\u0430<\/strong><\/p>\n<p>\u041f\u0440\u0438 decode-\u0444\u0430\u0437\u0435 LLM SDMA \u043a\u0430\u0436\u0434\u044b\u0439 \u0448\u0430\u0433 \u043f\u0435\u0440\u0435\u0437\u0430\u0433\u0440\u0443\u0436\u0430\u0435\u0442 \u0432\u0435\u0441\u0430 \u0432\u0441\u0435\u0445 \u0441\u043b\u043e\u0451\u0432 \u0438\u0437 CMM \u0432 OCM. \u0414\u043b\u044f <strong>Qwen3-2B-VL<\/strong> <em>(~2.5 GB \u0432\u0435\u0441\u043e\u0432, 28 \u0441\u043b\u043e\u0451\u0432)<\/em> \u043f\u0440\u0438 <strong>10 \u0442\u043e\u043a\u0435\u043d\u0430\u0445 \u0432 \u0441\u0435\u043a\u0443\u043d\u0434\u0443<\/strong> \u044d\u0442\u043e <strong>~25 GB\/s<\/strong> \u0442\u043e\u043b\u044c\u043a\u043e \u043d\u0430 \u0437\u0430\u0433\u0440\u0443\u0437\u043a\u0443 \u0432\u0435\u0441\u043e\u0432 NPU \u2014 <strong>73%<\/strong> \u043e\u0442 \u0432\u0441\u0435\u0433\u043e \u0434\u043e\u0441\u0442\u0443\u043f\u043d\u043e\u0433\u043e bandwidth. \u0418\u043c\u0435\u043d\u043d\u043e \u043f\u043e\u044d\u0442\u043e\u043c\u0443 \u0440\u0435\u0430\u043b\u044c\u043d\u0430\u044f \u0441\u043a\u043e\u0440\u043e\u0441\u0442\u044c <strong>Qwen3-2B-VL<\/strong> \u043d\u0430 <strong>AX650N<\/strong> \u0441\u043e\u0441\u0442\u0430\u0432\u043b\u044f\u0435\u0442 \u043e\u043a\u043e\u043b\u043e <strong>10-12<\/strong> \u0442\u043e\u043a\u0435\u043d\u0430\u0445 \u0432 \u0441\u0435\u043a\u0443\u043d\u0434\u0443: \u043e\u0433\u0440\u0430\u043d\u0438\u0447\u0438\u0432\u0430\u0435\u0442 \u043d\u0435 <strong>MAC<\/strong>-\u043f\u0440\u043e\u0438\u0437\u0432\u043e\u0434\u0438\u0442\u0435\u043b\u044c\u043d\u043e\u0441\u0442\u044c <strong>EU<\/strong> \u0431\u043b\u043e\u043a\u043e\u0432, \u0430 \u043f\u0440\u043e\u043f\u0443\u0441\u043a\u043d\u0430\u044f \u0441\u043f\u043e\u0441\u043e\u0431\u043d\u043e\u0441\u0442\u044c <strong>SDMA &lt;-&gt; DDR<\/strong>. \u0415\u0441\u043b\u0438 \u0432 \u044d\u0442\u043e\u0442 \u043c\u043e\u043c\u0435\u043d\u0442 <strong>ISP<\/strong> \u043f\u0438\u0448\u0435\u0442 \u0432 CMM \u043a\u0430\u0434\u0440\u044b \u0441 \u0447\u0435\u0442\u044b\u0440\u0451\u0445 \u043a\u0430\u043c\u0435\u0440 <em>(\u0435\u0449\u0451 ~3 GB\/s)<\/em>, \u043a\u043e\u043d\u043a\u0443\u0440\u0435\u043d\u0446\u0438\u044f \u0437\u0430 <strong>DDR<\/strong> \u0441\u0442\u0430\u043d\u043e\u0432\u0438\u0442\u0441\u044f \u043a\u0440\u0438\u0442\u0438\u0447\u0435\u0441\u043a\u043e\u0439.<\/p>\n<p><strong>2. \u041c\u0430\u043b\u0435\u043d\u044c\u043a\u0438\u0439 OCM \u043a\u0430\u043a \u043c\u0443\u043b\u044c\u0442\u0438\u043f\u043b\u0438\u043a\u0430\u0442\u043e\u0440 DDR-\u043d\u0430\u0433\u0440\u0443\u0437\u043a\u0438<\/strong><\/p>\n<p>11 MB <strong>OCM<\/strong> \u0434\u043e\u0441\u0442\u0430\u0442\u043e\u0447\u043d\u043e \u0434\u043b\u044f \u043d\u0435\u0431\u043e\u043b\u044c\u0448\u0438\u0445 <strong>CNN-\u043c\u043e\u0434\u0435\u043b\u0435\u0439<\/strong>. \u0414\u043b\u044f <strong>LLM<\/strong> 11 MB <strong>OCM<\/strong> \u2014 \u044d\u0442\u043e \u0431\u0443\u043a\u0432\u0430\u043b\u044c\u043d\u043e \u043d\u0438\u0447\u0442\u043e: \u043e\u0434\u0438\u043d <strong>FFN<\/strong>-\u0441\u043b\u043e\u0439 <strong>Qwen3-2B<\/strong> <em>(hidden=2560, ffn=6912)<\/em> &#8212; \u044d\u0442\u043e 2560x6912x1 \u0431\u0430\u0439\u0442 \u2248 <strong>17 MB<\/strong>. <strong>OCM<\/strong> \u043f\u0440\u0438 LLM \u0438\u043d\u0444\u0435\u0440\u0435\u043d\u0441\u0435 \u0438\u0441\u043f\u043e\u043b\u044c\u0437\u0443\u0435\u0442\u0441\u044f \u043b\u0438\u0448\u044c \u043a\u0430\u043a <em>staging-\u0431\u0443\u0444\u0435\u0440<\/em> \u0434\u043b\u044f \u043e\u0434\u043d\u043e\u0433\u043e \u0442\u0430\u0439\u043b\u0430 \u0437\u0430 \u0440\u0430\u0437, \u0447\u0442\u043e \u043e\u0437\u043d\u0430\u0447\u0430\u0435\u0442 \u043d\u0430\u0433\u0440\u0443\u0437\u043a\u0443 \u043d\u0430 <strong>SDMA<\/strong> \u0438 \u043f\u043e-\u0444\u0430\u043a\u0442\u0443 <strong>OCM<\/strong> \u043f\u0435\u0440\u0435\u0441\u0442\u0430\u0451\u0442 \u0438\u0441\u043f\u043e\u043b\u044c\u0437\u043e\u0432\u0430\u0442\u044c\u0441\u044f \u043f\u043e \u043d\u0430\u0437\u043d\u0430\u0447\u0435\u043d\u0438\u044e.<\/p>\n<h3>\u0418\u0442\u043e\u0433\u0438<\/h3>\n<p><strong>AX650N<\/strong> \u2014 \u044d\u0442\u043e \u043c\u043e\u0449\u043d\u044b\u0439 Vision-\u043a\u043e\u043d\u0432\u0435\u0439\u0435\u0440. \u041a\u0430\u0436\u0434\u044b\u0439 \u043a\u043b\u0430\u0441\u0441 \u0437\u0430\u0434\u0430\u0447 \u043f\u043e\u043b\u0443\u0447\u0430\u0435\u0442 \u0441\u0432\u043e\u0439 \u0430\u043f\u043f\u0430\u0440\u0430\u0442\u043d\u044b\u0439 \u0431\u043b\u043e\u043a: \u0441\u0432\u0451\u0440\u0442\u043a\u0438 \u2014 <strong>Conv EU<\/strong>, \u0433\u0435\u043e\u043c\u0435\u0442\u0440\u0438\u0447\u0435\u0441\u043a\u0438\u0435 \u043f\u0440\u0435\u043e\u0431\u0440\u0430\u0437\u043e\u0432\u0430\u043d\u0438\u044f \u2014 <strong>GDC<\/strong>, \u0432\u0438\u0434\u0435\u043e \u043f\u0440\u0435\u043f\u0440\u043e\u0446\u0435\u0441\u0441\u0438\u043d\u0433 \u2014 <strong>VPU<\/strong>. \u042d\u0442\u043e \u0434\u0430\u0451\u0442 \u0440\u0435\u0430\u043b\u044c\u043d\u044b\u0439 \u043f\u0430\u0440\u0430\u043b\u043b\u0435\u043b\u0438\u0437\u043c \u0431\u0435\u0437 \u043a\u043e\u043d\u043a\u0443\u0440\u0435\u043d\u0446\u0438\u0438 \u0437\u0430 \u0432\u044b\u0447\u0438\u0441\u043b\u0438\u0442\u0435\u043b\u044c\u043d\u044b\u0435 \u0440\u0435\u0441\u0443\u0440\u0441\u044b, \u043d\u043e \u043d\u0435 \u043b\u0438\u0448\u0430\u0435\u0442 \u043e\u0442 \u043f\u0440\u043e\u0431\u043b\u0435\u043c \u0441 \u043f\u0440\u043e\u043f\u0443\u0441\u043a\u043d\u043e\u0439 \u0441\u043f\u043e\u0441\u043e\u0431\u043d\u043e\u0441\u0442\u044c\u044e \u043f\u0430\u043c\u044f\u0442\u0438.<\/p>\n<p>\u0418\u0437 \u043f\u0435\u0440\u0432\u044b\u0445 \u0442\u0435\u0441\u0442\u043e\u0432 \u043c\u043e\u0436\u043d\u043e \u0441\u0434\u0435\u043b\u0430\u0442\u044c \u0432\u044b\u0432\u043e\u0434, \u0447\u0442\u043e \u0447\u0438\u043f \u0445\u043e\u0440\u043e\u0448\u043e \u0441\u043f\u0440\u0430\u0432\u043b\u044f\u0435\u0442\u0441\u044f \u0441 Vision \u0437\u0430\u0434\u0430\u0447\u0430\u043c\u0438, \u043f\u043e\u0434 \u043a\u043e\u0442\u043e\u0440\u044b\u0435 \u0438 \u0431\u044b\u043b \u0430\u0434\u0430\u043f\u0442\u0438\u0440\u043e\u0432\u0430\u043d. YOLO \u0440\u0430\u0431\u043e\u0442\u0430\u0435\u0442 \u0432 <strong>125\u2013140 FPS<\/strong>, Depth Anything \u2014 \u043e\u043a\u043e\u043b\u043e <strong>20\u201323 FPS<\/strong>, SuperPoint \u2014 <strong>35 FPS<\/strong>. Qwen3-VL-2B \u0433\u0435\u043d\u0435\u0440\u0438\u0440\u0443\u0435\u0442 <strong>~12 \u0442\u043e\u043a\u0435\u043d\u043e\u0432 \u0432 \u0441\u0435\u043a\u0443\u043d\u0434\u0443<\/strong>, \u0447\u0442\u043e \u0443\u0436\u0435 \u0434\u043e\u0441\u0442\u0430\u0442\u043e\u0447\u043d\u043e \u0434\u043b\u044f \u0437\u0430\u0434\u0430\u0447 \u0432\u044b\u0441\u043e\u043a\u043e\u0443\u0440\u043e\u0432\u043d\u0435\u0432\u043e\u0433\u043e \u043f\u043b\u0430\u043d\u0438\u0440\u043e\u0432\u0430\u043d\u0438\u044f \u0432 \u0440\u043e\u0431\u043e\u0442\u043e\u0442\u0435\u0445\u043d\u0438\u043a\u0435, \u043f\u0443\u0441\u0442\u044c \u0438 \u043d\u0435 \u0432 \u0440\u0435\u0430\u043b\u044c\u043d\u043e\u043c \u0432\u0440\u0435\u043c\u0435\u043d\u0438. \u0421 \u0442\u043e\u0447\u043a\u0438 \u0437\u0440\u0435\u043d\u0438\u044f \u0440\u0430\u0431\u043e\u0442\u044b <strong>LLM<\/strong> \u0438 <strong>VLM<\/strong> \u0433\u043b\u0430\u0432\u043d\u044b\u043c \u043e\u0433\u0440\u0430\u043d\u0438\u0447\u0435\u043d\u0438\u0435\u043c \u043d\u0430 \u043c\u043e\u0439 \u0432\u0437\u0433\u043b\u044f\u0434 \u044f\u0432\u043b\u044f\u0435\u0442\u0441\u044f \u043d\u0435 \u0441\u0430\u043c\u0430\u044f \u044d\u0444\u0444\u0435\u043a\u0442\u0438\u0432\u043d\u0430\u044f \u0440\u0435\u0430\u043b\u0438\u0437\u0430\u0446\u0438\u044f \u043e\u043f\u0435\u0440\u0430\u0442\u0438\u0432\u043d\u043e\u0439 \u043f\u0430\u043c\u044f\u0442\u0438. \u0412 \u0441\u043b\u0435\u0434\u0443\u044e\u0449\u0435\u0439 \u0441\u0442\u0430\u0442\u044c\u0435 \u044f \u043f\u0440\u043e\u0434\u0435\u043c\u043e\u043d\u0441\u0442\u0440\u0438\u0440\u0443\u044e \u0440\u0435\u0437\u0443\u043b\u044c\u0442\u0430\u0442\u044b \u0431\u0435\u043d\u0447\u043c\u0430\u0440\u043a\u043e\u0432 <em>(\u043a\u0430\u043a \u0441\u043a\u043e\u0440\u043e\u0441\u0442\u0438 \u0440\u0430\u0431\u043e\u0442\u044b, \u0442\u0430\u043a \u0438 \u043a\u0430\u0447\u0435\u0441\u0442\u0432\u0430)<\/em> \u0434\u043b\u044f \u0440\u0430\u0437\u043b\u0438\u0447\u043d\u044b\u0445 \u0441\u0432\u0451\u0440\u0442\u043e\u0447\u043d\u044b\u0445 \u043c\u043e\u0434\u0435\u043b\u0435\u0439, \u0432\u0438\u0437\u0443\u0430\u043b\u044c\u043d\u044b\u0445 \u0442\u0440\u0430\u043d\u0441\u0444\u043e\u0440\u043c\u0435\u0440\u043e\u0432 \u0438 <strong>LLM\/VLM<\/strong>.<\/p>\n<p>\u041f\u0440\u043e\u043c\u0435\u0436\u0443\u0442\u043e\u0447\u043d\u044b\u0435 \u0440\u0435\u0437\u0443\u043b\u044c\u0442\u0430\u0442\u044b \u0438 \u0442\u0435\u0441\u0442\u044b \u0440\u0430\u0437\u043d\u043e\u0433\u043e \u0434\u0440\u0443\u0433\u043e\u0433\u043e \u0436\u0435\u043b\u0435\u0437\u0430 \u044f \u0438\u043d\u043e\u0433\u0434\u0430 \u043f\u0443\u0431\u043b\u0438\u043a\u0443\u044e \u0443 \u0441\u0435\u0431\u044f \u0432 <a href=\"https:\/\/t.me\/step_zhdanov\" rel=\"noopener noreferrer nofollow\">Telegram \u043a\u0430\u043d\u0430\u043b\u0435<\/a>.<\/p>\n<\/div>\n<p>\u0441\u0441\u044b\u043b\u043a\u0430 \u043d\u0430 \u043e\u0440\u0438\u0433\u0438\u043d\u0430\u043b \u0441\u0442\u0430\u0442\u044c\u0438 <a href=\"https:\/\/habr.com\/ru\/articles\/1035776\/\">https:\/\/habr.com\/ru\/articles\/1035776\/<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>Maix4 HAT\u0411\u043e\u043b\u044c\u0448\u0438\u043d\u0441\u0442\u0432\u043e \u0437\u0430\u0434\u0430\u0447 \u0441\u043e\u0432\u0440\u0435\u043c\u0435\u043d\u043d\u043e\u0439 \u0440\u043e\u0431\u043e\u0442\u043e\u0442\u0435\u0445\u043d\u0438\u043a\u0438 \u0442\u0430\u043a \u0438\u043b\u0438 \u0438\u043d\u0430\u0447\u0435 \u0437\u0430\u0432\u044f\u0437\u0430\u043d\u044b \u043d\u0430 \u043d\u0435\u0439\u0440\u043e\u043d\u043d\u044b\u0445 \u0441\u0435\u0442\u044f\u0445: \u0434\u0435\u0442\u0435\u043a\u0446\u0438\u044f \u043e\u0431\u044a\u0435\u043a\u0442\u043e\u0432, \u043e\u0446\u0435\u043d\u043a\u0430 \u0433\u043b\u0443\u0431\u0438\u043d\u044b, \u043b\u043e\u043a\u0430\u043b\u0438\u0437\u0430\u0446\u0438\u044f, \u043f\u043b\u0430\u043d\u0438\u0440\u043e\u0432\u0430\u043d\u0438\u0435. \u0412\u0441\u0451 \u044d\u0442\u043e \u0440\u0435\u0441\u0443\u0440\u0441\u043e\u0451\u043c\u043a\u043e, \u0438 \u0432\u043e\u043f\u0440\u043e\u0441 \u0432\u044b\u0431\u043e\u0440\u0430 \u043a\u043e\u043c\u043f\u0430\u043a\u0442\u043d\u043e\u0433\u043e \u0432\u044b\u0447\u0438\u0441\u043b\u0438\u0442\u0435\u043b\u044f (\u0434\u043e\u0441\u0442\u0430\u0442\u043e\u0447\u043d\u043e \u0447\u0430\u0441\u0442\u043e \u0430\u043b\u0433\u043e\u0440\u0438\u0442\u043c\u044b \u0434\u043e\u043b\u0436\u043d\u044b \u0440\u0430\u0431\u043e\u0442\u0430\u0442\u044c \u043b\u043e\u043a\u0430\u043b\u044c\u043d\u043e) \u0432\u0441\u0442\u0430\u0435\u0442 \u0434\u043e\u0432\u043e\u043b\u044c\u043d\u043e \u043e\u0441\u0442\u0440\u043e. \u041d\u0430 \u043f\u0440\u0430\u043a\u0442\u0438\u043a\u0435 \u0432\u044b\u0431\u043e\u0440 \u0441\u0432\u043e\u0434\u0438\u0442\u0441\u044f \u043a \u0442\u0440\u0451\u043c \u043a\u043b\u0430\u0441\u0441\u0430\u043c \u0443\u0441\u0442\u0440\u043e\u0439\u0441\u0442\u0432: NVIDIA Jetson, \u0432\u043d\u0435\u0448\u043d\u0438\u0439 \u0443\u0441\u043a\u043e\u0440\u0438\u0442\u0435\u043b\u044c (\u043e\u0434\u0438\u043d \u0438\u0437 \u0441\u0430\u043c\u044b\u0445 \u043f\u043e\u043f\u0443\u043b\u044f\u0440\u043d\u044b\u0445 \u2014 Hailo) \u0438 \u043a\u0438\u0442\u0430\u0439\u0441\u043a\u0438\u0439 (\u043d\u0435 \u0432\u0441\u0435\u0433\u0434\u0430, \u043a\u043e\u043d\u0435\u0447\u043d\u043e, \u043d\u043e \u0432 \u0441\u043e\u0432\u0440\u0435\u043c\u0435\u043d\u043d\u044b\u0445 \u0440\u0435\u0430\u043b\u0438\u044f\u0445 \u043e\u0431\u044b\u0447\u043d\u043e \u043a\u0438\u0442\u0430\u0439\u0441\u043a\u0438\u0439) SoC \u0441 \u0438\u043d\u0442\u0435\u0433\u0440\u0438\u0440\u043e\u0432\u0430\u043d\u043d\u044b\u043c NPU. \u0412 \u044d\u0442\u043e\u0439 \u0441\u0442\u0430\u0442\u044c\u0435 \u044f \u0440\u0430\u0441\u0441\u043c\u043e\u0442\u0440\u044e \u043f\u0440\u0435\u0434\u0441\u0442\u0430\u0432\u0438\u0442\u0435\u043b\u044f \u0442\u0440\u0435\u0442\u044c\u0435\u0433\u043e \u043a\u043b\u0430\u0441\u0441\u0430 \u2014 Axera AX650N, \u0430 NVIDIA Jetson \u0431\u0443\u0434\u0435\u0442 \u0438\u0441\u043f\u043e\u043b\u044c\u0437\u043e\u0432\u0430\u0442\u044c\u0441\u044f \u0434\u043b\u044f \u0441\u0440\u0430\u0432\u043d\u0435\u043d\u0438\u044f, \u0442\u0430\u043a \u043a\u0430\u043a \u044d\u0442\u043e \u0435\u0434\u0438\u043d\u0441\u0442\u0432\u0435\u043d\u043d\u043e\u0435 \u043c\u0430\u0441\u0441\u043e\u0432\u043e\u0435 edge-\u0440\u0435\u0448\u0435\u043d\u0438\u0435 \u0441 \u0443\u043d\u0438\u0432\u0435\u0440\u0441\u0430\u043b\u044c\u043d\u044b\u043c\u0438 \u0432\u044b\u0447\u0438\u0441\u043b\u0438\u0442\u0435\u043b\u044c\u043d\u044b\u043c\u0438 \u044f\u0434\u0440\u0430\u043c\u0438 (CUDA).\u042d\u0442\u043e \u043f\u0435\u0440\u0432\u0430\u044f \u0447\u0430\u0441\u0442\u044c \u0446\u0438\u043a\u043b\u0430. \u0417\u0434\u0435\u0441\u044c \u044f \u0440\u0430\u0437\u0431\u0435\u0440\u0443 \u0430\u043f\u043f\u0430\u0440\u0430\u0442\u043d\u0443\u044e \u0430\u0440\u0445\u0438\u0442\u0435\u043a\u0442\u0443\u0440\u0443 \u0441\u0430\u043c\u043e\u0433\u043e AX650N \u2014 CPU, NPU, DSP, ISP, \u043f\u0430\u043c\u044f\u0442\u044c \u2014 \u0438 \u043f\u043e\u0434\u0435\u043b\u044e\u0441\u044c \u0440\u0435\u0437\u0443\u043b\u044c\u0442\u0430\u0442\u0430\u043c\u0438 \u043f\u0435\u0440\u0432\u044b\u0445 \u0442\u0435\u0441\u0442\u043e\u0432: YOLO, Depth Anything, SuperPoint \u0438 \u043c\u0443\u043b\u044c\u0442\u0438\u043c\u043e\u0434\u0430\u043b\u044c\u043d\u044b\u0439 Qwen3. \u041f\u043e\u0434\u0440\u043e\u0431\u043d\u044b\u0435 \u0431\u0435\u043d\u0447\u043c\u0430\u0440\u043a\u0438 \u0438 \u0441\u0440\u0430\u0432\u043d\u0435\u043d\u0438\u044f \u2014 \u0432\u043e \u0432\u0442\u043e\u0440\u043e\u0439 \u0447\u0430\u0441\u0442\u0438.\u042f \u0442\u0435\u0441\u0442\u0438\u0440\u043e\u0432\u0430\u043b AX650N \u0432 \u0440\u0430\u043c\u043a\u0430\u0445 \u0433\u043e\u0442\u043e\u0432\u043e\u0433\u043e \u0443\u0441\u0442\u0440\u043e\u0439\u0441\u0442\u0432\u0430 \u043e\u0442 Sipeed \u2014 Maix4 Hat. \u041e\u043d \u0441\u043e\u0441\u0442\u043e\u0438\u0442 \u0438\u0437 \u0434\u0432\u0443\u0445 \u0447\u0430\u0441\u0442\u0435\u0439: SoM, \u043d\u0430 \u043a\u043e\u0442\u043e\u0440\u043e\u043c \u0440\u0430\u0441\u043f\u043e\u043b\u043e\u0436\u0435\u043d\u044b SoC \u0438 8 GB RAM (2&#215;4 GB, \u0442\u0430\u043a \u043a\u0430\u043a \u0443 AX650N \u0434\u0432\u0430 \u043e\u0442\u0434\u0435\u043b\u044c\u043d\u044b\u0445 DDR-\u043a\u043e\u043d\u0442\u0440\u043e\u043b\u043b\u0435\u0440\u0430), \u0438 baseboard \u043e\u0442 Sipeed \u0441 \u043c\u0438\u043d\u0438\u043c\u0430\u043b\u044c\u043d\u044b\u043c \u043a\u043e\u043b\u0438\u0447\u0435\u0441\u0442\u0432\u043e\u043c \u0438\u043d\u0442\u0435\u0440\u0444\u0435\u0439\u0441\u043e\u0432. \u0421\u043a\u0440\u043e\u043c\u043d\u043e\u0441\u0442\u044c \u0438\u043d\u0442\u0435\u0440\u0444\u0435\u0439\u0441\u043e\u0432 \u043e\u0431\u044a\u044f\u0441\u043d\u044f\u0435\u0442\u0441\u044f \u043f\u0440\u043e\u0441\u0442\u043e: baseboard \u2014 \u044d\u0442\u043e HAT \u0434\u043b\u044f Raspberry Pi 5, \u043f\u043e\u0434\u043a\u043b\u044e\u0447\u0430\u044e\u0449\u0438\u0439\u0441\u044f \u043f\u043e PCIe 2.0. \u0412 \u0442\u0430\u043a\u043e\u0439 \u043a\u043e\u043d\u0444\u0438\u0433\u0443\u0440\u0430\u0446\u0438\u0438 AX650N \u0440\u0430\u0431\u043e\u0442\u0430\u0435\u0442 \u043a\u0430\u043a \u0432\u043d\u0435\u0448\u043d\u0438\u0439 ML-\u0443\u0441\u043a\u043e\u0440\u0438\u0442\u0435\u043b\u044c, \u0430\u043d\u0430\u043b\u043e\u0433\u0438\u0447\u043d\u043e Hailo. \u0412 \u0440\u0430\u043c\u043a\u0430\u0445 \u044d\u0442\u043e\u0439 \u0438 \u043f\u043e\u0441\u043b\u0435\u0434\u0443\u044e\u0449\u0438\u0445 \u0441\u0442\u0430\u0442\u0435\u0439 \u044f \u0431\u0443\u0434\u0443 \u0438\u0441\u043f\u043e\u043b\u044c\u0437\u043e\u0432\u0430\u0442\u044c Maix4 Hat \u043a\u0430\u043a \u0441\u0430\u043c\u043e\u0441\u0442\u043e\u044f\u0442\u0435\u043b\u044c\u043d\u044b\u0439 \u043c\u0438\u043a\u0440\u043e\u043a\u043e\u043c\u043f\u044c\u044e\u0442\u0435\u0440.\u0423 Sipeed \u0435\u0441\u0442\u044c \u0438 \u043f\u043e\u043b\u043d\u043e\u0446\u0435\u043d\u043d\u044b\u0439 baseboard \u0434\u043b\u044f AX650N \u2014 M4N Dock. \u0412 \u043d\u0451\u043c \u0443\u0436\u0435 \u0432\u044b\u0432\u0435\u0434\u0435\u043d Ethernet, PCIe \u0438 2 \u0447\u0435\u0442\u044b\u0440\u0451\u0445-\u043b\u0435\u0439\u043d\u043e\u0432\u044b\u0445 MIPI CSI \u0434\u043b\u044f \u043a\u0430\u043c\u0435\u0440. \u041d\u043e \u0441\u0442\u043e\u0438\u0442 \u043e\u043d \u043f\u0440\u0438\u043c\u0435\u0440\u043d\u043e \u0432\u0434\u0432\u043e\u0435 \u0434\u043e\u0440\u043e\u0436\u0435 HAT-\u0432\u0435\u0440\u0441\u0438\u0438 (45 \u0442\u044b\u0441\u044f\u0447 \u0440\u0443\u0431\u043b\u0435\u0439). \u0412 \u043f\u043b\u0430\u043d\u0435 \u043f\u0440\u043e\u0438\u0437\u0432\u043e\u0434\u0438\u0442\u0435\u043b\u044c\u043d\u043e\u0441\u0442\u0438 \u0440\u0430\u0437\u043d\u0438\u0446\u044b \u043d\u0435\u0442 \u2014 SoM \u043e\u0434\u0438\u043d\u0430\u043a\u043e\u0432\u044b\u0439.\u0417\u0430\u044f\u0432\u043b\u0435\u043d\u043d\u044b\u0435 18 TOPS@INT8 \u2014 \u043c\u0430\u0440\u043a\u0435\u0442\u0438\u043d\u0433\u043e\u0432\u043e\u0435 \u0447\u0438\u0441\u043b\u043e, \u043a\u043e\u0442\u043e\u0440\u043e\u0435 \u0432\u043a\u043b\u044e\u0447\u0430\u0435\u0442 \u043f\u0440\u043e\u0438\u0437\u0432\u043e\u0434\u0438\u0442\u0435\u043b\u044c\u043d\u043e\u0441\u0442\u044c \u043d\u0435\u0441\u043a\u043e\u043b\u044c\u043a\u0438\u0445 \u043f\u043e\u0434\u0441\u0438\u0441\u0442\u0435\u043c \u0441\u0440\u0430\u0437\u0443. \u0412 \u0434\u0430\u0442\u0430\u0448\u0438\u0442\u0435 \u0434\u043b\u044f \u0441\u0430\u043c\u043e\u0433\u043e NPU \u0443\u043a\u0430\u0437\u0430\u043d\u043e 10.8 TOPS@INT8. \u0414\u0430\u043b\u0435\u0435 \u0432 \u0441\u0442\u0430\u0442\u044c\u0435 \u0431\u0443\u0434\u0435\u0442 \u0440\u0430\u0437\u043e\u0431\u0440\u0430\u043d\u043e, \u043e\u0442\u043a\u0443\u0434\u0430 \u0431\u0435\u0440\u0451\u0442\u0441\u044f \u0440\u0430\u0437\u043d\u0438\u0446\u0430 \u0438 \u043a\u0443\u0434\u0430 \u0434\u0435\u043b\u0438\u0441\u044c \u0435\u0449\u0451 7 TOPS. \u0412 \u043b\u044e\u0431\u043e\u043c \u0441\u043b\u0443\u0447\u0430\u0435, \u0434\u0430\u0436\u0435 10 TOPS \u0432\u043f\u043e\u043b\u043d\u0435 \u0434\u043e\u0441\u0442\u0430\u0442\u043e\u0447\u043d\u043e \u0434\u043b\u044f \u0438\u043d\u0444\u0435\u0440\u0435\u043d\u0441\u0430 \u0441\u0432\u0451\u0440\u0442\u043e\u0447\u043d\u044b\u0445 \u043c\u043e\u0434\u0435\u043b\u0435\u0439 (\u043d\u0430\u043f\u0440\u0438\u043c\u0435\u0440, YOLO), \u043d\u043e \u0432\u043e\u0442 \u043d\u0430\u0441\u043a\u043e\u043b\u044c\u043a\u043e \u044d\u0442\u043e\u0442 \u0447\u0438\u043f \u0441\u043f\u0440\u0430\u0432\u0438\u0442\u0441\u044f \u0441 \u0431\u043e\u043b\u0435\u0435 \u0442\u044f\u0436\u0451\u043b\u044b\u043c\u0438 \u0437\u0430\u0434\u0430\u0447\u0430\u043c\u0438 \u2014 \u0432\u0438\u0437\u0443\u0430\u043b\u044c\u043d\u044b\u043c\u0438 \u0442\u0440\u0430\u043d\u0441\u0444\u043e\u0440\u043c\u0435\u0440\u0430\u043c\u0438 \u0438 LLM\/VLM \u2014 \u043f\u0440\u0435\u0434\u0441\u0442\u043e\u0438\u0442 \u0432\u044b\u044f\u0441\u043d\u0438\u0442\u044c.\u0427\u0442\u043e \u0432\u043d\u0443\u0442\u0440\u0438: \u0430\u0440\u0445\u0438\u0442\u0435\u043a\u0442\u0443\u0440\u0430 AX650N\u0427\u0430\u0441\u0442\u044c \u0434\u043e\u043a\u0443\u043c\u0435\u043d\u0442\u0430\u0446\u0438\u0438 Sipeed \u0432\u044b\u043b\u043e\u0436\u0438\u043b\u0438 \u0443 \u0441\u0435\u0431\u044f. \u041d\u0430 \u043e\u0441\u043d\u043e\u0432\u0435 \u0434\u043e\u0441\u0442\u0443\u043f\u043d\u044b\u0445 \u0434\u0430\u0442\u0430\u0448\u0438\u0442\u043e\u0432 \u0438 \u0442\u0435\u0441\u0442\u043e\u0432 \u043d\u0430 \u0440\u0435\u0430\u043b\u044c\u043d\u043e\u043c \u0436\u0435\u043b\u0435\u0437\u0435 \u044f \u0434\u0430\u043b\u0435\u0435 \u043e\u043f\u0438\u0441\u0430\u043b \u043e\u0431\u0449\u0443\u044e \u0430\u0440\u0445\u0438\u0442\u0435\u043a\u0442\u0443\u0440\u0443 \u0438 \u043e\u0441\u043e\u0431\u0435\u043d\u043d\u043e\u0441\u0442\u0438 \u0440\u0430\u0431\u043e\u0442\u044b NPU.AX650N \u043e\u0431\u044a\u0435\u0434\u0438\u043d\u044f\u0435\u0442 \u0432\u043e\u0441\u0435\u043c\u044c \u044f\u0434\u0435\u0440 Cortex-A55, \u0441\u043e\u0431\u0441\u0442\u0432\u0435\u043d\u043d\u044b\u0439 NPU Axera Neutron, AI-ISP \u0441 \u0430\u043f\u043f\u0430\u0440\u0430\u0442\u043d\u043e\u0439 \u043d\u0435\u0439\u0440\u043e\u0441\u0435\u0442\u0435\u0432\u043e\u0439 \u043e\u0431\u0440\u0430\u0431\u043e\u0442\u043a\u043e\u0439, \u0434\u0432\u043e\u0439\u043d\u043e\u0439 DSP \u043d\u0430 \u044f\u0434\u0440\u0430\u0445 Tensilica Vision Q7, \u0430\u043f\u043f\u0430\u0440\u0430\u0442\u043d\u044b\u0435 \u0432\u0438\u0434\u0435\u043e\u043a\u043e\u0434\u0435\u043a\u0438 \u0438 \u0446\u0435\u043b\u044b\u0439 \u043d\u0430\u0431\u043e\u0440 \u0434\u043e\u043f\u043e\u043b\u043d\u0438\u0442\u0435\u043b\u044c\u043d\u044b\u0445 \u0441\u043f\u0435\u0446\u0438\u0430\u043b\u0438\u0437\u0438\u0440\u043e\u0432\u0430\u043d\u043d\u044b\u0445 Vision-\u043c\u043e\u0434\u0443\u043b\u0435\u0439. GPU \u043d\u0435\u0442 \u0432\u043e\u043e\u0431\u0449\u0435 \u2014 \u0438 \u044d\u0442\u043e \u043f\u0440\u0438\u043d\u0446\u0438\u043f\u0438\u0430\u043b\u044c\u043d\u044b\u0439 \u0432\u044b\u0431\u043e\u0440 \u0430\u0440\u0445\u0438\u0442\u0435\u043a\u0442\u0443\u0440\u044b (\u0438\u043b\u0438 \u043f\u0440\u043e\u0441\u0442\u043e \u0443\u0434\u0435\u0448\u0435\u0432\u043b\u0435\u043d\u0438\u0435), \u043a \u043e\u0431\u0441\u0443\u0436\u0434\u0435\u043d\u0438\u044e \u043a\u043e\u0442\u043e\u0440\u043e\u0433\u043e \u044f \u0432\u0435\u0440\u043d\u0443\u0441\u044c \u0432 \u043a\u043e\u043d\u0446\u0435 \u0441\u0442\u0430\u0442\u044c\u0438.CPU\u0412\u043e\u0441\u0435\u043c\u044c \u044f\u0434\u0435\u0440 Cortex-A55 \u043e\u0440\u0433\u0430\u043d\u0438\u0437\u043e\u0432\u0430\u043d\u044b \u0432 \u0435\u0434\u0438\u043d\u044b\u0439 \u043a\u043b\u0430\u0441\u0442\u0435\u0440 \u0441 \u0442\u0440\u0451\u0445\u0443\u0440\u043e\u0432\u043d\u0435\u0432\u044b\u043c \u043a\u044d\u0448\u0435\u043c:L1: 32 KB I-Cache + 32 KB D-Cache \u043d\u0430 \u043a\u0430\u0436\u0434\u043e\u0435 \u044f\u0434\u0440\u043eL2: 64 KB \u043d\u0430 \u043a\u0430\u0436\u0434\u043e\u0435 \u044f\u0434\u0440\u043eL3: \u043e\u0431\u0449\u0438\u0439 512 KB \u043d\u0430 \u0432\u0435\u0441\u044c \u043a\u043b\u0430\u0441\u0442\u0435\u0440\u0427\u0430\u0441\u0442\u043e\u0442\u0430: \u0434\u043e 1.7 GHz \u043f\u0440\u0438 \u043d\u0430\u043f\u0440\u044f\u0436\u0435\u043d\u0438\u0438 1.0 V; \u0434\u043e 1.1 GHz \u043f\u0440\u0438 0.8 VSIMD\/FPU: \u043f\u043e\u0434\u0434\u0435\u0440\u0436\u043a\u0430 NEON \u0438 \u0438\u043d\u0442\u0435\u0433\u0440\u0438\u0440\u043e\u0432\u0430\u043d\u043d\u044b\u0439 FPUCortex-A55 \u2014 \u043d\u0435 \u0441\u0430\u043c\u044b\u0435 \u043c\u043e\u0449\u043d\u044b\u0435 \u044f\u0434\u0440\u0430 \u0434\u043b\u044f 2026 \u0433\u043e\u0434\u0430 (\u043d\u0430\u043f\u0440\u0438\u043c\u0435\u0440, \u0443 RK3588 4xA76 + 4xA55), \u043d\u043e \u0434\u043b\u044f Vision-\u043f\u0430\u0439\u043f\u043b\u0430\u0439\u043d\u043e\u0432 \u044d\u0442\u043e\u0433\u043e \u043e\u0431\u044b\u0447\u043d\u043e \u0434\u043e\u0441\u0442\u0430\u0442\u043e\u0447\u043d\u043e. CPU \u0437\u0434\u0435\u0441\u044c \u0437\u0430\u043d\u0438\u043c\u0430\u0435\u0442\u0441\u044f \u043e\u0440\u043a\u0435\u0441\u0442\u0440\u0430\u0446\u0438\u0435\u0439: \u043d\u0430\u0441\u0442\u0440\u0430\u0438\u0432\u0430\u0435\u0442 DMA, \u043a\u043e\u043d\u0444\u0438\u0433\u0443\u0440\u0438\u0440\u0443\u0435\u0442 ISP \u0438 NPU, \u0437\u0430\u043f\u0443\u0441\u043a\u0430\u0435\u0442 MIPI CSI \u2014 \u0432\u0441\u0435 \u0440\u0435\u0441\u0443\u0440\u0441\u043e\u0451\u043c\u043a\u0438\u0435 \u043e\u043f\u0435\u0440\u0430\u0446\u0438\u0438 \u0432\u044b\u043f\u043e\u043b\u043d\u044f\u044e\u0442\u0441\u044f \u0430\u043f\u043f\u0430\u0440\u0430\u0442\u043d\u043e. \u0415\u0434\u0438\u043d\u0441\u0442\u0432\u0435\u043d\u043d\u043e\u0435 \u043c\u0435\u0441\u0442\u043e, \u0433\u0434\u0435 \u043f\u0440\u043e\u0446\u0435\u0441\u0441\u043e\u0440 \u0440\u0435\u0430\u043b\u044c\u043d\u043e \u043d\u0430\u0433\u0440\u0443\u0436\u0430\u0435\u0442\u0441\u044f \u2014 \u043f\u043e\u0441\u0442\u043f\u0440\u043e\u0446\u0435\u0441\u0441\u0438\u043d\u0433 \u0440\u0435\u0437\u0443\u043b\u044c\u0442\u0430\u0442\u043e\u0432 \u043c\u043e\u0434\u0435\u043b\u0435\u0439 (NMS, confidence filtering). \u041f\u043e\u044d\u0442\u043e\u043c\u0443 \u043d\u0435\u043e\u0431\u0445\u043e\u0434\u0438\u043c\u043e \u043c\u0430\u043a\u0441\u0438\u043c\u0430\u043b\u044c\u043d\u0443\u044e \u0447\u0430\u0441\u0442\u044c \u043f\u043e\u0441\u0442\u043f\u0440\u043e\u0446\u0435\u0441\u0441\u0438\u043d\u0433\u0430 \u043f\u0435\u0440\u0435\u043d\u043e\u0441\u0438\u0442\u044c \u043d\u0430 NPU \u0438\u043b\u0438 DSP, \u0447\u0442\u043e\u0431\u044b CPU \u043d\u0435 \u0441\u0442\u0430\u043b \u0443\u0437\u043a\u0438\u043c \u0433\u043e\u0440\u043b\u044b\u0448\u043a\u043e\u043c \u0432 \u0438\u0442\u043e\u0433\u043e\u0432\u043e\u0439 \u0441\u043a\u043e\u0440\u043e\u0441\u0442\u0438 \u0440\u0430\u0431\u043e\u0442\u044b \u043c\u043e\u0434\u0435\u043b\u0438.\u041f\u043e\u0434\u0441\u0438\u0441\u0442\u0435\u043c\u0430 \u043f\u0430\u043c\u044f\u0442\u0438: \u0434\u0432\u043e\u0439\u043d\u043e\u0439 DDR-\u043a\u043e\u043d\u0442\u0440\u043e\u043b\u043b\u0435\u0440AX650N \u043e\u0441\u043d\u0430\u0449\u0451\u043d \u0434\u0432\u0443\u043c\u044f \u043d\u0435\u0437\u0430\u0432\u0438\u0441\u0438\u043c\u044b\u043c\u0438 32-\u0431\u0438\u0442\u043d\u044b\u043c\u0438 \u043a\u043e\u043d\u0442\u0440\u043e\u043b\u043b\u0435\u0440\u0430\u043c\u0438 LPDDR4\/LPDDR4x \u2014 DDR0 \u0438 DDR1. \u041a\u0430\u0436\u0434\u044b\u0439 \u043a\u043e\u043d\u0442\u0440\u043e\u043b\u043b\u0435\u0440 \u0438\u043c\u0435\u0435\u0442 \u0441\u043e\u0431\u0441\u0442\u0432\u0435\u043d\u043d\u0443\u044e \u0444\u0438\u0437\u0438\u0447\u0435\u0441\u043a\u0443\u044e \u0448\u0438\u043d\u0443 \u0434\u0430\u043d\u043d\u044b\u0445 (32 \u0431\u0438\u0442\u0430), \u0441\u043e\u0431\u0441\u0442\u0432\u0435\u043d\u043d\u044b\u0439 \u043d\u0430\u0431\u043e\u0440 \u0441\u0438\u0433\u043d\u0430\u043b\u043e\u0432 CA, CK, CS, DQS, \u0430 \u0438\u0445 \u0430\u0434\u0440\u0435\u0441\u043d\u044b\u0435 \u043f\u0440\u043e\u0441\u0442\u0440\u0430\u043d\u0441\u0442\u0432\u0430 \u0432 \u0441\u0438\u0441\u0442\u0435\u043c\u043d\u043e\u0439 \u043a\u0430\u0440\u0442\u0435 \u0430\u0434\u0440\u0435\u0441\u043e\u0432 \u0440\u0430\u0437\u043d\u0435\u0441\u0435\u043d\u044b: DDR0-\u0441\u0438\u0441\u0442\u0435\u043c\u0430 \u0431\u0430\u0437\u0438\u0440\u0443\u0435\u0442\u0441\u044f \u043f\u043e \u0430\u0434\u0440\u0435\u0441\u0443 0x8000000, DDR1-\u0441\u0438\u0441\u0442\u0435\u043c\u0430 \u2014 \u043f\u043e 0xC000000.\u0414\u0432\u0430 \u043a\u043e\u043d\u0442\u0440\u043e\u043b\u043b\u0435\u0440\u0430 \u043d\u0435 \u043e\u0431\u0440\u0430\u0437\u0443\u044e\u0442 \u0435\u0434\u0438\u043d\u0443\u044e 64-\u0431\u0438\u0442\u043d\u0443\u044e \u0448\u0438\u043d\u0443 (\u043a\u0430\u043a \u0432 \u043a\u043b\u0430\u0441\u0441\u0438\u0447\u0435\u0441\u043a\u043e\u043c \u201cdual-channel ddr\u201d), \u0430 \u0440\u0430\u0431\u043e\u0442\u0430\u044e\u0442 \u043f\u0430\u0440\u0430\u043b\u043b\u0435\u043b\u044c\u043d\u043e \u0441 \u0440\u0430\u0437\u0434\u0435\u043b\u044c\u043d\u044b\u043c\u0438 \u0430\u0434\u0440\u0435\u0441\u043d\u044b\u043c\u0438 \u043f\u0440\u043e\u0441\u0442\u0440\u0430\u043d\u0441\u0442\u0432\u0430\u043c\u0438. \u042d\u0442\u043e \u043e\u0437\u043d\u0430\u0447\u0430\u0435\u0442, \u0447\u0442\u043e NPU, VDSP, ISP \u0438 CPU \u043c\u043e\u0433\u0443\u0442 \u0438\u043d\u0438\u0446\u0438\u0438\u0440\u043e\u0432\u0430\u0442\u044c \u0442\u0440\u0430\u043d\u0437\u0430\u043a\u0446\u0438\u0438 \u043a \u0440\u0430\u0437\u043d\u044b\u043c \u043a\u043e\u043d\u0442\u0440\u043e\u043b\u043b\u0435\u0440\u0430\u043c \u043e\u0434\u043d\u043e\u0432\u0440\u0435\u043c\u0435\u043d\u043d\u043e, \u043d\u0435 \u0441\u043e\u0437\u0434\u0430\u0432\u0430\u044f \u043a\u043e\u043d\u043a\u0443\u0440\u0435\u043d\u0446\u0438\u0438 \u0437\u0430 \u043e\u0434\u043d\u0443 \u0448\u0438\u043d\u0443. \u041f\u0440\u0438 \u0441\u043a\u043e\u0440\u043e\u0441\u0442\u0438 4266 Mbps \u043a\u0430\u0436\u0434\u044b\u0439 \u043a\u043e\u043d\u0442\u0440\u043e\u043b\u043b\u0435\u0440 \u043e\u0431\u0435\u0441\u043f\u0435\u0447\u0438\u0432\u0430\u0435\u0442 ~17 GB\/s, \u0441\u0443\u043c\u043c\u0430\u0440\u043d\u043e \u2014 \u0434\u043e ~34 GB\/s \u0442\u0435\u043e\u0440\u0435\u0442\u0438\u0447\u0435\u0441\u043a\u043e\u0439 \u043f\u0440\u043e\u043f\u0443\u0441\u043a\u043d\u043e\u0439 \u0441\u043f\u043e\u0441\u043e\u0431\u043d\u043e\u0441\u0442\u0438.\u041c\u0430\u043a\u0441\u0438\u043c\u0430\u043b\u044c\u043d\u044b\u0439 \u043e\u0431\u044a\u0451\u043c: \u0434\u043e 16 GB \u043d\u0430 \u043a\u0430\u043d\u0430\u043b (\u0434\u043e 32 GB \u0441\u0443\u043c\u043c\u0430\u0440\u043d\u043e). \u041f\u0440\u0438 \u044d\u0442\u043e\u043c \u0447\u0430\u0441\u0442\u044c \u043f\u0430\u043c\u044f\u0442\u0438 \u0440\u0435\u0437\u0435\u0440\u0432\u0438\u0440\u0443\u0435\u0442\u0441\u044f \u043a\u0430\u043a CMM (Contiguous Media Memory) \u2014 \u043d\u0435\u043f\u0440\u0435\u0440\u044b\u0432\u043d\u044b\u0439 \u0440\u0435\u0433\u0438\u043e\u043d, \u0434\u043e\u0441\u0442\u0443\u043f\u043d\u044b\u0439 NPU, VPU, ISP \u0438 DSP \u043d\u0430\u043f\u0440\u044f\u043c\u0443\u044e, \u043c\u0438\u043d\u0443\u044f MMU \u044f\u0434\u0440\u0430 Linux. \u041e\u0442\u043d\u043e\u0448\u0435\u043d\u0438\u0435 CMM \u043a Linux-\u043f\u0430\u043c\u044f\u0442\u0438 \u043a\u043e\u043d\u0444\u0438\u0433\u0443\u0440\u0438\u0440\u0443\u0435\u0442\u0441\u044f \u043f\u0440\u043e\u0433\u0440\u0430\u043c\u043c\u043d\u043e. \u041f\u043e \u0443\u043c\u043e\u043b\u0447\u0430\u043d\u0438\u044e \u043d\u0430 Maix4 HAT \u043f\u043e\u0434 \u0441\u0438\u0441\u0442\u0435\u043c\u0443 \u0432\u044b\u0434\u0435\u043b\u044f\u0435\u0442\u0441\u044f 2 GB, \u0430 \u043f\u043e\u0434 CMM \u043e\u0441\u0442\u0430\u0432\u0448\u0438\u0435\u0441\u044f 6 GB.NPU \u2014 Axera Neutron\u0421\u0440\u0430\u0437\u0443 \u0445\u043e\u0447\u0443 \u043e\u0431\u0440\u0430\u0442\u0438\u0442\u044c \u0432\u043d\u0438\u043c\u0430\u043d\u0438\u0435, \u0447\u0442\u043e \u0431\u043e\u043b\u044c\u0448\u0438\u043d\u0441\u0442\u0432\u043e \u041f\u041e \u0434\u043b\u044f NPU (\u043d\u0438\u0437\u043a\u043e\u0443\u0440\u043e\u0432\u043d\u0435\u0432\u044b\u0435 \u0434\u0440\u0430\u0439\u0432\u0435\u0440\u0430 \u0438 SDK \u0434\u043b\u044f \u043f\u043e\u0434\u0433\u043e\u0442\u043e\u0432\u043a\u0438 \u043c\u043e\u0434\u0435\u043b\u0435\u0439) \u044f\u0432\u043b\u044f\u0435\u0442\u0441\u044f \u043f\u0440\u043e\u043f\u0440\u0438\u0435\u0442\u0430\u0440\u043d\u044b\u043c, \u043f\u043e\u044d\u0442\u043e\u043c\u0443 \u0432\u0441\u0451 \u043e\u043f\u0438\u0441\u0430\u043d\u043d\u043e\u0435 \u0434\u0430\u043b\u044c\u0448\u0435 \u2014 \u043b\u043e\u0433\u0438\u0447\u0435\u0441\u043a\u0438\u0435 \u0432\u044b\u0432\u043e\u0434\u044b \u043d\u0430 \u043e\u0441\u043d\u043e\u0432\u0435 \u0434\u043e\u0441\u0442\u0443\u043f\u043d\u043e\u0439 \u0434\u043e\u043a\u0443\u043c\u0435\u043d\u0442\u0430\u0446\u0438\u0438.\u0417\u0430\u044f\u0432\u043b\u0435\u043d\u043d\u044b\u0435 72 TOPS@INT4 \/ 18 TOPS@INT8 \u2014 \u044d\u0442\u043e \u043c\u0430\u0440\u043a\u0435\u0442\u0438\u043d\u0433\u043e\u0432\u0430\u044f \u0441\u0443\u043c\u043c\u0430 \u043f\u043e \u043d\u0435\u0441\u043a\u043e\u043b\u044c\u043a\u0438\u043c \u043f\u043e\u0434\u0441\u0438\u0441\u0442\u0435\u043c\u0430\u043c \u0441\u0440\u0430\u0437\u0443. \u0412 \u0434\u0430\u0442\u0430\u0448\u0438\u0442\u0435 \u043e\u0442\u0434\u0435\u043b\u044c\u043d\u043e \u0434\u043b\u044f NPU \u0443\u043a\u0430\u0437\u0430\u043d\u043e: 43.2 TOPS@INT4 \/ 10.8 TOPS@INT8. \u041a\u0430\u043a \u0438\u043c\u0435\u043d\u043d\u043e \u0431\u044b\u043b\u0438 \u043f\u043e\u043b\u0443\u0447\u0435\u043d\u044b 18 TOPS@INT8 \u043d\u0435\u043f\u043e\u043d\u044f\u0442\u043d\u043e, \u043d\u043e \u043e\u0447\u0435\u0432\u0438\u0434\u043d\u043e, \u0447\u0442\u043e \u0432 \u043d\u0438\u0445 \u0432\u0445\u043e\u0434\u0438\u0442 \u0441\u0443\u043c\u043c\u0430 NPU, DSP, MAU \u0438 \u0434\u0440\u0443\u0433\u0438\u0445 IP-\u0431\u043b\u043e\u043a\u043e\u0432. \u041f\u0440\u0438 \u044d\u0442\u043e\u043c, \u043d\u0430\u043f\u0440\u0438\u043c\u0435\u0440, \u0434\u0432\u0443\u0445\u044a\u044f\u0434\u0435\u0440\u043d\u044b\u0439 DSP \u043d\u0435 \u0443\u0447\u0430\u0441\u0442\u0432\u0443\u0435\u0442 \u0432 \u0438\u043d\u0444\u0435\u0440\u0435\u043d\u0441\u0435 \u043c\u043e\u0434\u0435\u043b\u0435\u0439. \u042d\u0442\u043e \u0431\u044b\u043b\u043e \u043f\u0440\u043e\u0432\u0435\u0440\u0435\u043d\u043e \u043f\u0440\u0438 \u0437\u0430\u043f\u0443\u0449\u0435\u043d\u043d\u043e\u0439 YOLO \u0447\u0435\u0440\u0435\u0437 \u0434\u0435\u0431\u0430\u0433 \u0438\u043d\u0444\u043e\u0440\u043c\u0430\u0446\u0438\u044e \u043e\u0442 \u0434\u0440\u0430\u0439\u0432\u0435\u0440\u0430 DSP \u0432 sysfs (\/proc\/ax_proc\/dspX\/info).NPU \u043f\u043e\u0434\u0434\u0435\u0440\u0436\u0438\u0432\u0430\u0435\u0442 \u0448\u0438\u0440\u043e\u043a\u0438\u0439 \u043d\u0430\u0431\u043e\u0440 \u043a\u0432\u0430\u043d\u0442\u043e\u0432\u0430\u043d\u0438\u044f (\u043f\u0440\u0438\u0447\u0451\u043c \u0438\u043c\u0435\u0435\u0442\u0441\u044f \u043f\u043e\u0434\u0434\u0435\u0440\u0436\u043a\u0430 \u0441\u043c\u0435\u0448\u0430\u043d\u043d\u044b\u0445 \u0442\u0438\u043f\u043e\u0432 \u043c\u0435\u0436\u0434\u0443 \u0441\u043b\u043e\u044f\u043c\u0438): \u0446\u0435\u043b\u043e\u0447\u0438\u0441\u043b\u0435\u043d\u043d\u044b\u0435 INT4\/INT8\/INT16 \u0438 \u0444\u043b\u043e\u0430\u0442\u044b FP16\/FP32. \u0418 \u044d\u0442\u043e \u043d\u0430 \u0441\u0430\u043c\u043e\u043c \u0434\u0435\u043b\u0435 \u043e\u0447\u0435\u043d\u044c \u0432\u0430\u0436\u043d\u043e, \u043f\u043e\u0442\u043e\u043c\u0443 \u0447\u0442\u043e \u0435\u0441\u043b\u0438 \u0442\u0440\u0435\u0431\u0443\u0435\u0442\u0441\u044f \u0434\u0435\u0442\u0435\u043a\u0442\u0438\u0440\u043e\u0432\u0430\u0442\u044c \u043e\u0431\u044a\u0435\u043a\u0442\u044b \u0441 \u0432\u044b\u0441\u043e\u043a\u043e\u0439 \u0442\u043e\u0447\u043d\u043e\u0441\u0442\u044c\u044e \u0433\u0440\u0430\u043d\u0438\u0446, \u0442\u043e \u0432 INT8 \u043a\u0432\u0430\u043d\u0442\u043e\u0432\u0430\u043d\u0438\u0438 YOLO \u043c\u043e\u0434\u0435\u043b\u0438 \u043c\u043e\u0436\u0435\u0442 \u043d\u0435 \u0445\u0432\u0430\u0442\u0430\u0442\u044c.\u0417\u0430\u044f\u0432\u043b\u044f\u0435\u0442\u0441\u044f, \u0447\u0442\u043e INT4 \u0432 4 \u0440\u0430\u0437\u0430 \u0431\u044b\u0441\u0442\u0440\u0435\u0435, \u0447\u0435\u043c INT8 (10.8 vs 43.2). \u041e\u0431\u044b\u0447\u043d\u043e INT4 \u0434\u0430\u0451\u0442 \u0442\u0435\u043e\u0440\u0435\u0442\u0438\u0447\u0435\u0441\u043a\u0438\u0439 \u043f\u0440\u0438\u0440\u043e\u0441\u0442 \u043d\u0435 \u0431\u043e\u043b\u0435\u0435 \u0447\u0435\u043c \u0432 2 \u0440\u0430\u0437\u0430 \u043f\u043e \u0441\u0440\u0430\u0432\u043d\u0435\u043d\u0438\u044e \u0441 INT8 (\u043d\u0430\u043f\u0440\u0438\u043c\u0435\u0440, \u0434\u043b\u044f RK3588). \u0418\u0437 \u044d\u0442\u043e\u0433\u043e \u043c\u043e\u0436\u043d\u043e \u0441\u0434\u0435\u043b\u0430\u0442\u044c \u0432\u044b\u0432\u043e\u0434, \u0447\u0442\u043e \u043f\u043e\u0434\u0434\u0435\u0440\u0436\u043a\u0430 INT4 \u0440\u0435\u0430\u043b\u0438\u0437\u043e\u0432\u0430\u043d\u0430 \u0430\u043f\u043f\u0430\u0440\u0430\u0442\u043d\u043e. \u0421\u043a\u043e\u0440\u0435\u0435 \u0432\u0441\u0435\u0433\u043e \u0434\u0432\u0430 INT4 \u0443\u043f\u0430\u043a\u043e\u0432\u044b\u0432\u0430\u044e\u0442\u0441\u044f \u0432 \u043e\u0434\u043d\u0443 INT8 MAC \u0434\u043e\u0440\u043e\u0436\u043a\u0443, \u043f\u0440\u0438 \u044d\u0442\u043e\u043c \u0432 \u043e\u0442\u043b\u0438\u0447\u0438\u0435 \u043e\u0442 \u0442\u043e\u0433\u043e \u0436\u0435 RK3588 \u0441\u0447\u0438\u0442\u0430\u044e\u0442\u0441\u044f \u043f\u0430\u0440\u0430\u043b\u043b\u0435\u043b\u044c\u043d\u043e (\u0434\u0432\u0430 INT4 \u043f\u0435\u0440\u0435\u043c\u043d\u043e\u0436\u0435\u043d\u0438\u044f \u0437\u0430 \u043e\u0434\u0438\u043d \u0442\u0430\u043a\u0442). \u041a\u0432\u0430\u043d\u0442\u043e\u0432\u0430\u043d\u0438\u0435 \u0432 INT4 \u043e\u0431\u044b\u0447\u043d\u043e \u043d\u0435 \u0438\u0441\u043f\u043e\u043b\u044c\u0437\u0443\u0435\u0442\u0441\u044f \u0434\u043b\u044f \u0441\u0432\u0451\u0440\u0442\u043e\u0447\u043d\u044b\u0445 \u0432\u0438\u0437\u0443\u0430\u043b\u044c\u043d\u044b\u0445 \u043c\u043e\u0434\u0435\u043b\u0435\u0439, \u043d\u043e \u0432\u043e\u0442 \u0432\u0435\u0441\u0430 LLM \u0430\u043a\u0442\u0438\u0432\u043d\u043e \u043a\u043e\u043d\u0432\u0435\u0440\u0442\u0438\u0440\u0443\u044e\u0442 \u0432 INT4.\u041f\u043e \u043a\u0430\u0440\u0442\u0435 \u0430\u0434\u0440\u0435\u0441\u043e\u0432 \u043c\u043e\u0436\u043d\u043e \u0447\u0430\u0441\u0442\u0438\u0447\u043d\u043e \u0432\u043e\u0441\u0441\u0442\u0430\u043d\u043e\u0432\u0438\u0442\u044c \u0442\u043e\u043f\u043e\u043b\u043e\u0433\u0438\u044e \u043f\u043e\u0434\u0441\u0438\u0441\u0442\u0435\u043c\u044b NPU:\u00a0\u0411\u043b\u043e\u043a\u0410\u0434\u0440\u0435\u0441\u041a\u043e\u043b-\u0432\u043e\u041d\u0430\u0437\u043d\u0430\u0447\u0435\u043d\u0438\u0435OCM0x14000000\u20130x14AFFFFF1On-Chip Memory, ~11.5 MBSDMA0x16000000\u20130x162FFFFF3System DMA \u043c\u0435\u0436\u0434\u0443 CMM \u0438 OCMEU0\u2013EU120x16400000\u20130x164CFFFF13\u0418\u0441\u043f\u043e\u043b\u043d\u0438\u0442\u0435\u043b\u044c\u043d\u044b\u0435 \u0431\u043b\u043e\u043a\u0438sync0x164D00001\u0421\u0438\u043d\u0445\u0440\u043e\u043d\u0438\u0437\u0430\u0446\u0438\u044f \u043c\u0435\u0436\u0434\u0443 EUfab0x164E00001\u041c\u0435\u0436\u0441\u043e\u0435\u0434\u0438\u043d\u0438\u0442\u0435\u043b\u044c\u043d\u0430\u044f \u0444\u0430\u0431\u0440\u0438\u043a\u0430fbcdc0x165000001Frame Buffer Compress\/Decompress\u0422\u0440\u0438\u043d\u0430\u0434\u0446\u0430\u0442\u044c EU \u0440\u0430\u0437\u0434\u0435\u043b\u0435\u043d\u044b \u043d\u0430 \u0442\u0440\u0438 \u0433\u0440\u0443\u043f\u043f\u044b vNPU. \u0418\u0437\u043d\u0430\u0447\u0430\u043b\u044c\u043d\u043e \u044f \u043f\u0440\u0435\u0434\u043f\u043e\u043b\u043e\u0436\u0438\u043b \u0440\u0430\u0441\u043f\u0440\u0435\u0434\u0435\u043b\u0435\u043d\u0438\u0435:Convolution Unit (3x) \u2014 \u0432\u044b\u043f\u043e\u043b\u043d\u044f\u0435\u0442 \u0432\u0441\u0435 \u0432\u0438\u0434\u044b \u0441\u0432\u0451\u0440\u0442\u043e\u043a: standard, depthwise, grouped, dilated, transposed convolutionComputer Vision Unit (3x) \u2014 \u043d\u043e\u0440\u043c\u0430\u043b\u0438\u0437\u0430\u0446\u0438\u044f \u0438\u0437\u043e\u0431\u0440\u0430\u0436\u0435\u043d\u0438\u0439, \u0440\u0435\u0441\u0430\u0439\u0437, remap\/warpTensor Unit (3x) \u2014 \u0444\u0443\u043d\u043a\u0446\u0438\u0438 \u0430\u043a\u0442\u0438\u0432\u0430\u0446\u0438\u0438, pooling, elementwise-\u043e\u043f\u0435\u0440\u0430\u0446\u0438\u0438, reduce\u0417\u0430\u0442\u0435\u043c \u0432 \u043a\u0430\u0447\u0435\u0441\u0442\u0432\u0435 \u044d\u043a\u0441\u043f\u0435\u0440\u0438\u043c\u0435\u043d\u0442\u0430 \u044f \u044d\u043a\u0441\u043f\u043e\u0440\u0442\u0438\u0440\u043e\u0432\u0430\u043b \u043e\u0434\u043d\u0443 \u0438 \u0442\u0443 \u0436\u0435 YOLO-\u043c\u043e\u0434\u0435\u043b\u044c \u0447\u0435\u0440\u0435\u0437 Pulsar 2 (\u0438\u043d\u0441\u0442\u0440\u0443\u043c\u0435\u043d\u0442 \u043f\u043e\u0434\u0433\u043e\u0442\u043e\u0432\u043a\u0438 \u043c\u043e\u0434\u0435\u043b\u0435\u0439 \u0434\u043b\u044f Axera NPU) \u043f\u043e\u0434 \u0442\u0440\u0438 \u0440\u0430\u0437\u043d\u044b\u0445 \u0440\u0435\u0436\u0438\u043c\u0430 \u0440\u0430\u0431\u043e\u0442\u044b. Pulsar 2 \u043c\u043e\u0436\u0435\u0442 \u0433\u0435\u043d\u0435\u0440\u0438\u0440\u043e\u0432\u0430\u0442\u044c \u0442\u0440\u0435\u0439\u0441 \u043c\u043e\u0434\u0435\u043b\u0438, \u0438\u0437 \u043a\u043e\u0442\u043e\u0440\u043e\u0433\u043e \u0432\u0438\u0434\u043d\u043e \u0432 \u043a\u0430\u043a\u043e\u043c \u043f\u043e\u0440\u044f\u0434\u043a\u0435 \u0447\u0442\u043e \u0432\u044b\u043f\u043e\u043b\u043d\u044f\u0435\u0442\u0441\u044f \u0438 \u043d\u0430 \u043a\u0430\u043a\u0438\u0445 EU.\u0412 \u0440\u0435\u0436\u0438\u043c\u0435 NPU1 (\u043e\u0434\u043d\u043e \u044f\u0434\u0440\u043e \u0438\u0437 \u0442\u0440\u0451\u0445) \u0442\u0440\u0435\u0439\u0441 \u0434\u0435\u043c\u043e\u043d\u0441\u0442\u0440\u0438\u0440\u0443\u0435\u0442 \u043c\u0438\u043d\u0438\u043c\u0430\u043b\u044c\u043d\u044b\u0439 \u043d\u0430\u0431\u043e\u0440 \u0430\u043a\u0442\u0438\u0432\u043d\u044b\u0445 \u0442\u0440\u0435\u043a\u043e\u0432: \u043e\u0434\u0438\u043d teng, \u0434\u0432\u0430 conv, \u043e\u0434\u0438\u043d sdma \u0438 \u043e\u0442\u0434\u0435\u043b\u044c\u043d\u044b\u0439 cv.\u0412 \u0440\u0435\u0436\u0438\u043c\u0435 NPU2 (Big-Little, \u0434\u0432\u0430 \u0432\u0438\u0440\u0442\u0443\u0430\u043b\u044c\u043d\u044b\u0445 \u044f\u0434\u0440\u0430) \u043a\u043e\u043b\u0438\u0447\u0435\u0441\u0442\u0432\u043e \u0430\u043a\u0442\u0438\u0432\u043d\u044b\u0445 \u0442\u0440\u0435\u043a\u043e\u0432 \u0443\u0434\u0432\u0430\u0438\u0432\u0430\u0435\u0442\u0441\u044f \u0441\u0438\u043c\u043c\u0435\u0442\u0440\u0438\u0447\u043d\u043e: \u043f\u043e\u044f\u0432\u043b\u044f\u044e\u0442\u0441\u044f teng4\/teng5, conv0-3, sdma8\/sdma9. \u041f\u0430\u0442\u0442\u0435\u0440\u043d \u043f\u043e\u0434\u0442\u0432\u0435\u0440\u0436\u0434\u0430\u0435\u0442, \u0447\u0442\u043e \u0432\u0442\u043e\u0440\u043e\u0439 vNPU \u0434\u043e\u0431\u0430\u0432\u043b\u044f\u0435\u0442 \u0430\u043d\u0430\u043b\u043e\u0433\u0438\u0447\u043d\u044b\u0439 \u043a\u043e\u043c\u043f\u043b\u0435\u043a\u0442 EU, \u0430 \u043d\u0435 \u00ab\u0434\u043e\u0441\u0442\u0440\u0430\u0438\u0432\u0430\u0435\u0442\u00bb \u043f\u0443\u043b \u043a\u0430\u043a\u043e\u0433\u043e-\u0442\u043e \u043e\u0434\u043d\u043e\u0433\u043e \u0442\u0438\u043f\u0430.\u0412 \u0440\u0435\u0436\u0438\u043c\u0435 NPU3 (\u0432\u0441\u0435 \u0442\u0440\u0438 \u044f\u0434\u0440\u0430 \u043e\u0431\u044a\u0435\u0434\u0438\u043d\u0435\u043d\u044b) \u0442\u0440\u0435\u0439\u0441 \u043d\u0430\u0438\u0431\u043e\u043b\u0435\u0435 \u043f\u043e\u043b\u043d\u044b\u0439: teng6\/teng7\/teng8, conv0-5, sdma12-14.\u00a0\u0418\u0437 \u044d\u0442\u043e\u0433\u043e \u043c\u043e\u0436\u043d\u043e \u0441\u0434\u0435\u043b\u0430\u0442\u044c \u0432\u044b\u0432\u043e\u0434, \u0447\u0442\u043e \u043a\u0430\u0436\u0434\u044b\u0439 vNPU \u0441\u043e\u0434\u0435\u0440\u0436\u0438\u0442 2 Convolution EU + 1 Tensor EU + 1 SDMA, \u0430 \u0432\u0441\u0435\u0433\u043e Conv EU \u2014\u00a0 6, \u0430 \u043d\u0435 3, \u043a\u0430\u043a \u044f \u043f\u0440\u0435\u0434\u043f\u043e\u043b\u043e\u0436\u0438\u043b \u0438\u0437\u043d\u0430\u0447\u0430\u043b\u044c\u043d\u043e. \u0422\u0430\u043a\u0436\u0435 \u043c\u043e\u0436\u043d\u043e \u0437\u0430\u043c\u0435\u0442\u0438\u0442\u044c, \u0447\u0442\u043e \u0432\u0441\u0435 \u0432\u044b\u0447\u0438\u0441\u043b\u0435\u043d\u0438\u044f \u043c\u043e\u0434\u0435\u043b\u0438 \u0440\u0430\u0437\u0434\u0435\u043b\u044f\u044e\u0442\u0441\u044f \u043c\u0435\u0436\u0434\u0443 NPU EU  \u2014 \u0437\u0430\u0434\u0430\u0447 \u0434\u043b\u044f MAU \u0438\u043b\u0438 DSP \u043d\u0435\u0442. \u042d\u0442\u043e \u0435\u0449\u0451 \u0440\u0430\u0437 \u043f\u043e\u0434\u0442\u0432\u0435\u0440\u0436\u0434\u0430\u0435\u0442, \u0447\u0442\u043e \u043d\u0430 \u0438\u043d\u0444\u0435\u0440\u0435\u043d\u0441 \u0441\u0442\u0430\u043d\u0434\u0430\u0440\u0442\u043d\u044b\u043c\u0438 \u0441\u0440\u0435\u0434\u0441\u0442\u0432\u0430\u043c\u0438 \u0432\u044b\u0434\u0435\u043b\u044f\u0435\u0442\u0441\u044f \u043c\u0430\u043a\u0441\u0438\u043c\u0443\u043c 10.8 TOPS.\u0415\u0441\u043b\u0438 \u043f\u043e\u0441\u0447\u0438\u0442\u0430\u0442\u044c \u043a\u043e\u043b\u0438\u0447\u0435\u0441\u0442\u0432\u043e EU \u0434\u043b\u044f NPU3 \u0440\u0435\u0436\u0438\u043c\u0430, \u0442\u043e \u0438\u0445 \u043f\u043e\u043b\u0443\u0447\u0438\u0442\u0441\u044f 12, \u0445\u043e\u0442\u044f \u043f\u043e \u043a\u0430\u0440\u0442\u0435 \u0430\u0434\u0440\u0435\u0441\u043e\u0432 \u0438\u0445 \u0434\u043e\u043b\u0436\u043d\u043e \u0431\u044b\u0442\u044c 13. \u0421\u043d\u0430\u0447\u0430\u043b\u0430 \u044f \u043f\u0440\u0435\u0434\u043f\u043e\u043b\u043e\u0436\u0438\u043b, \u0447\u0442\u043e \u043e\u0434\u0438\u043d \u043e\u0441\u0442\u0430\u0432\u0448\u0438\u0439\u0441\u044f EU \u2014 \u044d\u0442\u043e MAU (Matrix Operation Unit), \u043e\u0442\u0434\u0435\u043b\u044c\u043d\u044b\u0439 \u0431\u043b\u043e\u043a \u0434\u043b\u044f \u043f\u0435\u0440\u0435\u043c\u043d\u043e\u0436\u0435\u043d\u0438\u044f \u043c\u0430\u0442\u0440\u0438\u0446. \u0412 \u0442\u0440\u0435\u0439\u0441\u0435 \u0435\u0433\u043e \u043d\u0435\u0442, \u043f\u043e\u0442\u043e\u043c\u0443 \u0447\u0442\u043e \u0434\u043b\u044f YOLO \u043e\u043d \u0432 \u0446\u0435\u043b\u043e\u043c \u0438 \u043d\u0435 \u043d\u0443\u0436\u0435\u043d, \u0432\u0441\u0451 \u043b\u043e\u0436\u0438\u0442\u0441\u044f \u043d\u0430 Conv \u0438 Tensor EU. \u0414\u0430\u0442\u0430\u0448\u0438\u0442 \u043a\u043e\u0441\u0432\u0435\u043d\u043d\u043e \u044d\u0442\u043e \u043f\u043e\u0434\u0442\u0432\u0435\u0440\u0436\u0434\u0430\u0435\u0442, \u043f\u043e\u0442\u043e\u043c\u0443 \u0447\u0442\u043e MAU \u043e\u043f\u0438\u0441\u0430\u043d \u0432 \u0447\u0430\u0441\u0442\u0438, \u043f\u043e\u0441\u0432\u044f\u0449\u0435\u043d\u043d\u043e\u0439 NPU:\u0422\u043e\u0447\u043d\u043e\u0433\u043e \u043a\u043e\u043b\u0438\u0447\u0435\u0441\u0442\u0432\u0430 MAC-\u0431\u043b\u043e\u043a\u043e\u0432 \u0432 \u043a\u0430\u0436\u0434\u043e\u043c EU \u043d\u0435 \u0443\u043a\u0430\u0437\u0430\u043d\u043e, \u043d\u043e \u0437\u043d\u0430\u044f \u0447\u0442\u043e \u043e\u0431\u0449\u0430\u044f \u043f\u0438\u043a\u043e\u0432\u0430\u044f \u043f\u0440\u043e\u0438\u0437\u0432\u043e\u0434\u0438\u0442\u0435\u043b\u044c\u043d\u043e\u0441\u0442\u044c 10.8 TOPS \u043c\u043e\u0436\u043d\u043e \u043d\u0430\u0439\u0442\u0438 \u043a\u043e\u043b\u0438\u0447\u0435\u0441\u0442\u0432\u043e MAC \u0432 \u043a\u0430\u0436\u0434\u043e\u043c EU, \u0435\u0441\u043b\u0438 \u043f\u0440\u0435\u0434\u043f\u043e\u043b\u043e\u0436\u0438\u0442\u044c, \u0447\u0442\u043e \u043e\u043d\u043e \u043e\u0434\u0438\u043d\u0430\u043a\u043e\u0432\u043e \u0434\u043b\u044f \u043a\u0430\u0436\u0434\u043e\u0433\u043e. \u0412 \u0442\u0430\u043a\u043e\u043c \u0441\u043b\u0443\u0447\u0430\u0435 \u043f\u043e\u043b\u0443\u0447\u0430\u0435\u0442\u0441\u044f:\u0422\u043e\u0433\u0434\u0430 \u043f\u0440\u043e\u0438\u0437\u0432\u043e\u0434\u0438\u0442\u0435\u043b\u044c\u043d\u043e\u0441\u0442\u044c \u043a\u0430\u0436\u0434\u043e\u0433\u043e EU \u043d\u0435\u043c\u043d\u043e\u0433\u043e \u0431\u043e\u043b\u044c\u0448\u0435 0.8 TOPS. \u0412\u043e\u0437\u043c\u043e\u0436\u043d\u043e \u0432 \u043a\u0430\u0436\u0434\u043e\u043c EU \u0431\u043e\u043b\u0435\u0435 \u043a\u0440\u0443\u0433\u043b\u043e\u0435 \u0437\u043d\u0430\u0447\u0435\u043d\u0438\u0435 MAC, \u0430 \u0438\u043c\u0435\u043d\u043d\u043e \u043f\u043e 512 MAC, \u043d\u043e \u043e\u043d\u0438 \u043d\u0435\u0440\u0430\u0432\u043d\u043e\u043c\u0435\u0440\u043d\u043e \u0440\u0430\u0441\u043f\u0440\u0435\u0434\u0435\u043b\u0435\u043d\u044b \u0438 \u0432 \u043a\u0430\u043a\u0438\u0445-\u0442\u043e \u043e\u0442\u0434\u0435\u043b\u044c\u043d\u044b\u0445 EU \u0431\u043e\u043b\u044c\u0448\u0435 MAC. \u041b\u0438\u0431\u043e \u0447\u0430\u0441\u0442\u043e\u0442\u0430 NPU \u043d\u0435\u043c\u043d\u043e\u0433\u043e \u0432\u044b\u0448\u0435, \u0447\u0435\u043c 800 MHz (\u043d\u0430\u043f\u0440\u0438\u043c\u0435\u0440, 810 MHz \u0434\u0430\u0451\u0442 \u043f\u0440\u0430\u043a\u0442\u0438\u0447\u0435\u0441\u043a\u0438 10.8 TOPS). \u0417\u043d\u0430\u0447\u0435\u043d\u0438\u0435 \u0432 800 MHz \u044f \u0432\u0437\u044f\u043b \u0438\u0437 \u0434\u0435\u0431\u0430\u0433\u0430 sysfs NPU \u0434\u0440\u0430\u0439\u0432\u0435\u0440\u0430 (\u043e\u043d\u0430 \u0432\u0441\u0435\u0433\u0434\u0430 \u0444\u0438\u043a\u0441\u0438\u0440\u043e\u0432\u0430\u043d\u043d\u0430\u044f \u0438 \u0443\u0432\u0435\u043b\u0438\u0447\u0438\u0442\u044c\/\u0443\u043c\u0435\u043d\u044c\u0448\u0438\u0442\u044c \u0435\u0451 \u043d\u0435\u043b\u044c\u0437\u044f):cat \/proc\/ax_proc\/npu\/clk800M\u041d\u043e \u043f\u043e\u0442\u043e\u043c \u044f \u043d\u0430\u0448\u0451\u043b \u0440\u0435\u0444\u0435\u0440\u0435\u043d\u0441 \u043c\u0430\u043d\u0443\u0430\u043b \u043d\u0430 IVE (Intelligent Video Engine \u2014 \u043e\u0442\u0434\u0435\u043b\u044c\u043d\u0430\u044f \u043f\u043e\u0434\u0441\u0438\u0441\u0442\u0435\u043c\u0430, \u0434\u043e\u043a\u0443\u043c\u0435\u043d\u0442 42, AX IVE API), \u0432 \u043a\u043e\u0442\u043e\u0440\u043e\u043c \u0431\u044b\u043b\u0430 \u043e\u043f\u0438\u0441\u0430\u043d\u0430 \u043f\u0440\u043e\u0438\u0437\u0432\u043e\u0434\u0438\u0442\u0435\u043b\u044c\u043d\u043e\u0441\u0442\u044c MAU:25.6 GOPS \u2014 \u044d\u0442\u043e \u043f\u0440\u0438\u043c\u0435\u0440\u043d\u043e \u0432 30 \u0440\u0430\u0437 \u043c\u0435\u043d\u044c\u0448\u0435, \u043c\u043e\u0449\u043d\u043e\u0441\u0442\u0438 \u043e\u0434\u043d\u043e\u0433\u043e EU (~800 GOPS). \u041f\u043e\u043b\u0443\u0447\u0430\u0435\u0442\u0441\u044f, MAU \u2014 \u044d\u0442\u043e \u043d\u0435 \u043f\u043e\u043b\u043d\u043e\u0446\u0435\u043d\u043d\u044b\u0439 EU, \u0430 \u043e\u0442\u0434\u0435\u043b\u044c\u043d\u044b\u0439 \u043d\u0435\u0431\u043e\u043b\u044c\u0448\u043e\u0439 \u0431\u043b\u043e\u043a \u0434\u043b\u044f \u043c\u0430\u0442\u0440\u0438\u0447\u043d\u044b\u0445 \u043e\u043f\u0435\u0440\u0430\u0446\u0438\u0439. \u042f \u043f\u043e\u043a\u0430 \u043d\u0435 \u043d\u0430\u0448\u0451\u043b \u0441\u043f\u043e\u0441\u043e\u0431\u0430 \u0438\u0437\u043c\u0435\u0440\u0438\u0442\u044c \u043d\u0430\u0433\u0440\u0443\u0437\u043a\u0443 \u043d\u0430 \u043a\u0430\u0436\u0434\u044b\u0439 \u0438\u0437 EU \u043f\u043e \u043e\u0442\u0434\u0435\u043b\u044c\u043d\u043e\u0441\u0442\u0438 (\u0434\u0440\u0430\u0439\u0432\u0435\u0440 \u0434\u0430\u0451\u0442 \u0442\u043e\u043b\u044c\u043a\u043e \u043e\u0446\u0435\u043d\u043a\u0443 \u0441\u0443\u043c\u043c\u0430\u0440\u043d\u043e\u0439 \u0443\u0442\u0438\u043b\u0438\u0437\u0430\u0446\u0438\u0438 \u0432\u0441\u0435\u0445 EU). \u041f\u043e\u044d\u0442\u043e\u043c\u0443 13-\u0439 EU, \u0441\u043a\u043e\u0440\u0435\u0435 \u0432\u0441\u0435\u0433\u043e, \u043b\u0438\u0431\u043e \u044f\u0432\u043b\u044f\u0435\u0442\u0441\u044f \u0447\u0430\u0441\u0442\u044c\u044e \u043e\u0434\u043d\u043e\u0433\u043e \u0438\u0437 Conv\/teng-\u0431\u043b\u043e\u043a\u043e\u0432, \u043b\u0438\u0431\u043e Pulsar 2 \u043f\u0440\u043e\u0441\u0442\u043e \u043d\u0435 \u0437\u0430\u0434\u0435\u0439\u0441\u0442\u0432\u043e\u0432\u0430\u043b \u0435\u0433\u043e \u0434\u043b\u044f YOLO.\u00a0\u0422\u0435\u043f\u0435\u0440\u044c \u043e\u0431\u0441\u0443\u0434\u0438\u043c \u0432\u043e\u0437\u043c\u043e\u0436\u043d\u043e\u0441\u0442\u0438 EU \u0431\u043b\u043e\u043a\u043e\u0432. \u0412\u043e-\u043f\u0435\u0440\u0432\u044b\u0445 \u044d\u0442\u043e \u043d\u0435 \u0443\u043d\u0438\u0432\u0435\u0440\u0441\u0430\u043b\u044c\u043d\u044b\u0435 \u0432\u044b\u0447\u0438\u0441\u043b\u0438\u0442\u0435\u043b\u044c\u043d\u044b\u0435 \u0431\u043b\u043e\u043a\u0438, \u043f\u043e\u0434\u043e\u0431\u043d\u043e CUDA. \u042d\u0442\u043e \u0441\u043f\u0435\u0446\u0438\u0444\u0438\u0447\u043d\u044b\u0435 \u0431\u043b\u043e\u043a\u0438, \u0440\u0435\u0430\u043b\u0438\u0437\u0443\u044e\u0449\u0438\u0435&#8230;<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[],"tags":[],"class_list":["post-479915","post","type-post","status-publish","format-standard","hentry"],"_links":{"self":[{"href":"https:\/\/savepearlharbor.com\/index.php?rest_route=\/wp\/v2\/posts\/479915","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/savepearlharbor.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/savepearlharbor.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/savepearlharbor.com\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/savepearlharbor.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=479915"}],"version-history":[{"count":0,"href":"https:\/\/savepearlharbor.com\/index.php?rest_route=\/wp\/v2\/posts\/479915\/revisions"}],"wp:attachment":[{"href":"https:\/\/savepearlharbor.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=479915"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/savepearlharbor.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=479915"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/savepearlharbor.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=479915"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}