{"id":349067,"date":"2023-06-20T09:04:06","date_gmt":"2023-06-20T09:04:06","guid":{"rendered":"http:\/\/savepearlharbor.com\/?p=349067"},"modified":"-0001-11-30T00:00:00","modified_gmt":"-0001-11-29T21:00:00","slug":"","status":"publish","type":"post","link":"https:\/\/savepearlharbor.com\/?p=349067","title":{"rendered":"\u041f\u043e\u0434\u0445\u043e\u0434\u0438\u0442 \u043b\u0438 Nvidia RTX A4000 ADA \u0434\u043b\u044f \u043c\u0430\u0448\u0438\u043d\u043d\u043e\u0433\u043e \u043e\u0431\u0443\u0447\u0435\u043d\u0438\u044f?<\/span>"},"content":{"rendered":"
<\/div>\n
\n
\n
\n
\n
<\/figure>\n

\u0412 \u0430\u043f\u0440\u0435\u043b\u0435 \u043a\u043e\u043c\u043f\u0430\u043d\u0438\u044f NVIDIA \u0432\u044b\u043f\u0443\u0441\u0442\u0438\u043b\u0430 \u043d\u0430 \u0440\u044b\u043d\u043e\u043a \u043d\u043e\u0432\u044b\u0439 \u043f\u0440\u043e\u0434\u0443\u043a\u0442 \u2014 \u0433\u0440\u0430\u0444\u0438\u0447\u0435\u0441\u043a\u0438\u0439 \u043f\u0440\u043e\u0446\u0435\u0441\u0441\u043e\u0440 \u043c\u0430\u043b\u043e\u0433\u043e \u0444\u043e\u0440\u043c-\u0444\u0430\u043a\u0442\u043e\u0440\u0430 RTX A4000 ADA, \u043f\u0440\u0435\u0434\u043d\u0430\u0437\u043d\u0430\u0447\u0435\u043d\u043d\u044b\u0439 \u0434\u043b\u044f \u043f\u0440\u0438\u043c\u0435\u043d\u0435\u043d\u0438\u044f \u0432 \u0440\u0430\u0431\u043e\u0447\u0438\u0445 \u0441\u0442\u0430\u043d\u0446\u0438\u044f\u0445. \u042d\u0442\u043e\u0442 \u043f\u0440\u043e\u0446\u0435\u0441\u0441\u043e\u0440 \u043f\u0440\u0438\u0448\u0435\u043b \u043d\u0430 \u0441\u043c\u0435\u043d\u0443 A2000 \u0438 \u043c\u043e\u0436\u0435\u0442 \u0431\u044b\u0442\u044c \u0438\u0441\u043f\u043e\u043b\u044c\u0437\u043e\u0432\u0430\u043d \u0434\u043b\u044f \u0432\u044b\u043f\u043e\u043b\u043d\u0435\u043d\u0438\u044f \u0441\u043b\u043e\u0436\u043d\u044b\u0445 \u0437\u0430\u0434\u0430\u0447, \u0432 \u0442\u043e\u043c \u0447\u0438\u0441\u043b\u0435 \u0434\u043b\u044f \u043d\u0430\u0443\u0447\u043d\u043e-\u0438\u0441\u0441\u043b\u0435\u0434\u043e\u0432\u0430\u0442\u0435\u043b\u044c\u0441\u043a\u0438\u0445 \u0438 \u0438\u043d\u0436\u0435\u043d\u0435\u0440\u043d\u044b\u0445 \u0440\u0430\u0441\u0447\u0435\u0442\u043e\u0432 \u0438 \u0434\u043b\u044f \u0432\u0438\u0437\u0443\u0430\u043b\u0438\u0437\u0430\u0446\u0438\u0438 \u0434\u0430\u043d\u043d\u044b\u0445.<\/p>\n

RTX A4000 ADA \u043e\u0441\u043d\u0430\u0449\u0435\u043d\u0430 6144 \u044f\u0434\u0440\u0430\u043c\u0438 CUDA, 192 \u0442\u0435\u043d\u0437\u043e\u0440\u0430\u043c\u0438 \u0438 48 \u044f\u0434\u0440\u0430\u043c\u0438 RT, \u043e\u043f\u0435\u0440\u0430\u0442\u0438\u0432\u043d\u043e\u0439 \u043f\u0430\u043c\u044f\u0442\u044c\u044e GDDR6 ECC VRAM \u043e\u0431\u044a\u0435\u043c\u043e\u043c 20 \u0413\u0431. \u041e\u0434\u043d\u043e \u0438\u0437 \u043a\u043b\u044e\u0447\u0435\u0432\u044b\u0445 \u043f\u0440\u0435\u0438\u043c\u0443\u0449\u0435\u0441\u0442\u0432 \u043d\u043e\u0432\u043e\u0433\u043e \u0433\u0440\u0430\u0444\u0438\u0447\u0435\u0441\u043a\u043e\u0433\u043e \u043f\u0440\u043e\u0446\u0435\u0441\u0441\u043e\u0440\u0430 \u2014 \u0435\u0433\u043e \u044d\u043d\u0435\u0440\u0433\u043e\u044d\u0444\u0444\u0435\u043a\u0442\u0438\u0432\u043d\u043e\u0441\u0442\u044c: RTX A4000 ADA \u043f\u043e\u0442\u0440\u0435\u0431\u043b\u044f\u0435\u0442 \u0432\u0441\u0435\u0433\u043e 70 \u0412\u0442, \u0447\u0442\u043e \u0441\u043d\u0438\u0436\u0430\u0435\u0442 \u0437\u0430\u0442\u0440\u0430\u0442\u044b \u043d\u0430 \u044d\u043b\u0435\u043a\u0442\u0440\u043e\u044d\u043d\u0435\u0440\u0433\u0438\u044e \u0438 \u0443\u043c\u0435\u043d\u044c\u0448\u0430\u0435\u0442 \u0442\u0435\u043f\u043b\u043e\u0432\u044b\u0434\u0435\u043b\u0435\u043d\u0438\u0435 \u0432 \u0441\u0438\u0441\u0442\u0435\u043c\u0435. \u0413\u0440\u0430\u0444\u0438\u0447\u0435\u0441\u043a\u0438\u0439 \u043f\u0440\u043e\u0446\u0435\u0441\u0441\u043e\u0440 \u0442\u0430\u043a\u0436\u0435 \u043f\u043e\u0437\u0432\u043e\u043b\u044f\u0435\u0442 \u0443\u043f\u0440\u0430\u0432\u043b\u044f\u0442\u044c \u043d\u0435\u0441\u043a\u043e\u043b\u044c\u043a\u0438\u043c\u0438 \u0434\u0438\u0441\u043f\u043b\u0435\u044f\u043c\u0438 \u0431\u043b\u0430\u0433\u043e\u0434\u0430\u0440\u044f \u043f\u043e\u0434\u043a\u043b\u044e\u0447\u0435\u043d\u0438\u044e 4x Mini-DisplayPort 1.4a.<\/p>\n

<\/figure>\n

\u041f\u0440\u0438 \u0441\u0440\u0430\u0432\u043d\u0435\u043d\u0438\u0438 \u0433\u0440\u0430\u0444\u0438\u0447\u0435\u0441\u043a\u0438\u0445 \u043f\u0440\u043e\u0446\u0435\u0441\u0441\u043e\u0440\u043e\u0432 RTX 4000 SFF Ada \u0441 \u0434\u0440\u0443\u0433\u0438\u043c\u0438 \u0443\u0441\u0442\u0440\u043e\u0439\u0441\u0442\u0432\u0430\u043c\u0438 \u0442\u043e\u0433\u043e \u0436\u0435 \u043a\u043b\u0430\u0441\u0441\u0430 \u043c\u043e\u0436\u043d\u043e \u043e\u0442\u043c\u0435\u0442\u0438\u0442\u044c, \u0447\u0442\u043e \u043f\u0440\u0438 \u0440\u0430\u0431\u043e\u0442\u0435 \u0432 \u0440\u0435\u0436\u0438\u043c\u0435 \u043e\u0434\u0438\u043d\u0430\u0440\u043d\u043e\u0439 \u0442\u043e\u0447\u043d\u043e\u0441\u0442\u0438 \u0434\u0430\u043d\u043d\u044b\u0439 \u043f\u0440\u043e\u0434\u0443\u043a\u0442 \u043f\u043e\u043a\u0430\u0437\u044b\u0432\u0430\u0435\u0442 \u043f\u0440\u043e\u0438\u0437\u0432\u043e\u0434\u0438\u0442\u0435\u043b\u044c\u043d\u043e\u0441\u0442\u044c, \u0430\u043d\u0430\u043b\u043e\u0433\u0438\u0447\u043d\u0443\u044e \u043f\u043e\u0441\u043b\u0435\u0434\u043d\u0435\u043c\u0443 \u043f\u043e\u043a\u043e\u043b\u0435\u043d\u0438\u044e \u0433\u0440\u0430\u0444\u0438\u0447\u0435\u0441\u043a\u043e\u0433\u043e \u043f\u0440\u043e\u0446\u0435\u0441\u0441\u043e\u0440\u0430 RTX A4000, \u043a\u043e\u0442\u043e\u0440\u044b\u0439 \u043f\u043e\u0442\u0440\u0435\u0431\u043b\u044f\u0435\u0442 \u0432\u0434\u0432\u043e\u0435 \u0431\u043e\u043b\u044c\u0448\u0435 \u044d\u043d\u0435\u0440\u0433\u0438\u0438 (140 \u0412\u0442 \u043f\u0440\u043e\u0442\u0438\u0432 70 \u0412\u0442).\u00a0<\/p>\n

<\/figure>\n

RTX 4000 SFF Ada \u043f\u043e\u0441\u0442\u0440\u043e\u0435\u043d\u0430 \u043d\u0430 \u0430\u0440\u0445\u0438\u0442\u0435\u043a\u0442\u0443\u0440\u0435 Ada Lovelace \u0438 \u0442\u0435\u0445\u043f\u0440\u043e\u0446\u0435\u0441\u0441\u0435 5 \u043d\u043c. \u042d\u0442\u043e \u043f\u043e\u0437\u0432\u043e\u043b\u044f\u0435\u0442 \u0438\u0441\u043f\u043e\u043b\u044c\u0437\u043e\u0432\u0430\u0442\u044c \u044f\u0434\u0440\u0430 Tensor Core \u043d\u043e\u0432\u043e\u0433\u043e \u043f\u043e\u043a\u043e\u043b\u0435\u043d\u0438\u044f \u0438 \u044f\u0434\u0440\u0430 \u0442\u0440\u0430\u0441\u0441\u0438\u0440\u043e\u0432\u043a\u0438 \u043b\u0443\u0447\u0435\u0439, \u043a\u043e\u0442\u043e\u0440\u044b\u0435 \u0437\u043d\u0430\u0447\u0438\u0442\u0435\u043b\u044c\u043d\u043e \u043f\u043e\u0432\u044b\u0448\u0430\u044e\u0442 \u043f\u0440\u043e\u0438\u0437\u0432\u043e\u0434\u0438\u0442\u0435\u043b\u044c\u043d\u043e\u0441\u0442\u044c, \u043e\u0431\u0435\u0441\u043f\u0435\u0447\u0438\u0432\u0430\u044f \u0431\u043e\u043b\u0435\u0435 \u0431\u044b\u0441\u0442\u0440\u0443\u044e \u0438 \u044d\u0444\u0444\u0435\u043a\u0442\u0438\u0432\u043d\u0443\u044e \u0440\u0430\u0431\u043e\u0442\u0443 \u0441 \u0442\u0440\u0430\u0441\u0441\u0438\u0440\u043e\u0432\u043a\u043e\u0439 \u043b\u0443\u0447\u0435\u0439 \u0438 \u0442\u0435\u043d\u0437\u043e\u0440\u043d\u044b\u043c\u0438 \u044f\u0434\u0440\u0430\u043c\u0438, \u0447\u0435\u043c RTX A4000. \u041a\u0440\u043e\u043c\u0435 \u0442\u043e\u0433\u043e, RTX 4000 SFF Ada \u0443\u043f\u0430\u043a\u043e\u0432\u0430\u043d \u0432 \u043d\u0435\u0431\u043e\u043b\u044c\u0448\u043e\u0439 \u043a\u043e\u0440\u043f\u0443\u0441 \u2014 \u0434\u043b\u0438\u043d\u0430 \u043a\u0430\u0440\u0442\u044b 168 \u043c\u043c, \u0442\u043e\u043b\u0449\u0438\u043d\u0430 \u0440\u0430\u0432\u043d\u0430 \u0434\u0432\u0443\u043c \u0441\u043b\u043e\u0442\u0430\u043c \u0440\u0430\u0441\u0448\u0438\u0440\u0435\u043d\u0438\u044f.<\/p>\n

<\/figure>\n

\u0423\u043b\u0443\u0447\u0448\u0435\u043d\u0438\u0435 \u044f\u0434\u0435\u0440 \u0442\u0440\u0430\u0441\u0441\u0438\u0440\u043e\u0432\u043a\u0438 \u043b\u0443\u0447\u0435\u0439 \u043e\u0431\u0435\u0441\u043f\u0435\u0447\u0438\u0432\u0430\u0435\u0442 \u044d\u0444\u0444\u0435\u043a\u0442\u0438\u0432\u043d\u0443\u044e \u0440\u0430\u0431\u043e\u0442\u0443 \u0432 \u0441\u0440\u0435\u0434\u0430\u0445, \u0433\u0434\u0435 \u0438\u0441\u043f\u043e\u043b\u044c\u0437\u0443\u0435\u0442\u0441\u044f \u044d\u0442\u0430 \u0442\u0435\u0445\u043d\u043e\u043b\u043e\u0433\u0438\u044f, \u0442\u0430\u043a\u0438\u0445 \u043a\u0430\u043a 3D-\u0434\u0438\u0437\u0430\u0439\u043d \u0438 \u0440\u0435\u043d\u0434\u0435\u0440\u0438\u043d\u0433. \u041e\u0431\u044a\u0435\u043c \u043f\u0430\u043c\u044f\u0442\u0438 \u043d\u043e\u0432\u043e\u0433\u043e GPU (20 \u0413\u0431) \u043f\u043e\u0437\u0432\u043e\u043b\u044f\u0435\u0442 \u0441\u043f\u0440\u0430\u0432\u043b\u044f\u0442\u044c\u0441\u044f \u0441 \u0431\u043e\u043b\u044c\u0448\u0438\u043c\u0438 \u0441\u0440\u0435\u0434\u0430\u043c\u0438.<\/p>\n

<\/figure>\n

\u0421\u043e\u0433\u043b\u0430\u0441\u043d\u043e \u0437\u0430\u044f\u0432\u043b\u0435\u043d\u0438\u044f\u043c \u043f\u0440\u043e\u0438\u0437\u0432\u043e\u0434\u0438\u0442\u0435\u043b\u044f<\/u><\/a>, \u0442\u0435\u043d\u0437\u043e\u0440\u043d\u044b\u0435 \u044f\u0434\u0440\u0430 \u0447\u0435\u0442\u0432\u0435\u0440\u0442\u043e\u0433\u043e \u043f\u043e\u043a\u043e\u043b\u0435\u043d\u0438\u044f \u043e\u0431\u0435\u0441\u043f\u0435\u0447\u0438\u0432\u0430\u044e\u0442 \u0432\u044b\u0441\u043e\u043a\u0443\u044e \u043f\u0440\u043e\u0438\u0437\u0432\u043e\u0434\u0438\u0442\u0435\u043b\u044c\u043d\u043e\u0441\u0442\u044c \u0432\u044b\u0447\u0438\u0441\u043b\u0435\u043d\u0438\u0439 \u0418\u0418 \u2014 \u0434\u0432\u0443\u043a\u0440\u0430\u0442\u043d\u043e\u0435 \u0443\u0432\u0435\u043b\u0438\u0447\u0435\u043d\u0438\u0435 \u043f\u0440\u043e\u0438\u0437\u0432\u043e\u0434\u0438\u0442\u0435\u043b\u044c\u043d\u043e\u0441\u0442\u0438 \u043f\u043e \u0441\u0440\u0430\u0432\u043d\u0435\u043d\u0438\u044e \u0441 \u043f\u0440\u0435\u0434\u044b\u0434\u0443\u0449\u0438\u043c \u043f\u043e\u043a\u043e\u043b\u0435\u043d\u0438\u0435\u043c. \u041d\u043e\u0432\u044b\u0435 \u0442\u0435\u043d\u0437\u043e\u0440\u043d\u044b\u0435 \u044f\u0434\u0440\u0430 \u043f\u043e\u0434\u0434\u0435\u0440\u0436\u0438\u0432\u0430\u044e\u0442 \u0443\u0441\u043a\u043e\u0440\u0435\u043d\u0438\u0435 FP8. \u042d\u0442\u0430 \u043e\u0441\u043e\u0431\u0435\u043d\u043d\u043e\u0441\u0442\u044c \u043d\u043e\u0432\u043e\u0433\u043e \u0433\u0440\u0430\u0444\u0438\u0447\u0435\u0441\u043a\u043e\u0433\u043e \u043f\u0440\u043e\u0446\u0435\u0441\u0441\u043e\u0440\u0430 \u043c\u043e\u0436\u0435\u0442 \u0445\u043e\u0440\u043e\u0448\u043e \u043f\u043e\u0434\u043e\u0439\u0442\u0438 \u0442\u0435\u043c, \u043a\u0442\u043e \u0440\u0430\u0437\u0440\u0430\u0431\u0430\u0442\u044b\u0432\u0430\u0435\u0442 \u0438 \u0440\u0430\u0437\u0432\u0435\u0440\u0442\u044b\u0432\u0430\u0435\u0442 \u043c\u043e\u0434\u0435\u043b\u0438 \u0418\u0418 \u0432 \u0442\u0430\u043a\u0438\u0445 \u0441\u0440\u0435\u0434\u0430\u0445, \u043a\u0430\u043a \u0433\u0435\u043d\u043e\u043c\u0438\u043a\u0430 \u0438 \u043a\u043e\u043c\u043f\u044c\u044e\u0442\u0435\u0440\u043d\u043e\u0435 \u0437\u0440\u0435\u043d\u0438\u0435.<\/p>\n

\u0422\u0430\u043a\u0436\u0435 \u0441\u0442\u043e\u0438\u0442 \u043e\u0442\u043c\u0435\u0442\u0438\u0442\u044c, \u0447\u0442\u043e \u0443\u0432\u0435\u043b\u0438\u0447\u0435\u043d\u0438\u0435 \u043a\u043e\u043b\u0438\u0447\u0435\u0441\u0442\u0432\u0430 \u043c\u0435\u0445\u0430\u043d\u0438\u0437\u043c\u043e\u0432 \u043a\u043e\u0434\u0438\u0440\u043e\u0432\u0430\u043d\u0438\u044f \u0438 \u0434\u0435\u043a\u043e\u0434\u0438\u0440\u043e\u0432\u0430\u043d\u0438\u044f \u0434\u0435\u043b\u0430\u0435\u0442 RTX 4000 SFF Ada \u0445\u043e\u0440\u043e\u0448\u0438\u043c \u0440\u0435\u0448\u0435\u043d\u0438\u0435\u043c \u0434\u043b\u044f \u043c\u0443\u043b\u044c\u0442\u0438\u043c\u0435\u0434\u0438\u0439\u043d\u044b\u0445 \u0440\u0430\u0431\u043e\u0447\u0438\u0445 \u043d\u0430\u0433\u0440\u0443\u0437\u043e\u043a, \u0442\u0430\u043a\u0438\u0445 \u043a\u0430\u043a \u0440\u0430\u0431\u043e\u0442\u0430 \u0441 \u0432\u0438\u0434\u0435\u043e.<\/p>\n

\u0422\u0435\u0445\u043d\u0438\u0447\u0435\u0441\u043a\u0438\u0435 \u0445\u0430\u0440\u0430\u043a\u0442\u0435\u0440\u0438\u0441\u0442\u0438\u043a\u0438 \u0432\u0438\u0434\u0435\u043e\u043a\u0430\u0440\u0442 NVIDIA RTX A4000 \u0438 RTX A5000, RTX 3090<\/strong><\/p>\n

\n
\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
\n

\n<\/td>\n

\n

RTX A4000 ADA\u00a0<\/strong><\/p>\n<\/td>\n

\n

NVIDIA RTX A4000<\/strong><\/p>\n<\/td>\n

\n

NVIDIA RTX A5000<\/strong><\/p>\n<\/td>\n

\n

RTX 3090<\/strong><\/p>\n<\/td>\n<\/tr>\n

\n

\u0410\u0440\u0445\u0438\u0442\u0435\u043a\u0442\u0443\u0440\u0430<\/strong><\/p>\n<\/td>\n

\n

Ada Lovelace<\/p>\n<\/td>\n

\n

Ampere<\/p>\n<\/td>\n

\n

Ampere<\/p>\n<\/td>\n

\n

Ampere<\/p>\n<\/td>\n<\/tr>\n

\n

\u0422\u0435\u0445\u043f\u0440\u043e\u0446\u0435\u0441\u0441<\/strong><\/p>\n<\/td>\n

\n

5 \u043d\u043c<\/p>\n<\/td>\n

\n

8 \u043d\u043c<\/p>\n<\/td>\n

\n

8 \u043d\u043c<\/p>\n<\/td>\n

\n

8 \u043d\u043c<\/p>\n<\/td>\n<\/tr>\n

\n

\u0413\u0440\u0430\u0444\u0438\u0447\u0435\u0441\u043a\u0438\u0439 \u043f\u0440\u043e\u0446\u0435\u0441\u0441\u043e\u0440<\/strong><\/p>\n<\/td>\n

\n

AD104<\/p>\n<\/td>\n

\n

GA102<\/p>\n<\/td>\n

\n

GA104<\/p>\n<\/td>\n

\n

GA102<\/p>\n<\/td>\n<\/tr>\n

\n

\u041a\u043e\u043b\u0438\u0447\u0435\u0441\u0442\u0432\u043e \u0442\u0440\u0430\u043d\u0437\u0438\u0441\u0442\u043e\u0440\u043e\u0432 (\u043c\u043b\u043d)<\/strong><\/p>\n<\/td>\n

\n

35,800<\/p>\n<\/td>\n

\n

17,400<\/p>\n<\/td>\n

\n

28,300<\/p>\n<\/td>\n

\n

28,300<\/p>\n<\/td>\n<\/tr>\n

\n

\u041f\u0440\u043e\u043f\u0443\u0441\u043a\u043d\u0430\u044f \u0441\u043f\u043e\u0441\u043e\u0431\u043d\u043e\u0441\u0442\u044c \u043f\u0430\u043c\u044f\u0442\u0438 (\u0413\u0431\/\u0441)<\/strong><\/p>\n<\/td>\n

\n

280.0<\/p>\n<\/td>\n

\n

448<\/p>\n<\/td>\n

\n

768<\/p>\n<\/td>\n

\n

936.2<\/p>\n<\/td>\n<\/tr>\n

\n

\u0420\u0430\u0437\u0440\u044f\u0434\u043d\u043e\u0441\u0442\u044c \u0448\u0438\u043d\u044b \u0432\u0438\u0434\u0435\u043e\u043f\u0430\u043c\u044f\u0442\u0438 (\u0431\u0438\u0442)<\/strong><\/p>\n<\/td>\n

\n

160<\/p>\n<\/td>\n

\n

256<\/p>\n<\/td>\n

\n

384<\/p>\n<\/td>\n

\n

384<\/p>\n<\/td>\n<\/tr>\n

\n

\u041f\u0430\u043c\u044f\u0442\u044c GPU (\u0413\u0431)<\/strong><\/p>\n<\/td>\n

\n

20<\/p>\n<\/td>\n

\n

16<\/p>\n<\/td>\n

\n

24<\/p>\n<\/td>\n

\n

24<\/p>\n<\/td>\n<\/tr>\n

\n

\u0422\u0438\u043f \u043f\u0430\u043c\u044f\u0442\u0438<\/strong><\/p>\n<\/td>\n

\n

GDDR6<\/p>\n<\/td>\n

\n

GDDR6<\/p>\n<\/td>\n

\n

GDDR6<\/p>\n<\/td>\n

\n

GDDR6X<\/p>\n<\/td>\n<\/tr>\n

\n

\u042f\u0434\u0440\u0430 CUDA<\/strong><\/p>\n<\/td>\n

\n

6,144<\/p>\n<\/td>\n

\n

6 144<\/p>\n<\/td>\n

\n

8192<\/p>\n<\/td>\n

\n

10496<\/p>\n<\/td>\n<\/tr>\n

\n

\u0422\u0435\u043d\u0437\u043e\u0440\u043d\u044b\u0435 \u044f\u0434\u0440\u0430<\/strong><\/p>\n<\/td>\n

\n

192<\/p>\n<\/td>\n

\n

192<\/p>\n<\/td>\n

\n

256<\/p>\n<\/td>\n

\n

328<\/p>\n<\/td>\n<\/tr>\n

\n

\u042f\u0434\u0440\u0430 RT<\/strong><\/p>\n<\/td>\n

\n

48<\/p>\n<\/td>\n

\n

48<\/p>\n<\/td>\n

\n

64<\/p>\n<\/td>\n

\n

82<\/p>\n<\/td>\n<\/tr>\n

\n

SP perf (\u0442\u0435\u0440\u0430\u0444\u043b\u043e\u043f\u0441)<\/strong><\/p>\n<\/td>\n

\n

19.2<\/p>\n<\/td>\n

\n

19,2\u00a0<\/p>\n<\/td>\n

\n

27,8<\/p>\n<\/td>\n

\n

35,6\u00a0<\/p>\n<\/td>\n<\/tr>\n

\n

RT Core performance (\u0442\u0435\u0440\u0430\u0444\u043b\u043e\u043f\u0441)<\/strong><\/p>\n<\/td>\n

\n

44.3<\/p>\n<\/td>\n

\n

37,4<\/p>\n<\/td>\n

\n

54,2<\/p>\n<\/td>\n

\n

69,5<\/p>\n<\/td>\n<\/tr>\n

\n

Tensor performance (\u0442\u0435\u0440\u0430\u0444\u043b\u043e\u043f\u0441)<\/strong><\/p>\n<\/td>\n

\n

306.8<\/p>\n<\/td>\n

\n

153,4<\/p>\n<\/td>\n

\n

222,2<\/p>\n<\/td>\n

\n

285<\/p>\n<\/td>\n<\/tr>\n

\n

\u041c\u0430\u043a\u0441\u0438\u043c\u0430\u043b\u044c\u043d\u0430\u044f \u043c\u043e\u0449\u043d\u043e\u0441\u0442\u044c (\u0412\u0442)<\/strong><\/p>\n<\/td>\n

\n

70<\/p>\n<\/td>\n

\n

140<\/p>\n<\/td>\n

\n

230<\/p>\n<\/td>\n

\n

350<\/p>\n<\/td>\n<\/tr>\n

\n

\u0418\u043d\u0442\u0435\u0440\u0444\u0435\u0439\u0441<\/strong><\/p>\n<\/td>\n

\n

PCIe 4.0 x 16<\/p>\n<\/td>\n

\n

PCI-E 4.0 x16<\/p>\n<\/td>\n

\n

PCI-E 4.0 x16<\/p>\n<\/td>\n

\n

PCIe 4.0 x16<\/p>\n<\/td>\n<\/tr>\n

\n

\u0420\u0430\u0437\u044a\u0435\u043c\u044b<\/strong><\/p>\n<\/td>\n

\n

4x Mini DisplayPort 1.4a<\/p>\n<\/td>\n

\n

\u0414\u041f 1.4 (4)<\/p>\n<\/td>\n

\n

\u0414\u041f 1.4 (4)<\/p>\n<\/td>\n

\n

\u0414\u041f 1.4 (4)<\/p>\n<\/td>\n<\/tr>\n

\n

\u0424\u043e\u0440\u043c-\u0444\u0430\u043a\u0442\u043e\u0440<\/strong><\/p>\n<\/td>\n

\n

2 \u0441\u043b\u043e\u0442\u0430<\/p>\n<\/td>\n

\n

1 \u0441\u043b\u043e\u0442<\/p>\n<\/td>\n

\n

2 \u0441\u043b\u043e\u0442\u0430<\/p>\n<\/td>\n

\n

2-3 \u0441\u043b\u043e\u0442\u0430<\/p>\n<\/td>\n<\/tr>\n

\n

\u041f\u0440\u043e\u0433\u0440\u0430\u043c\u043c\u043d\u043e\u0435 \u043e\u0431\u0435\u0441\u043f\u0435\u0447\u0435\u043d\u0438\u0435 vGPU<\/strong><\/p>\n<\/td>\n

\n

\u043d\u0435\u0442<\/p>\n<\/td>\n

\n

\u043d\u0435\u0442<\/p>\n<\/td>\n

\n

\u0435\u0441\u0442\u044c \u043d\u0435\u043e\u0433\u0440\u0430\u043d\u0438\u0447\u0435\u043d\u043d\u043e<\/p>\n<\/td>\n

\n

\u0435\u0441\u0442\u044c \u0441 \u043e\u0433\u0440\u0430\u043d\u0438\u0447\u0435\u043d\u0438\u044f\u043c\u0438<\/p>\n<\/td>\n<\/tr>\n

\n

Nvlink<\/strong><\/p>\n<\/td>\n

\n

\u043d\u0435\u0442<\/p>\n<\/td>\n

\n

\u043d\u0435\u0442<\/p>\n<\/td>\n

\n

2x RTX A5000<\/p>\n<\/td>\n

\n

\u0435\u0441\u0442\u044c<\/p>\n<\/td>\n<\/tr>\n

\n

\u041f\u043e\u0434\u0434\u0435\u0440\u0436\u043a\u0430 CUDA<\/strong><\/p>\n<\/td>\n

\n

11.6<\/p>\n<\/td>\n

\n

8.6<\/p>\n<\/td>\n

\n

8.6<\/p>\n<\/td>\n

\n

8.6<\/p>\n<\/td>\n<\/tr>\n

\n

\u041f\u043e\u0434\u0434\u0435\u0440\u0436\u043a\u0430 VULKAN<\/strong><\/p>\n<\/td>\n

\n

1.3<\/p>\n<\/td>\n

\n

\u0435\u0441\u0442\u044c<\/p>\n<\/td>\n

\n

\u0435\u0441\u0442\u044c<\/p>\n<\/td>\n

\n

\u0435\u0441\u0442\u044c, 1.2<\/p>\n<\/td>\n<\/tr>\n

\n

\u0426\u0435\u043d\u0430 (\u0440\u0443\u0431.)<\/strong><\/p>\n<\/td>\n

\n

100 000<\/p>\n<\/td>\n

\n

125 000\u00a0<\/p>\n<\/td>\n

\n

220 000\u00a0<\/p>\n<\/td>\n

\n

100 000<\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<\/div>\n

\u041e\u043f\u0438\u0441\u0430\u043d\u0438\u0435 \u0442\u0435\u0441\u0442\u043e\u0432\u043e\u0439 \u0441\u0440\u0435\u0434\u044b<\/strong><\/p>\n

\n
\n\n\n\n\n\n\n\n\n
\n

\n<\/td>\n

\n

RTX A4000 ADA<\/p>\n<\/td>\n

\n

RTX A4000<\/p>\n<\/td>\n<\/tr>\n

\n

\u041f\u0440\u043e\u0446\u0435\u0441\u0441\u043e\u0440<\/p>\n<\/td>\n

\n

AMD Ryzen 9 5950X 3.4GHz (16 cores)<\/p>\n<\/td>\n

\n

OctaCore Intel Xeon E-2288G, 3,5 GHz<\/p>\n<\/td>\n<\/tr>\n

\n

\u041e\u043f\u0435\u0440\u0430\u0442\u0438\u0432\u043d\u0430\u044f \u043f\u0430\u043c\u044f\u0442\u044c<\/p>\n<\/td>\n

\n

4x 32 Gb DDR4 ECC SO-DIMM<\/p>\n<\/td>\n

\n

2x 32 GB DDR4-3200 ECC DDR4 SDRAM 1600 \u041c\u0413\u0446<\/p>\n<\/td>\n<\/tr>\n

\n

\u041d\u0430\u043a\u043e\u043f\u0438\u0442\u0435\u043b\u044c<\/p>\n<\/td>\n

\n

1Tb NVMe SSD<\/p>\n<\/td>\n

\n

Samsung SSD 980 PRO 1TB<\/p>\n<\/td>\n<\/tr>\n

\n

\u041c\u0430\u0442\u0435\u0440\u0438\u043d\u0441\u043a\u0430\u044f \u043f\u043b\u0430\u0442\u0430<\/p>\n<\/td>\n

\n

ASRock X570D4I-2T<\/p>\n<\/td>\n

\n

Asus P11C-I Series<\/p>\n<\/td>\n<\/tr>\n

\n

\u041e\u043f\u0435\u0440\u0430\u0446\u0438\u043e\u043d\u043d\u0430\u044f \u0441\u0438\u0441\u0442\u0435\u043c\u0430\u00a0<\/p>\n<\/td>\n

\n

Microsoft Windows 10<\/p>\n<\/td>\n

\n

Microsoft Windows 10<\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<\/div>\n

\u0420\u0435\u0437\u0443\u043b\u044c\u0442\u0430\u0442\u044b \u0432 \u0442\u0435\u0441\u0442\u0430\u0445<\/h2>\n

V-Ray 5 Benchmark<\/strong><\/p>\n

\"Points<\/p>\n
Points scored<\/figcaption><\/div>\n<\/figure>\n
\"Points<\/p>\n
Points scored<\/figcaption><\/div>\n<\/figure>\n

\u0422\u0435\u0441\u0442\u044b V-Ray GPU CUDA \u0438 RTX \u043f\u043e\u0437\u0432\u043e\u043b\u044f\u044e\u0442 \u0438\u0437\u043c\u0435\u0440\u0438\u0442\u044c \u043e\u0442\u043d\u043e\u0441\u0438\u0442\u0435\u043b\u044c\u043d\u0443\u044e \u043f\u0440\u043e\u0438\u0437\u0432\u043e\u0434\u0438\u0442\u0435\u043b\u044c\u043d\u043e\u0441\u0442\u044c GPU \u043f\u0440\u0438 \u0440\u0435\u043d\u0434\u0435\u0440\u0438\u043d\u0433\u0435. GPU RTX A4000 \u043d\u0435\u0437\u043d\u0430\u0447\u0438\u0442\u0435\u043b\u044c\u043d\u043e \u0443\u0441\u0442\u0443\u043f\u0430\u0435\u0442 \u043f\u043e \u043f\u0440\u043e\u0438\u0437\u0432\u043e\u0434\u0438\u0442\u0435\u043b\u044c\u043d\u043e\u0441\u0442\u0438 RTX A4000 ADA (4% \u0438 11% \u0441\u043e\u043e\u0442\u0432\u0435\u0442\u0441\u0442\u0432\u0435\u043d\u043d\u043e).\u00a0<\/p>\n

\u041c\u0430\u0448\u0438\u043d\u043d\u043e\u0435 \u043e\u0431\u0443\u0447\u0435\u043d\u0438\u0435<\/h2>\n

\u00ab\u0421\u043e\u0431\u0430\u043a\u0438 \u043f\u0440\u043e\u0442\u0438\u0432 \u043a\u043e\u0448\u0435\u043a\u00bb<\/strong><\/p>\n

\u0414\u043b\u044f \u0441\u0440\u0430\u0432\u043d\u0435\u043d\u0438\u044f \u043f\u0440\u043e\u0438\u0437\u0432\u043e\u0434\u0438\u0442\u0435\u043b\u044c\u043d\u043e\u0441\u0442\u0438 GPU \u0434\u043b\u044f \u043d\u0435\u0439\u0440\u043e\u0441\u0435\u0442\u0435\u0439 \u043c\u044b \u0438\u0441\u043f\u043e\u043b\u044c\u0437\u0443\u0435\u043c \u043d\u0430\u0431\u043e\u0440 \u0434\u0430\u043d\u043d\u044b\u0445 \u00ab\u0421\u043e\u0431\u0430\u043a\u0438 \u043f\u0440\u043e\u0442\u0438\u0432 \u043a\u043e\u0448\u0435\u043a\u00bb \u2014 \u0442\u0435\u0441\u0442 \u0430\u043d\u0430\u043b\u0438\u0437\u0438\u0440\u0443\u0435\u0442 \u0441\u043e\u0434\u0435\u0440\u0436\u0438\u043c\u043e\u0435 \u0444\u043e\u0442\u043e\u0433\u0440\u0430\u0444\u0438\u0438 \u0438 \u0440\u0430\u0437\u043b\u0438\u0447\u0430\u0435\u0442, \u0438\u0437\u043e\u0431\u0440\u0430\u0436\u0435\u043d\u0430 \u043d\u0430 \u0444\u043e\u0442\u043e \u043a\u043e\u0448\u043a\u0430 \u0438\u043b\u0438 \u0441\u043e\u0431\u0430\u043a\u0430. \u0412\u0441\u0435 \u043d\u0435\u043e\u0431\u0445\u043e\u0434\u0438\u043c\u044b\u0435 \u0438\u0441\u0445\u043e\u0434\u043d\u044b\u0435 \u0434\u0430\u043d\u043d\u044b\u0435 \u043d\u0430\u0445\u043e\u0434\u044f\u0442\u0441\u044f \u0437\u0434\u0435\u0441\u044c<\/u><\/a>. \u041c\u044b \u0437\u0430\u043f\u0443\u0441\u043a\u0430\u043b\u0438 \u044d\u0442\u043e\u0442 \u0442\u0435\u0441\u0442 \u043d\u0430 \u0440\u0430\u0437\u043d\u044b\u0445 GPU \u0438 \u0432 \u0440\u0430\u0437\u043b\u0438\u0447\u043d\u044b\u0445 \u043e\u0431\u043b\u0430\u0447\u043d\u044b\u0445 \u0441\u0435\u0440\u0432\u0438\u0441\u0430\u0445 \u0438 \u043f\u043e\u043b\u0443\u0447\u0438\u043b\u0438 \u0441\u043b\u0435\u0434\u0443\u044e\u0449\u0438\u0435 \u0440\u0435\u0437\u0443\u043b\u044c\u0442\u0430\u0442\u044b:<\/p>\n

\"Points<\/p>\n
Points scored<\/figcaption><\/div>\n<\/figure>\n

\u0412 \u044d\u0442\u043e\u043c \u0442\u0435\u0441\u0442\u0435 \u0440\u0435\u0437\u0443\u043b\u044c\u0442\u0430\u0442 RTX A4000 ADA \u043d\u0435\u0437\u043d\u0430\u0447\u0438\u0442\u0435\u043b\u044c\u043d\u043e \u043f\u0440\u0435\u0432\u0437\u043e\u0448\u0435\u043b RTX A4000 (9%), \u043d\u043e \u0441\u043b\u0435\u0434\u0443\u0435\u0442 \u043f\u043e\u043c\u043d\u0438\u0442\u044c \u043e \u043d\u0435\u0431\u043e\u043b\u044c\u0448\u043e\u043c \u0440\u0430\u0437\u043c\u0435\u0440\u0435 \u0438 \u043d\u0438\u0437\u043a\u043e\u043c \u044d\u043d\u0435\u0440\u0433\u043e\u043f\u043e\u0442\u0440\u0435\u0431\u043b\u0435\u043d\u0438\u0438 \u043d\u043e\u0432\u043e\u0433\u043e GPU.<\/p>\n

AI-Benchmark<\/u><\/a><\/p>\n

AI-Benchmark \u043f\u043e\u0437\u0432\u043e\u043b\u044f\u0435\u0442 \u0438\u0437\u043c\u0435\u0440\u0438\u0442\u044c \u043f\u0440\u043e\u0438\u0437\u0432\u043e\u0434\u0438\u0442\u0435\u043b\u044c\u043d\u043e\u0441\u0442\u044c \u0443\u0441\u0442\u0440\u043e\u0439\u0441\u0442\u0432\u0430 \u0432\u043e \u0432\u0440\u0435\u043c\u044f \u0432\u044b\u043f\u043e\u043b\u043d\u0435\u043d\u0438\u044f \u0437\u0430\u0434\u0430\u0447\u0438 \u0432\u044b\u0432\u043e\u0434\u0430 AI-\u043c\u043e\u0434\u0435\u043b\u0435\u0439. \u0415\u0434\u0438\u043d\u0438\u0446\u044b \u0438\u0437\u043c\u0435\u0440\u0435\u043d\u0438\u044f \u043c\u043e\u0433\u0443\u0442 \u0437\u0430\u0432\u0438\u0441\u0435\u0442\u044c \u043e\u0442 \u0442\u0435\u0441\u0442\u0430, \u043d\u043e \u043e\u0431\u044b\u0447\u043d\u043e \u044d\u0442\u043e \u043a\u043e\u043b\u0438\u0447\u0435\u0441\u0442\u0432\u043e \u043e\u043f\u0435\u0440\u0430\u0446\u0438\u0439 \u0432 \u0441\u0435\u043a\u0443\u043d\u0434\u0443 (OPS) \u0438\u043b\u0438 \u043a\u043e\u043b\u0438\u0447\u0435\u0441\u0442\u0432\u043e \u0438\u0437\u043e\u0431\u0440\u0430\u0436\u0435\u043d\u0438\u0439 \u0432 \u0441\u0435\u043a\u0443\u043d\u0434\u0443 (FPS).<\/p>\n

\"Points<\/p>\n
Points scored<\/figcaption><\/div>\n<\/figure>\n
\n
\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
<\/td>\n\n

RTX A4000<\/strong><\/p>\n<\/td>\n

\n

RTX A4000 ADA<\/strong><\/p>\n<\/td>\n<\/tr>\n

\n

1\/19. MobileNet-V2<\/strong><\/p>\n<\/td>\n

\n

1.1 \u2014 inference | batch=50, size=224×224: 38.5 \u00b1 2.4 ms<\/p>\n

1.2 \u2014 training | batch=50, size=224×224: 109 \u00b1 4 ms<\/p>\n<\/td>\n

\n

1.1 \u2014 inference | batch=50, size=224×224: 53.5 \u00b1 0.7 ms<\/p>\n

1.2 \u2014 training | batch=50, size=224×224: 130.1 \u00b1 0.6 ms<\/p>\n<\/td>\n<\/tr>\n

\n

2\/19. Inception-V3<\/strong><\/p>\n<\/td>\n

\n

2.1 \u2014 inference | batch=20, size=346×346: 36.1 \u00b1 1.8 ms<\/p>\n

2.2 \u2014 training | batch=20, size=346×346: 137.4 \u00b1 0.6 ms<\/p>\n<\/td>\n

\n

2.1 \u2014 inference | batch=20, size=346×346: 36.8 \u00b1 1.1 ms<\/p>\n

2.2 \u2014 training | batch=20, size=346×346: 147.5 \u00b1 0.8 ms<\/p>\n<\/td>\n<\/tr>\n

\n

3\/19. Inception-V4<\/strong><\/p>\n<\/td>\n

\n

3.1 \u2014 inference | batch=10, size=346×346: 34.0 \u00b1 0.9 ms<\/p>\n

3.2 \u2014 training | batch=10, size=346×346: 139.4 \u00b1 1.0 ms<\/p>\n<\/td>\n

\n

3.1 \u2014 inference | batch=10, size=346×346: 33.0 \u00b1 0.8 ms<\/p>\n

3.2 \u2014 training | batch=10, size=346×346: 135.7 \u00b1 0.9 ms<\/p>\n<\/td>\n<\/tr>\n

\n

4\/19. Inception-ResNet-V2<\/strong><\/p>\n<\/td>\n

\n

4.1 \u2014 inference | batch=10, size=346×346: 45.7 \u00b1 0.6 ms<\/p>\n

4.2 \u2014 training | batch=8, size=346×346: 153.4 \u00b1 0.8 ms<\/p>\n<\/td>\n

\n

4.1 \u2014 inference batch=10, size=346×346: 33.6 \u00b1 0.7 ms<\/p>\n

4.2 \u2014 training batch=8, size=346×346: 132 \u00b1 1 ms<\/p>\n<\/td>\n<\/tr>\n

\n

5\/19. ResNet-V2-50<\/strong><\/p>\n<\/td>\n

\n

5.1 \u2014 inference | batch=10, size=346×346: 25.3 \u00b1 0.5 ms<\/p>\n

5.2 \u2014 training | batch=10, size=346×346: 91.1 \u00b1 0.8 ms<\/p>\n<\/td>\n

\n

5.1 \u2014 inference | batch=10, size=346×346: 26.1 \u00b1 0.5 ms<\/p>\n

5.2 \u2014 training | batch=10, size=346×346: 92.3 \u00b1 0.6 ms<\/p>\n<\/td>\n<\/tr>\n

\n

6\/19. ResNet-V2-152<\/strong><\/p>\n<\/td>\n

\n

6.1 \u2014 inference | batch=10, size=256×256: 32.4 \u00b1 0.5 ms<\/p>\n

6.2 \u2014 training | batch=10, size=256×256: 131.4 \u00b1 0.7 ms<\/p>\n<\/td>\n

\n

6.1 \u2014 inference | batch=10, size=256×256: 23.7 \u00b1 0.6 ms<\/p>\n

6.2 \u2014 training | batch=10, size=256×256: 107.1 \u00b1 0.9 ms<\/p>\n<\/td>\n<\/tr>\n

\n

7\/19. VGG-16<\/strong><\/p>\n<\/td>\n

\n

7.1 \u2014 inference | batch=20, size=224×224: 54.9 \u00b1 0.9 ms<\/p>\n

7.2 \u2014 training | batch=2, size=224×224: 83.6 \u00b1 0.7 ms<\/p>\n<\/td>\n

\n

7.1 \u2014 inference | batch=20, size=224×224: 66.3 \u00b1 0.9 ms<\/p>\n

7.2 \u2014 training | batch=2, size=224×224: 109.3 \u00b1 0.8 ms<\/p>\n<\/td>\n<\/tr>\n

\n

8\/19. SRCNN 9-5-5<\/strong><\/p>\n<\/td>\n

\n

8.1 \u2014 inference | batch=10, size=512×512: 51.5 \u00b1 0.9 ms<\/p>\n

8.2 \u2014 inference | batch=1, size=1536×1536: 45.7 \u00b1 0.9 ms<\/p>\n

8.3 \u2014 training | batch=10, size=512×512: 183 \u00b1 1 ms<\/p>\n<\/td>\n

\n

8.1 \u2014 inference | batch=10, size=512×512: 59.9 \u00b1 1.6 ms<\/p>\n

8.2 \u2014 inference | batch=1, size=1536×1536: 53.1 \u00b1 0.7 ms<\/p>\n

8.3 \u2014 training | batch=10, size=512×512: 176 \u00b1 2 ms<\/p>\n<\/td>\n<\/tr>\n

\n

9\/19. VGG-19 Super-Res<\/strong><\/p>\n<\/td>\n

\n

9.1 \u2014 inference | batch=10, size=256×256: 99.5 \u00b1 0.8 ms<\/p>\n

9.2 \u2014 inference | batch=1, size=1024×1024: 162 \u00b1 1 ms<\/p>\n

9.3 \u2014 training | batch=10, size=224×224: 204 \u00b1 2 ms<\/p>\n<\/td>\n

\n

\n<\/td>\n<\/tr>\n

\n

10\/19. ResNet-SRGAN<\/strong><\/p>\n<\/td>\n

\n

10.1 \u2014 inference | batch=10, size=512×512: 85.8 \u00b1 0.6 ms<\/p>\n

10.2 \u2014 inference | batch=1, size=1536×1536: 82.4 \u00b1 1.9 ms<\/p>\n

10.3 \u2014 training | batch=5, size=512×512: 133 \u00b1 1 ms<\/p>\n<\/td>\n

\n

10.1 \u2014 inference | batch=10, size=512×512: 98.9 \u00b1 0.8 ms<\/p>\n

10.2 \u2014 inference | batch=1, size=1536×1536: 86.1 \u00b1 0.6 ms<\/p>\n

10.3 \u2014 training | batch=5, size=512×512: 130.9 \u00b1 0.6 ms<\/p>\n<\/td>\n<\/tr>\n

\n

11\/19. ResNet-DPED<\/strong><\/p>\n<\/td>\n

\n

11.1 \u2014 inference | batch=10, size=256×256: 114.9 \u00b1 0.6 ms<\/p>\n

11.2 \u2014 inference | batch=1, size=1024×1024: 182 \u00b1 2 ms<\/p>\n

11.3 \u2014 training | batch=15, size=128×128: 178.1 \u00b1 0.8 ms<\/p>\n<\/td>\n

\n

11.1 \u2014 inference | batch=10, size=256×256: 146.4 \u00b1 0.5 ms<\/p>\n

11.2 \u2014 inference | batch=1, size=1024×1024: 234.3 \u00b1 0.5 ms<\/p>\n

11.3 \u2014 training | batch=15, size=128×128: 234.7 \u00b1 0.6 ms<\/p>\n<\/td>\n<\/tr>\n

\n

12\/19. U-Net<\/strong><\/p>\n<\/td>\n

\n

12.1 \u2014 inference | batch=4, size=512×512: 180.8 \u00b1 0.7 ms<\/p>\n

12.2 \u2014 inference | batch=1, size=1024×1024: 177.0 \u00b1 0.4 ms<\/p>\n

12.3 \u2014 training | batch=4, size=256×256: 198.6 \u00b1 0.5 ms<\/p>\n<\/td>\n

\n

12.1 \u2014 inference | batch=4, size=512×512: 222.9 \u00b1 0.5 ms<\/p>\n

12.2 \u2014 inference | batch=1, size=1024×1024: 220.4 \u00b1 0.6 ms<\/p>\n

12.3 \u2014 training | batch=4, size=256×256: 229.1 \u00b1 0.7 ms<\/p>\n<\/td>\n<\/tr>\n

\n

13\/19. Nvidia-SPADE<\/strong><\/p>\n<\/td>\n

\n

13.1 \u2014 inference | batch=5, size=128×128: 54.5 \u00b1 0.5 ms<\/p>\n

13.2 \u2014 training | batch=1, size=128×128: 103.6 \u00b1 0.6 ms<\/p>\n<\/td>\n

\n

13.1 \u2014 inference | batch=5, size=128×128: 59.6 \u00b1 0.6 ms<\/p>\n

13.2 \u2014 training | batch=1, size=128×128: 94.6 \u00b1 0.6 ms<\/p>\n<\/td>\n<\/tr>\n

\n

14\/19. ICNet<\/strong><\/p>\n<\/td>\n

\n

14.1 \u2014 inference | batch=5, size=1024×1536: 126.3 \u00b1 0.8 ms<\/p>\n

14.2 \u2014 training | batch=10, size=1024×1536: 426 \u00b1 9 ms<\/p>\n<\/td>\n

\n

14.1 \u2014 inference | batch=5, size=1024×1536: 144 \u00b1 4 ms<\/p>\n

14.2 \u2014 training | batch=10, size=1024×1536: 475 \u00b1 17 ms<\/p>\n<\/td>\n<\/tr>\n

\n

15\/19. PSPNet<\/strong><\/p>\n<\/td>\n

\n

15.1 \u2014 inference | batch=5, size=720×720: 249 \u00b1 12 ms<\/p>\n

15.2 \u2014 training | batch=1, size=512×512: 104.6 \u00b1 0.6 ms<\/p>\n<\/td>\n

\n

15.1 \u2014 inference | batch=5, size=720×720: 291.4 \u00b1 0.5 ms<\/p>\n

15.2 \u2014 training | batch=1, size=512×512: 99.8 \u00b1 0.9 ms<\/p>\n<\/td>\n<\/tr>\n

\n

16\/19. DeepLab<\/strong><\/p>\n<\/td>\n

\n

16.1 \u2014 inference | batch=2, size=512×512: 71.7 \u00b1 0.6 ms<\/p>\n

16.2 \u2014 training | batch=1, size=384×384: 84.9 \u00b1 0.5 ms<\/p>\n<\/td>\n

\n

16.1 \u2014 inference | batch=2, size=512×512: 71.5 \u00b1 0.7 ms<\/p>\n

16.2 \u2014 training | batch=1, size=384×384: 69.4 \u00b1 0.6 ms<\/p>\n<\/td>\n<\/tr>\n

\n

17\/19. Pixel-RNN<\/strong><\/p>\n<\/td>\n

\n

17.1 \u2014 inference | batch=50, size=64×64: 299 \u00b1 14 ms<\/p>\n

17.2 \u2014 training | batch=10, size=64×64: 1258 \u00b1 64 ms<\/p>\n<\/td>\n

\n

17.1 \u2014 inference | batch=50, size=64×64: 321 \u00b1 30 ms<\/p>\n

17.2 \u2014 training | batch=10, size=64×64: 1278 \u00b1 74 ms<\/p>\n<\/td>\n<\/tr>\n

\n

18\/19. LSTM-Sentiment<\/strong><\/p>\n<\/td>\n

\n

18.1 \u2014 inference | batch=100, size=1024×300: 395 \u00b1 11 ms<\/p>\n

18.2 \u2014 training | batch=10, size=1024×300: 676 \u00b1 15 ms<\/p>\n<\/td>\n

\n

18.1 \u2014 inference | batch=100, size=1024×300: 345 \u00b1 10 ms<\/p>\n

18.2 \u2014 training | batch=10, size=1024×300: 774 \u00b1 17 ms<\/p>\n<\/td>\n<\/tr>\n

\n

19\/19. GNMT-Translation<\/strong><\/p>\n<\/td>\n

\n

19.1 \u2014 inference | batch=1, size=1×20: 119 \u00b1 2 ms<\/p>\n<\/td>\n

\n

19.1 \u2014 inference | batch=1, size=1×20: 156 \u00b1 1 ms<\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<\/div>\n

\u0420\u0435\u0437\u0443\u043b\u044c\u0442\u0430\u0442\u044b \u044d\u0442\u043e\u0433\u043e \u0442\u0435\u0441\u0442\u0430 \u043f\u043e\u043a\u0430\u0437\u044b\u0432\u0430\u044e\u0442, \u0447\u0442\u043e \u043f\u0440\u043e\u0438\u0437\u0432\u043e\u0434\u0438\u0442\u0435\u043b\u044c\u043d\u043e\u0441\u0442\u044c RTX A4000 \u043d\u0435\u0437\u043d\u0430\u0447\u0438\u0442\u0435\u043b\u044c\u043d\u043e (\u043d\u0430 6%) \u0432\u044b\u0448\u0435, \u0447\u0435\u043c \u0443 RTX A4000 ADA. \u041e\u0434\u043d\u0430\u043a\u043e, \u0441\u0442\u043e\u0438\u0442 \u043e\u0442\u043c\u0435\u0442\u0438\u0442\u044c, \u0447\u0442\u043e \u0440\u0435\u0437\u0443\u043b\u044c\u0442\u0430\u0442\u044b \u0442\u0435\u0441\u0442\u043e\u0432 \u043c\u043e\u0433\u0443\u0442 \u0440\u0430\u0437\u043b\u0438\u0447\u0430\u0442\u044c\u0441\u044f \u0432 \u0437\u0430\u0432\u0438\u0441\u0438\u043c\u043e\u0441\u0442\u0438 \u043e\u0442 \u043a\u043e\u043d\u043a\u0440\u0435\u0442\u043d\u044b\u0445 \u0437\u0430\u0434\u0430\u0447 \u0438 \u0443\u0441\u043b\u043e\u0432\u0438\u0439 \u0440\u0430\u0431\u043e\u0442\u044b.<\/p>\n

PyTorch<\/u><\/a><\/p>\n

\n\u0420\u0435\u0437\u0443\u043b\u044c\u0442\u0430\u0442\u044b \u0441 RTX A 4000<\/summary>\n
\n
\n
\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
\n

Benchmarking<\/strong><\/p>\n<\/td>\n

\n

Model average train time (ms)<\/strong><\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type mnasnet0_5<\/p>\n<\/td>\n

\n

62.995805740356445<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type mnasnet0_75<\/p>\n<\/td>\n

\n

98.39066505432129<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type mnasnet1_0<\/p>\n<\/td>\n

\n

126.60405158996582<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type mnasnet1_3<\/p>\n<\/td>\n

\n

186.89460277557373<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type resnet18<\/p>\n<\/td>\n

\n

428.08079719543457<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type resnet34<\/p>\n<\/td>\n

\n

883.5790348052979<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type resnet50<\/p>\n<\/td>\n

\n

1016.3950300216675<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type resnet101<\/p>\n<\/td>\n

\n

1927.2308254241943<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type resnet152<\/p>\n<\/td>\n

\n

2815.663013458252<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type resnext50_32x4d<\/p>\n<\/td>\n

\n

1075.4373741149902<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type resnext101_32x8d<\/p>\n<\/td>\n

\n

4050.0641918182373<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type wide_resnet50_2<\/p>\n<\/td>\n

\n

2615.9953451156616<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type wide_resnet101_2<\/p>\n<\/td>\n

\n

5218.524832725525<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type densenet121<\/p>\n<\/td>\n

\n

751.9759511947632<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type densenet169<\/p>\n<\/td>\n

\n

910.3225564956665<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type densenet201<\/p>\n<\/td>\n

\n

1163.036551475525<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type densenet161<\/p>\n<\/td>\n

\n

2141.505298614502<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type squeezenet1_0<\/p>\n<\/td>\n

\n

203.14435005187988<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type squeezenet1_1<\/p>\n<\/td>\n

\n

98.04857730865479<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type vgg11<\/p>\n<\/td>\n

\n

1697.710485458374<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type vgg11_bn<\/p>\n<\/td>\n

\n

1729.2972660064697<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type vgg13<\/p>\n<\/td>\n

\n

2491.615080833435<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type vgg13_bn<\/p>\n<\/td>\n

\n

2545.1631927490234<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type vgg16<\/p>\n<\/td>\n

\n

3371.1953449249268<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type vgg16_bn<\/p>\n<\/td>\n

\n

3423.8639068603516<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type vgg19_bn<\/p>\n<\/td>\n

\n

4314.5153522491455<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type vgg19<\/p>\n<\/td>\n

\n

4249.422650337219<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type mobilenet_v3_large<\/p>\n<\/td>\n

\n

105.54619789123535<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type mobilenet_v3_small<\/p>\n<\/td>\n

\n

37.6680850982666<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type shufflenet_v2_x0_5<\/p>\n<\/td>\n

\n

26.51611328125<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type shufflenet_v2_x1_0<\/p>\n<\/td>\n

\n

61.260504722595215<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type shufflenet_v2_x1_5<\/p>\n<\/td>\n

\n

105.30067920684814<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type shufflenet_v2_x2_0<\/p>\n<\/td>\n

\n

181.03694438934326<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type mnasnet0_5<\/p>\n<\/td>\n

\n

17.397074699401855<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type mnasnet0_75<\/p>\n<\/td>\n

\n

28.902697563171387<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type mnasnet1_0<\/p>\n<\/td>\n

\n

38.387718200683594<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type mnasnet1_3<\/p>\n<\/td>\n

\n

58.228821754455566<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type resnet18<\/p>\n<\/td>\n

\n

147.95727252960205<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type resnet34<\/p>\n<\/td>\n

\n

293.519492149353<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type resnet50<\/p>\n<\/td>\n

\n

336.44991874694824<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type resnet101<\/p>\n<\/td>\n

\n

637.9982376098633<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type resnet152<\/p>\n<\/td>\n

\n

948.9351654052734<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type resnext50_32x4d<\/p>\n<\/td>\n

\n

372.80876636505127<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type resnext101_32x8d<\/p>\n<\/td>\n

\n

1385.1624917984009<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type wide_resnet50_2<\/p>\n<\/td>\n

\n

873.048791885376<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type wide_resnet101_2<\/p>\n<\/td>\n

\n

1729.2765426635742<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type densenet121<\/p>\n<\/td>\n

\n

270.13323307037354<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type densenet169<\/p>\n<\/td>\n

\n

327.1932888031006<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type densenet201<\/p>\n<\/td>\n

\n

414.733362197876<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type densenet161<\/p>\n<\/td>\n

\n

766.3542318344116<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type squeezenet1_0<\/p>\n<\/td>\n

\n

74.86292839050293<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type squeezenet1_1<\/p>\n<\/td>\n

\n

34.04905319213867<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type vgg11<\/p>\n<\/td>\n

\n

576.3767147064209<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type vgg11_bn<\/p>\n<\/td>\n

\n

580.5839586257935<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type vgg13<\/p>\n<\/td>\n

\n

853.4365510940552<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type vgg13_bn<\/p>\n<\/td>\n

\n

860.3136301040649<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type vgg16<\/p>\n<\/td>\n

\n

1145.091052055359<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type vgg16_bn<\/p>\n<\/td>\n

\n

1152.8028392791748<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type vgg19_bn<\/p>\n<\/td>\n

\n

1444.9562692642212<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type vgg19<\/p>\n<\/td>\n

\n

1437.0987701416016<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type mobilenet_v3_large<\/p>\n<\/td>\n

\n

30.876317024230957<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type mobilenet_v3_small<\/p>\n<\/td>\n

\n

11.234536170959473<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type shufflenet_v2_x0_5<\/p>\n<\/td>\n

\n

7.425284385681152<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type shufflenet_v2_x1_0<\/p>\n<\/td>\n

\n

18.25782299041748<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type shufflenet_v2_x1_5<\/p>\n<\/td>\n

\n

33.34946632385254<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type shufflenet_v2_x2_0<\/p>\n<\/td>\n

\n

57.84676551818848<\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<\/div>\n<\/div>\n<\/details>\n

\n\u0420\u0435\u0437\u0443\u043b\u044c\u0442\u0430\u0442\u044b \u0441 A4000 ADA<\/summary>\n
\n
\n
\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
\n

Benchmarking<\/strong><\/p>\n<\/td>\n

\n

Model average train time<\/strong><\/p>\n<\/td>\n<\/tr>\n

\n

Training half precision type mnasnet0_5<\/p>\n<\/td>\n

\n

20.266618728637695<\/p>\n<\/td>\n<\/tr>\n

\n

Training half precision type mnasnet0_75<\/p>\n<\/td>\n

\n

21.445374488830566<\/p>\n<\/td>\n<\/tr>\n

\n

Training half precision type mnasnet1_0<\/p>\n<\/td>\n

\n

26.714019775390625<\/p>\n<\/td>\n<\/tr>\n

\n

Training half precision type mnasnet1_3<\/p>\n<\/td>\n

\n

26.5126371383667<\/p>\n<\/td>\n<\/tr>\n

\n

Training half precision type resnet18<\/p>\n<\/td>\n

\n

19.624991416931152<\/p>\n<\/td>\n<\/tr>\n

\n

Training half precision type resnet34<\/p>\n<\/td>\n

\n

32.46446132659912<\/p>\n<\/td>\n<\/tr>\n

\n

Training half precision type resnet50<\/p>\n<\/td>\n

\n

57.17473030090332<\/p>\n<\/td>\n<\/tr>\n

\n

Training half precision type resnet101<\/p>\n<\/td>\n

\n

98.20127010345459<\/p>\n<\/td>\n<\/tr>\n

\n

Training half precision type resnet152<\/p>\n<\/td>\n

\n

138.18389415740967<\/p>\n<\/td>\n<\/tr>\n

\n

Training half precision type resnext50_32x4d<\/p>\n<\/td>\n

\n

75.56005001068115<\/p>\n<\/td>\n<\/tr>\n

\n

Training half precision type resnext101_32x8d<\/p>\n<\/td>\n

\n

228.8706636428833<\/p>\n<\/td>\n<\/tr>\n

\n

Training half precision type wide_resnet50_2<\/p>\n<\/td>\n

\n

113.76442432403564<\/p>\n<\/td>\n<\/tr>\n

\n

Training half precision type wide_resnet101_2<\/p>\n<\/td>\n

\n

204.17311191558838<\/p>\n<\/td>\n<\/tr>\n

\n

Training half precision type densenet121<\/p>\n<\/td>\n

\n

68.97401332855225<\/p>\n<\/td>\n<\/tr>\n

\n

Training half precision type densenet169<\/p>\n<\/td>\n

\n

85.16453742980957<\/p>\n<\/td>\n<\/tr>\n

\n

Training half precision type densenet201<\/p>\n<\/td>\n

\n

103.299241065979<\/p>\n<\/td>\n<\/tr>\n

\n

Training half precision type densenet161<\/p>\n<\/td>\n

\n

137.54578113555908<\/p>\n<\/td>\n<\/tr>\n

\n

Training half precision type squeezenet1_0<\/p>\n<\/td>\n

\n

16.71830177307129<\/p>\n<\/td>\n<\/tr>\n

\n

Training half precision type squeezenet1_1<\/p>\n<\/td>\n

\n

12.906527519226074<\/p>\n<\/td>\n<\/tr>\n

\n

Training half precision type vgg11<\/p>\n<\/td>\n

\n

51.7004919052124<\/p>\n<\/td>\n<\/tr>\n

\n

Training half precision type vgg11_bn<\/p>\n<\/td>\n

\n

57.63327598571777<\/p>\n<\/td>\n<\/tr>\n

\n

Training half precision type vgg13<\/p>\n<\/td>\n

\n

86.10869407653809<\/p>\n<\/td>\n<\/tr>\n

\n

Training half precision type vgg13_bn<\/p>\n<\/td>\n

\n

95.86676120758057<\/p>\n<\/td>\n<\/tr>\n

\n

Training half precision type vgg16<\/p>\n<\/td>\n

\n

102.91589260101318<\/p>\n<\/td>\n<\/tr>\n

\n

Training half precision type vgg16_bn<\/p>\n<\/td>\n

\n

113.74778270721436<\/p>\n<\/td>\n<\/tr>\n

\n

Training half precision type vgg19_bn<\/p>\n<\/td>\n

\n

131.56734943389893<\/p>\n<\/td>\n<\/tr>\n

\n

Training half precision type vgg19<\/p>\n<\/td>\n

\n

119.70191955566406<\/p>\n<\/td>\n<\/tr>\n

\n

Training half precision type mobilenet_v3_large<\/p>\n<\/td>\n

\n

31.30636692047119<\/p>\n<\/td>\n<\/tr>\n

\n

Training half precision type mobilenet_v3_small<\/p>\n<\/td>\n

\n

19.44464683532715<\/p>\n<\/td>\n<\/tr>\n

\n

Training half precision type shufflenet_v2_x0_5<\/p>\n<\/td>\n

\n

13.710575103759766<\/p>\n<\/td>\n<\/tr>\n

\n

Training half precision type shufflenet_v2_x1_0<\/p>\n<\/td>\n

\n

23.608479499816895<\/p>\n<\/td>\n<\/tr>\n

\n

Training half precision type shufflenet_v2_x1_5<\/p>\n<\/td>\n

\n

26.793746948242188<\/p>\n<\/td>\n<\/tr>\n

\n

Training half precision type shufflenet_v2_x2_0<\/p>\n<\/td>\n

\n

24.550962448120117<\/p>\n<\/td>\n<\/tr>\n

\n

Inference half precision type mnasnet0_5<\/p>\n<\/td>\n

\n

4.418272972106934<\/p>\n<\/td>\n<\/tr>\n

\n

Inference half precision type mnasnet0_75<\/p>\n<\/td>\n

\n

4.021778106689453<\/p>\n<\/td>\n<\/tr>\n

\n

Inference half precision type mnasnet1_0<\/p>\n<\/td>\n

\n

4.42598819732666<\/p>\n<\/td>\n<\/tr>\n

\n

Inference half precision type mnasnet1_3<\/p>\n<\/td>\n

\n

4.618926048278809<\/p>\n<\/td>\n<\/tr>\n

\n

Inference half precision type resnet18<\/p>\n<\/td>\n

\n

5.803341865539551<\/p>\n<\/td>\n<\/tr>\n

\n

Inference half precision type resnet34<\/p>\n<\/td>\n

\n

9.756693840026855<\/p>\n<\/td>\n<\/tr>\n

\n

Inference half precision type resnet50<\/p>\n<\/td>\n

\n

15.873079299926758<\/p>\n<\/td>\n<\/tr>\n

\n

Inference half precision type resnet101<\/p>\n<\/td>\n

\n

28.268003463745117<\/p>\n<\/td>\n<\/tr>\n

\n

Inference half precision type resnet152<\/p>\n<\/td>\n

\n

40.04594326019287<\/p>\n<\/td>\n<\/tr>\n

\n

Inference half precision type resnext50_32x4d<\/p>\n<\/td>\n

\n

19.53421115875244<\/p>\n<\/td>\n<\/tr>\n

\n

Inference half precision type resnext101_32x8d<\/p>\n<\/td>\n

\n

62.44826316833496<\/p>\n<\/td>\n<\/tr>\n

\n

Inference half precision type wide_resnet50_2<\/p>\n<\/td>\n

\n

33.533992767333984<\/p>\n<\/td>\n<\/tr>\n

\n

Inference half precision type wide_resnet101_2<\/p>\n<\/td>\n

\n

59.60897445678711<\/p>\n<\/td>\n<\/tr>\n

\n

Inference half precision type densenet121<\/p>\n<\/td>\n

\n

18.052735328674316<\/p>\n<\/td>\n<\/tr>\n

\n

Inference half precision type densenet169<\/p>\n<\/td>\n

\n

21.956982612609863<\/p>\n<\/td>\n<\/tr>\n

\n

Inference half precision type densenet201<\/p>\n<\/td>\n

\n

27.85182476043701<\/p>\n<\/td>\n<\/tr>\n

\n

Inference half precision type densenet161<\/p>\n<\/td>\n

\n

37.41891860961914<\/p>\n<\/td>\n<\/tr>\n

\n

Inference half precision type squeezenet1_0<\/p>\n<\/td>\n

\n

4.391803741455078<\/p>\n<\/td>\n<\/tr>\n

\n

Inference half precision type squeezenet1_1<\/p>\n<\/td>\n

\n

2.4281740188598633<\/p>\n<\/td>\n<\/tr>\n

\n

Inference half precision type vgg11<\/p>\n<\/td>\n

\n

17.11493968963623<\/p>\n<\/td>\n<\/tr>\n

\n

Inference half precision type vgg11_bn<\/p>\n<\/td>\n

\n

18.40585231781006<\/p>\n<\/td>\n<\/tr>\n

\n

Inference half precision type vgg13<\/p>\n<\/td>\n

\n

28.438148498535156<\/p>\n<\/td>\n<\/tr>\n

\n

Inference half precision type vgg13_bn<\/p>\n<\/td>\n

\n

30.672597885131836<\/p>\n<\/td>\n<\/tr>\n

\n

Inference half precision type vgg16<\/p>\n<\/td>\n

\n

34.43562984466553<\/p>\n<\/td>\n<\/tr>\n

\n

Inference half precision type vgg16_bn<\/p>\n<\/td>\n

\n

36.92122936248779<\/p>\n<\/td>\n<\/tr>\n

\n

Inference half precision type vgg19_bn<\/p>\n<\/td>\n

\n

43.144264221191406<\/p>\n<\/td>\n<\/tr>\n

\n

Inference half precision type vgg19<\/p>\n<\/td>\n

\n

40.5385684967041<\/p>\n<\/td>\n<\/tr>\n

\n

Inference half precision type mobilenet_v3_large<\/p>\n<\/td>\n

\n

5.350713729858398<\/p>\n<\/td>\n<\/tr>\n

\n

Inference half precision type mobilenet_v3_small<\/p>\n<\/td>\n

\n

4.016985893249512<\/p>\n<\/td>\n<\/tr>\n

\n

Inference half precision type shufflenet_v2_x0_5<\/p>\n<\/td>\n

\n

5.079126358032227<\/p>\n<\/td>\n<\/tr>\n

\n

Inference half precision type shufflenet_v2_x1_0<\/p>\n<\/td>\n

\n

5.593156814575195<\/p>\n<\/td>\n<\/tr>\n

\n

Inference half precision type shufflenet_v2_x1_5<\/p>\n<\/td>\n

\n

5.649552345275879<\/p>\n<\/td>\n<\/tr>\n

\n

Inference half precision type shufflenet_v2_x2_0<\/p>\n<\/td>\n

\n

5.355663299560547<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type mnasnet0_5<\/p>\n<\/td>\n

\n

50.2386999130249<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type mnasnet0_75<\/p>\n<\/td>\n

\n

80.66896915435791<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type mnasnet1_0<\/p>\n<\/td>\n

\n

103.32422733306885<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type mnasnet1_3<\/p>\n<\/td>\n

\n

154.6230697631836<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type resnet18<\/p>\n<\/td>\n

\n

337.94031620025635<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type resnet34<\/p>\n<\/td>\n

\n

677.7706575393677<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type resnet50<\/p>\n<\/td>\n

\n

789.9243211746216<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type resnet101<\/p>\n<\/td>\n

\n

1484.3351316452026<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type resnet152<\/p>\n<\/td>\n

\n

2170.570478439331<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type resnext50_32x4d<\/p>\n<\/td>\n

\n

877.3719882965088<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type resnext101_32x8d<\/p>\n<\/td>\n

\n

3652.4944639205933<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type wide_resnet50_2<\/p>\n<\/td>\n

\n

2154.612874984741<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type wide_resnet101_2<\/p>\n<\/td>\n

\n

4176.522083282471<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type densenet121<\/p>\n<\/td>\n

\n

607.8699731826782<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type densenet169<\/p>\n<\/td>\n

\n

744.6409797668457<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type densenet201<\/p>\n<\/td>\n

\n

962.677731513977<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type densenet161<\/p>\n<\/td>\n

\n

1759.772515296936<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type squeezenet1_0<\/p>\n<\/td>\n

\n

164.3690824508667<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type squeezenet1_1<\/p>\n<\/td>\n

\n

78.70647430419922<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type vgg11<\/p>\n<\/td>\n

\n

1362.6095294952393<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type vgg11_bn<\/p>\n<\/td>\n

\n

1387.2539138793945<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type vgg13<\/p>\n<\/td>\n

\n

2006.0230445861816<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type vgg13_bn<\/p>\n<\/td>\n

\n

2047.526364326477<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type vgg16<\/p>\n<\/td>\n

\n

2702.2086429595947<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type vgg16_bn<\/p>\n<\/td>\n

\n

2747.241234779358<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type vgg19_bn<\/p>\n<\/td>\n

\n

3447.1724700927734<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type vgg19<\/p>\n<\/td>\n

\n

3397.990345954895<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type mobilenet_v3_large<\/p>\n<\/td>\n

\n

84.65698719024658<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type mobilenet_v3_small<\/p>\n<\/td>\n

\n

29.816465377807617<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type shufflenet_v2_x0_5<\/p>\n<\/td>\n

\n

27.401342391967773<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type shufflenet_v2_x1_0<\/p>\n<\/td>\n

\n

48.322744369506836<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type shufflenet_v2_x1_5<\/p>\n<\/td>\n

\n

82.22103118896484<\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type shufflenet_v2_x2_0<\/p>\n<\/td>\n

\n

141.7021369934082<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type mnasnet0_5<\/p>\n<\/td>\n

\n

12.988653182983398<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type mnasnet0_75<\/p>\n<\/td>\n

\n

22.422199249267578<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type mnasnet1_0<\/p>\n<\/td>\n

\n

30.056486129760742<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type mnasnet1_3<\/p>\n<\/td>\n

\n

46.953935623168945<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type resnet18<\/p>\n<\/td>\n

\n

118.04479122161865<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type resnet34<\/p>\n<\/td>\n

\n

231.52336597442627<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type resnet50<\/p>\n<\/td>\n

\n

268.63497734069824<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type resnet101<\/p>\n<\/td>\n

\n

495.2010440826416<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type resnet152<\/p>\n<\/td>\n

\n

726.4922094345093<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type resnext50_32x4d<\/p>\n<\/td>\n

\n

291.47679328918457<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type resnext101_32x8d<\/p>\n<\/td>\n

\n

1055.10901927948<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type wide_resnet50_2<\/p>\n<\/td>\n

\n

690.6917667388916<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type wide_resnet101_2<\/p>\n<\/td>\n

\n

1347.5529861450195<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type densenet121<\/p>\n<\/td>\n

\n

224.35829639434814<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type densenet169<\/p>\n<\/td>\n

\n

268.9145278930664<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type densenet201<\/p>\n<\/td>\n

\n

343.1972026824951<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type densenet161<\/p>\n<\/td>\n

\n

635.866231918335<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type squeezenet1_0<\/p>\n<\/td>\n

\n

61.92759037017822<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type squeezenet1_1<\/p>\n<\/td>\n

\n

27.009410858154297<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type vgg11<\/p>\n<\/td>\n

\n

462.3375129699707<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type vgg11_bn<\/p>\n<\/td>\n

\n

468.4495782852173<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type vgg13<\/p>\n<\/td>\n

\n

692.8219032287598<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type vgg13_bn<\/p>\n<\/td>\n

\n

703.3538103103638<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type vgg16<\/p>\n<\/td>\n

\n

924.4353818893433<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type vgg16_bn<\/p>\n<\/td>\n

\n

936.5075063705444<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type vgg19_bn<\/p>\n<\/td>\n

\n

1169.098300933838<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type vgg19<\/p>\n<\/td>\n

\n

1156.3771772384644<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type mobilenet_v3_large<\/p>\n<\/td>\n

\n

24.2356014251709<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type mobilenet_v3_small<\/p>\n<\/td>\n

\n

8.85490894317627<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type shufflenet_v2_x0_5<\/p>\n<\/td>\n

\n

6.360034942626953<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type shufflenet_v2_x1_0<\/p>\n<\/td>\n

\n

14.301743507385254<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type shufflenet_v2_x1_5<\/p>\n<\/td>\n

\n

24.863481521606445<\/p>\n<\/td>\n<\/tr>\n

\n

Inference double precision type shufflenet_v2_x2_0<\/p>\n<\/td>\n

\n

43.8505744934082<\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<\/div>\n<\/div>\n<\/details>\n

\u0417\u0430\u043a\u043b\u044e\u0447\u0435\u043d\u0438\u0435<\/h2>\n

\u041d\u043e\u0432\u0430\u044f \u0432\u0438\u0434\u0435\u043e\u043a\u0430\u0440\u0442\u0430 \u043f\u043e\u043a\u0430\u0437\u0430\u043b\u0430 \u0441\u0435\u0431\u044f \u044d\u0444\u0444\u0435\u043a\u0442\u0438\u0432\u043d\u044b\u043c \u0440\u0435\u0448\u0435\u043d\u0438\u0435\u043c \u0434\u043b\u044f \u0432\u044b\u043f\u043e\u043b\u043d\u0435\u043d\u0438\u044f \u0440\u0430\u0437\u043b\u0438\u0447\u043d\u044b\u0445 \u0440\u0430\u0431\u043e\u0447\u0438\u0445 \u0437\u0430\u0434\u0430\u0447. \u0411\u043b\u0430\u0433\u043e\u0434\u0430\u0440\u044f \u0441\u0432\u043e\u0438\u043c \u043a\u043e\u043c\u043f\u0430\u043a\u0442\u043d\u044b\u043c \u0440\u0430\u0437\u043c\u0435\u0440\u0430\u043c \u043e\u043d\u0430 \u043e\u0442\u043b\u0438\u0447\u043d\u043e \u043f\u043e\u0434\u043e\u0439\u0434\u0435\u0442 \u0434\u043b\u044f \u043c\u043e\u0449\u043d\u044b\u0445 \u043a\u043e\u043c\u043f\u044c\u044e\u0442\u0435\u0440\u043e\u0432 \u0444\u043e\u0440\u043c-\u0444\u0430\u043a\u0442\u043e\u0440\u0430 SFF (Small Form Factor). \u0422\u0430\u043a\u0436\u0435 \u0441\u0442\u043e\u0438\u0442 \u043e\u0442\u043c\u0435\u0442\u0438\u0442\u044c, \u0447\u0442\u043e 6144 \u044f\u0434\u0440\u0430 CUDA \u0438 20 \u0413\u0411 \u043f\u0430\u043c\u044f\u0442\u0438 \u0441\u043e 160-\u0440\u0430\u0437\u0440\u044f\u0434\u043d\u043e\u0439 \u0448\u0438\u043d\u043e\u0439 \u0434\u0435\u043b\u0430\u044e\u0442 \u044d\u0442\u0443 \u043a\u0430\u0440\u0442\u0443 \u043e\u0434\u043d\u043e\u0439 \u0438\u0437 \u0441\u0430\u043c\u044b\u0445 \u043f\u0440\u043e\u0438\u0437\u0432\u043e\u0434\u0438\u0442\u0435\u043b\u044c\u043d\u044b\u0445 \u043d\u0430 \u0440\u044b\u043d\u043a\u0435. \u041f\u0440\u0438 \u044d\u0442\u043e\u043c \u043d\u0438\u0437\u043a\u043e\u0435 TDP \u0432 70 \u0412\u0442 \u043f\u043e\u0437\u0432\u043e\u043b\u044f\u0435\u0442 \u0441\u043d\u0438\u0437\u0438\u0442\u044c \u0437\u0430\u0442\u0440\u0430\u0442\u044b \u043d\u0430 \u044d\u043d\u0435\u0440\u0433\u043e\u043f\u043e\u0442\u0440\u0435\u0431\u043b\u0435\u043d\u0438\u0435. \u0427\u0435\u0442\u044b\u0440\u0435 \u043f\u043e\u0440\u0442\u0430 Mini-DisplayPort \u043f\u043e\u0437\u0432\u043e\u043b\u044f\u044e\u0442 \u0438\u0441\u043f\u043e\u043b\u044c\u0437\u043e\u0432\u0430\u0442\u044c \u043a\u0430\u0440\u0442\u0443 \u0441 \u043d\u0435\u0441\u043a\u043e\u043b\u044c\u043a\u0438\u043c\u0438 \u043c\u043e\u043d\u0438\u0442\u043e\u0440\u0430\u043c\u0438 \u0438\u043b\u0438 \u0432 \u043a\u0430\u0447\u0435\u0441\u0442\u0432\u0435 \u0440\u0435\u0448\u0435\u043d\u0438\u044f \u0434\u043b\u044f \u043c\u043d\u043e\u0433\u043e\u043a\u0430\u043d\u0430\u043b\u044c\u043d\u043e\u0439 \u0433\u0440\u0430\u0444\u0438\u043a\u0438.<\/p>\n

RTX 4000 SFF Ada \u043f\u0440\u0435\u0434\u0441\u0442\u0430\u0432\u043b\u044f\u0435\u0442 \u0441\u043e\u0431\u043e\u0439 \u0437\u043d\u0430\u0447\u0438\u0442\u0435\u043b\u044c\u043d\u044b\u0439 \u043f\u0440\u043e\u0433\u0440\u0435\u0441\u0441 \u043f\u043e \u0441\u0440\u0430\u0432\u043d\u0435\u043d\u0438\u044e \u0441 \u043f\u0440\u0435\u0434\u044b\u0434\u0443\u0449\u0438\u043c\u0438 \u043f\u043e\u043a\u043e\u043b\u0435\u043d\u0438\u044f\u043c\u0438, \u043e\u0431\u0435\u0441\u043f\u0435\u0447\u0438\u0432\u0430\u0435\u0442 \u043f\u0440\u043e\u0438\u0437\u0432\u043e\u0434\u0438\u0442\u0435\u043b\u044c\u043d\u043e\u0441\u0442\u044c, \u044d\u043a\u0432\u0438\u0432\u0430\u043b\u0435\u043d\u0442\u043d\u0443\u044e \u043a\u0430\u0440\u0442\u0435 \u0441 \u0443\u0434\u0432\u043e\u0435\u043d\u043d\u043e\u0439 \u043f\u043e\u0442\u0440\u0435\u0431\u043b\u044f\u0435\u043c\u043e\u0439 \u043c\u043e\u0449\u043d\u043e\u0441\u0442\u044c\u044e. \u0411\u043b\u0430\u0433\u043e\u0434\u0430\u0440\u044f \u043e\u0442\u0441\u0443\u0442\u0441\u0442\u0432\u0438\u044e \u0440\u0430\u0437\u044a\u0435\u043c\u0430 \u043f\u0438\u0442\u0430\u043d\u0438\u044f PCIe RTX 4000 SFF Ada \u043b\u0435\u0433\u043a\u043e \u0438\u043d\u0442\u0435\u0433\u0440\u0438\u0440\u043e\u0432\u0430\u0442\u044c \u0432 \u0440\u0430\u0431\u043e\u0447\u0438\u0435 \u0441\u0442\u0430\u043d\u0446\u0438\u0438 \u0441 \u043d\u0438\u0437\u043a\u0438\u043c \u044d\u043d\u0435\u0440\u0433\u043e\u043f\u043e\u0442\u0440\u0435\u0431\u043b\u0435\u043d\u0438\u0435\u043c \u043f\u0440\u0438 \u0441\u043e\u0445\u0440\u0430\u043d\u0435\u043d\u0438\u0438 \u0432\u044b\u0441\u043e\u043a\u043e\u0439 \u043f\u0440\u043e\u0438\u0437\u0432\u043e\u0434\u0438\u0442\u0435\u043b\u044c\u043d\u043e\u0441\u0442\u0438.<\/p>\n\n

\n

\u0410\u0440\u0435\u043d\u0434\u0443\u0439\u0442\u0435 \u0432\u044b\u0434\u0435\u043b\u0435\u043d\u043d\u044b\u0435 \u0438 \u0432\u0438\u0440\u0442\u0443\u0430\u043b\u044c\u043d\u044b\u0435 GPU \u0441\u0435\u0440\u0432\u0435\u0440\u044b \u0441 \u043f\u0440\u043e\u0444\u0435\u0441\u0441\u0438\u043e\u043d\u0430\u043b\u044c\u043d\u044b\u043c\u0438 \u0433\u0440\u0430\u0444\u0438\u0447\u0435\u0441\u043a\u0438\u043c\u0438 \u043a\u0430\u0440\u0442\u0430\u043c\u0438 NVIDIA RTX A5000 \/ A4000<\/a> \u0432 \u043d\u0430\u0434\u0435\u0436\u043d\u044b\u0445 \u0434\u0430\u0442\u0430-\u0446\u0435\u043d\u0442\u0440\u0430\u0445 \u043a\u043b\u0430\u0441\u0441\u0430 TIER III \u0432 \u041c\u043e\u0441\u043a\u0432\u0435 \u0438 \u041d\u0438\u0434\u0435\u0440\u043b\u0430\u043d\u0434\u0430\u0445. \u041f\u0440\u0438\u043d\u0438\u043c\u0430\u0435\u043c \u043e\u043f\u043b\u0430\u0442\u0443 \u0437\u0430 \u0443\u0441\u043b\u0443\u0433\u0438 HOSTKEY \u0432 \u041d\u0438\u0434\u0435\u0440\u043b\u0430\u043d\u0434\u0430\u0445 \u0432 \u0440\u0443\u0431\u043b\u044f\u0445 \u043d\u0430 \u0441\u0447\u0435\u0442 \u0440\u043e\u0441\u0441\u0438\u0439\u0441\u043a\u043e\u0439 \u043a\u043e\u043c\u043f\u0430\u043d\u0438\u0438. \u041e\u043f\u043b\u0430\u0442\u0430 \u0441 \u043f\u043e\u043c\u043e\u0449\u044c\u044e \u0431\u0430\u043d\u043a\u043e\u0432\u0441\u043a\u0438\u0445 \u043a\u0430\u0440\u0442, \u0432 \u0442\u043e\u043c \u0447\u0438\u0441\u043b\u0435 \u0438 \u043a\u0430\u0440\u0442\u043e\u0439 \u041c\u0418\u0420, \u0431\u0430\u043d\u043a\u043e\u0432\u0441\u043a\u043e\u0433\u043e \u043f\u0435\u0440\u0435\u0432\u043e\u0434\u0430 \u0438 \u044d\u043b\u0435\u043a\u0442\u0440\u043e\u043d\u043d\u044b\u0445 \u0434\u0435\u043d\u0435\u0433. <\/p>\n<\/blockquote>\n<\/div>\n<\/div>\n<\/div>\n

<\/div>\n


\u0441\u0441\u044b\u043b\u043a\u0430 \u043d\u0430 \u043e\u0440\u0438\u0433\u0438\u043d\u0430\u043b \u0441\u0442\u0430\u0442\u044c\u0438 https:\/\/habr.com\/ru\/companies\/hostkey\/articles\/741450\/<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"

<\/div>\n
\n
\n
\n
\n
<\/figure>\n

\u0412 \u0430\u043f\u0440\u0435\u043b\u0435 \u043a\u043e\u043c\u043f\u0430\u043d\u0438\u044f NVIDIA \u0432\u044b\u043f\u0443\u0441\u0442\u0438\u043b\u0430 \u043d\u0430 \u0440\u044b\u043d\u043e\u043a \u043d\u043e\u0432\u044b\u0439 \u043f\u0440\u043e\u0434\u0443\u043a\u0442 \u2014 \u0433\u0440\u0430\u0444\u0438\u0447\u0435\u0441\u043a\u0438\u0439 \u043f\u0440\u043e\u0446\u0435\u0441\u0441\u043e\u0440 \u043c\u0430\u043b\u043e\u0433\u043e \u0444\u043e\u0440\u043c-\u0444\u0430\u043a\u0442\u043e\u0440\u0430 RTX A4000 ADA, \u043f\u0440\u0435\u0434\u043d\u0430\u0437\u043d\u0430\u0447\u0435\u043d\u043d\u044b\u0439 \u0434\u043b\u044f \u043f\u0440\u0438\u043c\u0435\u043d\u0435\u043d\u0438\u044f \u0432 \u0440\u0430\u0431\u043e\u0447\u0438\u0445 \u0441\u0442\u0430\u043d\u0446\u0438\u044f\u0445. \u042d\u0442\u043e\u0442 \u043f\u0440\u043e\u0446\u0435\u0441\u0441\u043e\u0440 \u043f\u0440\u0438\u0448\u0435\u043b \u043d\u0430 \u0441\u043c\u0435\u043d\u0443 A2000 \u0438 \u043c\u043e\u0436\u0435\u0442 \u0431\u044b\u0442\u044c \u0438\u0441\u043f\u043e\u043b\u044c\u0437\u043e\u0432\u0430\u043d \u0434\u043b\u044f \u0432\u044b\u043f\u043e\u043b\u043d\u0435\u043d\u0438\u044f \u0441\u043b\u043e\u0436\u043d\u044b\u0445 \u0437\u0430\u0434\u0430\u0447, \u0432 \u0442\u043e\u043c \u0447\u0438\u0441\u043b\u0435 \u0434\u043b\u044f \u043d\u0430\u0443\u0447\u043d\u043e-\u0438\u0441\u0441\u043b\u0435\u0434\u043e\u0432\u0430\u0442\u0435\u043b\u044c\u0441\u043a\u0438\u0445 \u0438 \u0438\u043d\u0436\u0435\u043d\u0435\u0440\u043d\u044b\u0445 \u0440\u0430\u0441\u0447\u0435\u0442\u043e\u0432 \u0438 \u0434\u043b\u044f \u0432\u0438\u0437\u0443\u0430\u043b\u0438\u0437\u0430\u0446\u0438\u0438 \u0434\u0430\u043d\u043d\u044b\u0445.<\/p>\n

RTX A4000 ADA \u043e\u0441\u043d\u0430\u0449\u0435\u043d\u0430 6144 \u044f\u0434\u0440\u0430\u043c\u0438 CUDA, 192 \u0442\u0435\u043d\u0437\u043e\u0440\u0430\u043c\u0438 \u0438 48 \u044f\u0434\u0440\u0430\u043c\u0438 RT, \u043e\u043f\u0435\u0440\u0430\u0442\u0438\u0432\u043d\u043e\u0439 \u043f\u0430\u043c\u044f\u0442\u044c\u044e GDDR6 ECC VRAM \u043e\u0431\u044a\u0435\u043c\u043e\u043c 20 \u0413\u0431. \u041e\u0434\u043d\u043e \u0438\u0437 \u043a\u043b\u044e\u0447\u0435\u0432\u044b\u0445 \u043f\u0440\u0435\u0438\u043c\u0443\u0449\u0435\u0441\u0442\u0432 \u043d\u043e\u0432\u043e\u0433\u043e \u0433\u0440\u0430\u0444\u0438\u0447\u0435\u0441\u043a\u043e\u0433\u043e \u043f\u0440\u043e\u0446\u0435\u0441\u0441\u043e\u0440\u0430 \u2014 \u0435\u0433\u043e \u044d\u043d\u0435\u0440\u0433\u043e\u044d\u0444\u0444\u0435\u043a\u0442\u0438\u0432\u043d\u043e\u0441\u0442\u044c: RTX A4000 ADA \u043f\u043e\u0442\u0440\u0435\u0431\u043b\u044f\u0435\u0442 \u0432\u0441\u0435\u0433\u043e 70 \u0412\u0442, \u0447\u0442\u043e \u0441\u043d\u0438\u0436\u0430\u0435\u0442 \u0437\u0430\u0442\u0440\u0430\u0442\u044b \u043d\u0430 \u044d\u043b\u0435\u043a\u0442\u0440\u043e\u044d\u043d\u0435\u0440\u0433\u0438\u044e \u0438 \u0443\u043c\u0435\u043d\u044c\u0448\u0430\u0435\u0442 \u0442\u0435\u043f\u043b\u043e\u0432\u044b\u0434\u0435\u043b\u0435\u043d\u0438\u0435 \u0432 \u0441\u0438\u0441\u0442\u0435\u043c\u0435. \u0413\u0440\u0430\u0444\u0438\u0447\u0435\u0441\u043a\u0438\u0439 \u043f\u0440\u043e\u0446\u0435\u0441\u0441\u043e\u0440 \u0442\u0430\u043a\u0436\u0435 \u043f\u043e\u0437\u0432\u043e\u043b\u044f\u0435\u0442 \u0443\u043f\u0440\u0430\u0432\u043b\u044f\u0442\u044c \u043d\u0435\u0441\u043a\u043e\u043b\u044c\u043a\u0438\u043c\u0438 \u0434\u0438\u0441\u043f\u043b\u0435\u044f\u043c\u0438 \u0431\u043b\u0430\u0433\u043e\u0434\u0430\u0440\u044f \u043f\u043e\u0434\u043a\u043b\u044e\u0447\u0435\u043d\u0438\u044e 4x Mini-DisplayPort 1.4a.<\/p>\n

<\/figure>\n

\u041f\u0440\u0438 \u0441\u0440\u0430\u0432\u043d\u0435\u043d\u0438\u0438 \u0433\u0440\u0430\u0444\u0438\u0447\u0435\u0441\u043a\u0438\u0445 \u043f\u0440\u043e\u0446\u0435\u0441\u0441\u043e\u0440\u043e\u0432 RTX 4000 SFF Ada \u0441 \u0434\u0440\u0443\u0433\u0438\u043c\u0438 \u0443\u0441\u0442\u0440\u043e\u0439\u0441\u0442\u0432\u0430\u043c\u0438 \u0442\u043e\u0433\u043e \u0436\u0435 \u043a\u043b\u0430\u0441\u0441\u0430 \u043c\u043e\u0436\u043d\u043e \u043e\u0442\u043c\u0435\u0442\u0438\u0442\u044c, \u0447\u0442\u043e \u043f\u0440\u0438 \u0440\u0430\u0431\u043e\u0442\u0435 \u0432 \u0440\u0435\u0436\u0438\u043c\u0435 \u043e\u0434\u0438\u043d\u0430\u0440\u043d\u043e\u0439 \u0442\u043e\u0447\u043d\u043e\u0441\u0442\u0438 \u0434\u0430\u043d\u043d\u044b\u0439 \u043f\u0440\u043e\u0434\u0443\u043a\u0442 \u043f\u043e\u043a\u0430\u0437\u044b\u0432\u0430\u0435\u0442 \u043f\u0440\u043e\u0438\u0437\u0432\u043e\u0434\u0438\u0442\u0435\u043b\u044c\u043d\u043e\u0441\u0442\u044c, \u0430\u043d\u0430\u043b\u043e\u0433\u0438\u0447\u043d\u0443\u044e \u043f\u043e\u0441\u043b\u0435\u0434\u043d\u0435\u043c\u0443 \u043f\u043e\u043a\u043e\u043b\u0435\u043d\u0438\u044e \u0433\u0440\u0430\u0444\u0438\u0447\u0435\u0441\u043a\u043e\u0433\u043e \u043f\u0440\u043e\u0446\u0435\u0441\u0441\u043e\u0440\u0430 RTX A4000, \u043a\u043e\u0442\u043e\u0440\u044b\u0439 \u043f\u043e\u0442\u0440\u0435\u0431\u043b\u044f\u0435\u0442 \u0432\u0434\u0432\u043e\u0435 \u0431\u043e\u043b\u044c\u0448\u0435 \u044d\u043d\u0435\u0440\u0433\u0438\u0438 (140 \u0412\u0442 \u043f\u0440\u043e\u0442\u0438\u0432 70 \u0412\u0442).\u00a0<\/p>\n

<\/figure>\n

RTX 4000 SFF Ada \u043f\u043e\u0441\u0442\u0440\u043e\u0435\u043d\u0430 \u043d\u0430 \u0430\u0440\u0445\u0438\u0442\u0435\u043a\u0442\u0443\u0440\u0435 Ada Lovelace \u0438 \u0442\u0435\u0445\u043f\u0440\u043e\u0446\u0435\u0441\u0441\u0435 5 \u043d\u043c. \u042d\u0442\u043e \u043f\u043e\u0437\u0432\u043e\u043b\u044f\u0435\u0442 \u0438\u0441\u043f\u043e\u043b\u044c\u0437\u043e\u0432\u0430\u0442\u044c \u044f\u0434\u0440\u0430 Tensor Core \u043d\u043e\u0432\u043e\u0433\u043e \u043f\u043e\u043a\u043e\u043b\u0435\u043d\u0438\u044f \u0438 \u044f\u0434\u0440\u0430 \u0442\u0440\u0430\u0441\u0441\u0438\u0440\u043e\u0432\u043a\u0438 \u043b\u0443\u0447\u0435\u0439, \u043a\u043e\u0442\u043e\u0440\u044b\u0435 \u0437\u043d\u0430\u0447\u0438\u0442\u0435\u043b\u044c\u043d\u043e \u043f\u043e\u0432\u044b\u0448\u0430\u044e\u0442 \u043f\u0440\u043e\u0438\u0437\u0432\u043e\u0434\u0438\u0442\u0435\u043b\u044c\u043d\u043e\u0441\u0442\u044c, \u043e\u0431\u0435\u0441\u043f\u0435\u0447\u0438\u0432\u0430\u044f \u0431\u043e\u043b\u0435\u0435 \u0431\u044b\u0441\u0442\u0440\u0443\u044e \u0438 \u044d\u0444\u0444\u0435\u043a\u0442\u0438\u0432\u043d\u0443\u044e \u0440\u0430\u0431\u043e\u0442\u0443 \u0441 \u0442\u0440\u0430\u0441\u0441\u0438\u0440\u043e\u0432\u043a\u043e\u0439 \u043b\u0443\u0447\u0435\u0439 \u0438 \u0442\u0435\u043d\u0437\u043e\u0440\u043d\u044b\u043c\u0438 \u044f\u0434\u0440\u0430\u043c\u0438, \u0447\u0435\u043c RTX A4000. \u041a\u0440\u043e\u043c\u0435 \u0442\u043e\u0433\u043e, RTX 4000 SFF Ada \u0443\u043f\u0430\u043a\u043e\u0432\u0430\u043d \u0432 \u043d\u0435\u0431\u043e\u043b\u044c\u0448\u043e\u0439 \u043a\u043e\u0440\u043f\u0443\u0441 \u2014 \u0434\u043b\u0438\u043d\u0430 \u043a\u0430\u0440\u0442\u044b 168 \u043c\u043c, \u0442\u043e\u043b\u0449\u0438\u043d\u0430 \u0440\u0430\u0432\u043d\u0430 \u0434\u0432\u0443\u043c \u0441\u043b\u043e\u0442\u0430\u043c \u0440\u0430\u0441\u0448\u0438\u0440\u0435\u043d\u0438\u044f.<\/p>\n

<\/figure>\n

\u0423\u043b\u0443\u0447\u0448\u0435\u043d\u0438\u0435 \u044f\u0434\u0435\u0440 \u0442\u0440\u0430\u0441\u0441\u0438\u0440\u043e\u0432\u043a\u0438 \u043b\u0443\u0447\u0435\u0439 \u043e\u0431\u0435\u0441\u043f\u0435\u0447\u0438\u0432\u0430\u0435\u0442 \u044d\u0444\u0444\u0435\u043a\u0442\u0438\u0432\u043d\u0443\u044e \u0440\u0430\u0431\u043e\u0442\u0443 \u0432 \u0441\u0440\u0435\u0434\u0430\u0445, \u0433\u0434\u0435 \u0438\u0441\u043f\u043e\u043b\u044c\u0437\u0443\u0435\u0442\u0441\u044f \u044d\u0442\u0430 \u0442\u0435\u0445\u043d\u043e\u043b\u043e\u0433\u0438\u044f, \u0442\u0430\u043a\u0438\u0445 \u043a\u0430\u043a 3D-\u0434\u0438\u0437\u0430\u0439\u043d \u0438 \u0440\u0435\u043d\u0434\u0435\u0440\u0438\u043d\u0433. \u041e\u0431\u044a\u0435\u043c \u043f\u0430\u043c\u044f\u0442\u0438 \u043d\u043e\u0432\u043e\u0433\u043e GPU (20 \u0413\u0431) \u043f\u043e\u0437\u0432\u043e\u043b\u044f\u0435\u0442 \u0441\u043f\u0440\u0430\u0432\u043b\u044f\u0442\u044c\u0441\u044f \u0441 \u0431\u043e\u043b\u044c\u0448\u0438\u043c\u0438 \u0441\u0440\u0435\u0434\u0430\u043c\u0438.<\/p>\n

<\/figure>\n

\u0421\u043e\u0433\u043b\u0430\u0441\u043d\u043e \u0437\u0430\u044f\u0432\u043b\u0435\u043d\u0438\u044f\u043c \u043f\u0440\u043e\u0438\u0437\u0432\u043e\u0434\u0438\u0442\u0435\u043b\u044f<\/u><\/a>, \u0442\u0435\u043d\u0437\u043e\u0440\u043d\u044b\u0435 \u044f\u0434\u0440\u0430 \u0447\u0435\u0442\u0432\u0435\u0440\u0442\u043e\u0433\u043e \u043f\u043e\u043a\u043e\u043b\u0435\u043d\u0438\u044f \u043e\u0431\u0435\u0441\u043f\u0435\u0447\u0438\u0432\u0430\u044e\u0442 \u0432\u044b\u0441\u043e\u043a\u0443\u044e \u043f\u0440\u043e\u0438\u0437\u0432\u043e\u0434\u0438\u0442\u0435\u043b\u044c\u043d\u043e\u0441\u0442\u044c \u0432\u044b\u0447\u0438\u0441\u043b\u0435\u043d\u0438\u0439 \u0418\u0418 \u2014 \u0434\u0432\u0443\u043a\u0440\u0430\u0442\u043d\u043e\u0435 \u0443\u0432\u0435\u043b\u0438\u0447\u0435\u043d\u0438\u0435 \u043f\u0440\u043e\u0438\u0437\u0432\u043e\u0434\u0438\u0442\u0435\u043b\u044c\u043d\u043e\u0441\u0442\u0438 \u043f\u043e \u0441\u0440\u0430\u0432\u043d\u0435\u043d\u0438\u044e \u0441 \u043f\u0440\u0435\u0434\u044b\u0434\u0443\u0449\u0438\u043c \u043f\u043e\u043a\u043e\u043b\u0435\u043d\u0438\u0435\u043c. \u041d\u043e\u0432\u044b\u0435 \u0442\u0435\u043d\u0437\u043e\u0440\u043d\u044b\u0435 \u044f\u0434\u0440\u0430 \u043f\u043e\u0434\u0434\u0435\u0440\u0436\u0438\u0432\u0430\u044e\u0442 \u0443\u0441\u043a\u043e\u0440\u0435\u043d\u0438\u0435 FP8. \u042d\u0442\u0430 \u043e\u0441\u043e\u0431\u0435\u043d\u043d\u043e\u0441\u0442\u044c \u043d\u043e\u0432\u043e\u0433\u043e \u0433\u0440\u0430\u0444\u0438\u0447\u0435\u0441\u043a\u043e\u0433\u043e \u043f\u0440\u043e\u0446\u0435\u0441\u0441\u043e\u0440\u0430 \u043c\u043e\u0436\u0435\u0442 \u0445\u043e\u0440\u043e\u0448\u043e \u043f\u043e\u0434\u043e\u0439\u0442\u0438 \u0442\u0435\u043c, \u043a\u0442\u043e \u0440\u0430\u0437\u0440\u0430\u0431\u0430\u0442\u044b\u0432\u0430\u0435\u0442 \u0438 \u0440\u0430\u0437\u0432\u0435\u0440\u0442\u044b\u0432\u0430\u0435\u0442 \u043c\u043e\u0434\u0435\u043b\u0438 \u0418\u0418 \u0432 \u0442\u0430\u043a\u0438\u0445 \u0441\u0440\u0435\u0434\u0430\u0445, \u043a\u0430\u043a \u0433\u0435\u043d\u043e\u043c\u0438\u043a\u0430 \u0438 \u043a\u043e\u043c\u043f\u044c\u044e\u0442\u0435\u0440\u043d\u043e\u0435 \u0437\u0440\u0435\u043d\u0438\u0435.<\/p>\n

\u0422\u0430\u043a\u0436\u0435 \u0441\u0442\u043e\u0438\u0442 \u043e\u0442\u043c\u0435\u0442\u0438\u0442\u044c, \u0447\u0442\u043e \u0443\u0432\u0435\u043b\u0438\u0447\u0435\u043d\u0438\u0435 \u043a\u043e\u043b\u0438\u0447\u0435\u0441\u0442\u0432\u0430 \u043c\u0435\u0445\u0430\u043d\u0438\u0437\u043c\u043e\u0432 \u043a\u043e\u0434\u0438\u0440\u043e\u0432\u0430\u043d\u0438\u044f \u0438 \u0434\u0435\u043a\u043e\u0434\u0438\u0440\u043e\u0432\u0430\u043d\u0438\u044f \u0434\u0435\u043b\u0430\u0435\u0442 RTX 4000 SFF Ada \u0445\u043e\u0440\u043e\u0448\u0438\u043c \u0440\u0435\u0448\u0435\u043d\u0438\u0435\u043c \u0434\u043b\u044f \u043c\u0443\u043b\u044c\u0442\u0438\u043c\u0435\u0434\u0438\u0439\u043d\u044b\u0445 \u0440\u0430\u0431\u043e\u0447\u0438\u0445 \u043d\u0430\u0433\u0440\u0443\u0437\u043e\u043a, \u0442\u0430\u043a\u0438\u0445 \u043a\u0430\u043a \u0440\u0430\u0431\u043e\u0442\u0430 \u0441 \u0432\u0438\u0434\u0435\u043e.<\/p>\n

\u0422\u0435\u0445\u043d\u0438\u0447\u0435\u0441\u043a\u0438\u0435 \u0445\u0430\u0440\u0430\u043a\u0442\u0435\u0440\u0438\u0441\u0442\u0438\u043a\u0438 \u0432\u0438\u0434\u0435\u043e\u043a\u0430\u0440\u0442 NVIDIA RTX A4000 \u0438 RTX A5000, RTX 3090<\/strong><\/p>\n

\n
\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
\n

\n<\/td>\n

\n

RTX A4000 ADA\u00a0<\/strong><\/p>\n<\/td>\n

\n

NVIDIA RTX A4000<\/strong><\/p>\n<\/td>\n

\n

NVIDIA RTX A5000<\/strong><\/p>\n<\/td>\n

\n

RTX 3090<\/strong><\/p>\n<\/td>\n<\/tr>\n

\n

\u0410\u0440\u0445\u0438\u0442\u0435\u043a\u0442\u0443\u0440\u0430<\/strong><\/p>\n<\/td>\n

\n

Ada Lovelace<\/p>\n<\/td>\n

\n

Ampere<\/p>\n<\/td>\n

\n

Ampere<\/p>\n<\/td>\n

\n

Ampere<\/p>\n<\/td>\n<\/tr>\n

\n

\u0422\u0435\u0445\u043f\u0440\u043e\u0446\u0435\u0441\u0441<\/strong><\/p>\n<\/td>\n

\n

5 \u043d\u043c<\/p>\n<\/td>\n

\n

8 \u043d\u043c<\/p>\n<\/td>\n

\n

8 \u043d\u043c<\/p>\n<\/td>\n

\n

8 \u043d\u043c<\/p>\n<\/td>\n<\/tr>\n

\n

\u0413\u0440\u0430\u0444\u0438\u0447\u0435\u0441\u043a\u0438\u0439 \u043f\u0440\u043e\u0446\u0435\u0441\u0441\u043e\u0440<\/strong><\/p>\n<\/td>\n

\n

AD104<\/p>\n<\/td>\n

\n

GA102<\/p>\n<\/td>\n

\n

GA104<\/p>\n<\/td>\n

\n

GA102<\/p>\n<\/td>\n<\/tr>\n

\n

\u041a\u043e\u043b\u0438\u0447\u0435\u0441\u0442\u0432\u043e \u0442\u0440\u0430\u043d\u0437\u0438\u0441\u0442\u043e\u0440\u043e\u0432 (\u043c\u043b\u043d)<\/strong><\/p>\n<\/td>\n

\n

35,800<\/p>\n<\/td>\n

\n

17,400<\/p>\n<\/td>\n

\n

28,300<\/p>\n<\/td>\n

\n

28,300<\/p>\n<\/td>\n<\/tr>\n

\n

\u041f\u0440\u043e\u043f\u0443\u0441\u043a\u043d\u0430\u044f \u0441\u043f\u043e\u0441\u043e\u0431\u043d\u043e\u0441\u0442\u044c \u043f\u0430\u043c\u044f\u0442\u0438 (\u0413\u0431\/\u0441)<\/strong><\/p>\n<\/td>\n

\n

280.0<\/p>\n<\/td>\n

\n

448<\/p>\n<\/td>\n

\n

768<\/p>\n<\/td>\n

\n

936.2<\/p>\n<\/td>\n<\/tr>\n

\n

\u0420\u0430\u0437\u0440\u044f\u0434\u043d\u043e\u0441\u0442\u044c \u0448\u0438\u043d\u044b \u0432\u0438\u0434\u0435\u043e\u043f\u0430\u043c\u044f\u0442\u0438 (\u0431\u0438\u0442)<\/strong><\/p>\n<\/td>\n

\n

160<\/p>\n<\/td>\n

\n

256<\/p>\n<\/td>\n

\n

384<\/p>\n<\/td>\n

\n

384<\/p>\n<\/td>\n<\/tr>\n

\n

\u041f\u0430\u043c\u044f\u0442\u044c GPU (\u0413\u0431)<\/strong><\/p>\n<\/td>\n

\n

20<\/p>\n<\/td>\n

\n

16<\/p>\n<\/td>\n

\n

24<\/p>\n<\/td>\n

\n

24<\/p>\n<\/td>\n<\/tr>\n

\n

\u0422\u0438\u043f \u043f\u0430\u043c\u044f\u0442\u0438<\/strong><\/p>\n<\/td>\n

\n

GDDR6<\/p>\n<\/td>\n

\n

GDDR6<\/p>\n<\/td>\n

\n

GDDR6<\/p>\n<\/td>\n

\n

GDDR6X<\/p>\n<\/td>\n<\/tr>\n

\n

\u042f\u0434\u0440\u0430 CUDA<\/strong><\/p>\n<\/td>\n

\n

6,144<\/p>\n<\/td>\n

\n

6 144<\/p>\n<\/td>\n

\n

8192<\/p>\n<\/td>\n

\n

10496<\/p>\n<\/td>\n<\/tr>\n

\n

\u0422\u0435\u043d\u0437\u043e\u0440\u043d\u044b\u0435 \u044f\u0434\u0440\u0430<\/strong><\/p>\n<\/td>\n

\n

192<\/p>\n<\/td>\n

\n

192<\/p>\n<\/td>\n

\n

256<\/p>\n<\/td>\n

\n

328<\/p>\n<\/td>\n<\/tr>\n

\n

\u042f\u0434\u0440\u0430 RT<\/strong><\/p>\n<\/td>\n

\n

48<\/p>\n<\/td>\n

\n

48<\/p>\n<\/td>\n

\n

64<\/p>\n<\/td>\n

\n

82<\/p>\n<\/td>\n<\/tr>\n

\n

SP perf (\u0442\u0435\u0440\u0430\u0444\u043b\u043e\u043f\u0441)<\/strong><\/p>\n<\/td>\n

\n

19.2<\/p>\n<\/td>\n

\n

19,2\u00a0<\/p>\n<\/td>\n

\n

27,8<\/p>\n<\/td>\n

\n

35,6\u00a0<\/p>\n<\/td>\n<\/tr>\n

\n

RT Core performance (\u0442\u0435\u0440\u0430\u0444\u043b\u043e\u043f\u0441)<\/strong><\/p>\n<\/td>\n

\n

44.3<\/p>\n<\/td>\n

\n

37,4<\/p>\n<\/td>\n

\n

54,2<\/p>\n<\/td>\n

\n

69,5<\/p>\n<\/td>\n<\/tr>\n

\n

Tensor performance (\u0442\u0435\u0440\u0430\u0444\u043b\u043e\u043f\u0441)<\/strong><\/p>\n<\/td>\n

\n

306.8<\/p>\n<\/td>\n

\n

153,4<\/p>\n<\/td>\n

\n

222,2<\/p>\n<\/td>\n

\n

285<\/p>\n<\/td>\n<\/tr>\n

\n

\u041c\u0430\u043a\u0441\u0438\u043c\u0430\u043b\u044c\u043d\u0430\u044f \u043c\u043e\u0449\u043d\u043e\u0441\u0442\u044c (\u0412\u0442)<\/strong><\/p>\n<\/td>\n

\n

70<\/p>\n<\/td>\n

\n

140<\/p>\n<\/td>\n

\n

230<\/p>\n<\/td>\n

\n

350<\/p>\n<\/td>\n<\/tr>\n

\n

\u0418\u043d\u0442\u0435\u0440\u0444\u0435\u0439\u0441<\/strong><\/p>\n<\/td>\n

\n

PCIe 4.0 x 16<\/p>\n<\/td>\n

\n

PCI-E 4.0 x16<\/p>\n<\/td>\n

\n

PCI-E 4.0 x16<\/p>\n<\/td>\n

\n

PCIe 4.0 x16<\/p>\n<\/td>\n<\/tr>\n

\n

\u0420\u0430\u0437\u044a\u0435\u043c\u044b<\/strong><\/p>\n<\/td>\n

\n

4x Mini DisplayPort 1.4a<\/p>\n<\/td>\n

\n

\u0414\u041f 1.4 (4)<\/p>\n<\/td>\n

\n

\u0414\u041f 1.4 (4)<\/p>\n<\/td>\n

\n

\u0414\u041f 1.4 (4)<\/p>\n<\/td>\n<\/tr>\n

\n

\u0424\u043e\u0440\u043c-\u0444\u0430\u043a\u0442\u043e\u0440<\/strong><\/p>\n<\/td>\n

\n

2 \u0441\u043b\u043e\u0442\u0430<\/p>\n<\/td>\n

\n

1 \u0441\u043b\u043e\u0442<\/p>\n<\/td>\n

\n

2 \u0441\u043b\u043e\u0442\u0430<\/p>\n<\/td>\n

\n

2-3 \u0441\u043b\u043e\u0442\u0430<\/p>\n<\/td>\n<\/tr>\n

\n

\u041f\u0440\u043e\u0433\u0440\u0430\u043c\u043c\u043d\u043e\u0435 \u043e\u0431\u0435\u0441\u043f\u0435\u0447\u0435\u043d\u0438\u0435 vGPU<\/strong><\/p>\n<\/td>\n

\n

\u043d\u0435\u0442<\/p>\n<\/td>\n

\n

\u043d\u0435\u0442<\/p>\n<\/td>\n

\n

\u0435\u0441\u0442\u044c \u043d\u0435\u043e\u0433\u0440\u0430\u043d\u0438\u0447\u0435\u043d\u043d\u043e<\/p>\n<\/td>\n

\n

\u0435\u0441\u0442\u044c \u0441 \u043e\u0433\u0440\u0430\u043d\u0438\u0447\u0435\u043d\u0438\u044f\u043c\u0438<\/p>\n<\/td>\n<\/tr>\n

\n

Nvlink<\/strong><\/p>\n<\/td>\n

\n

\u043d\u0435\u0442<\/p>\n<\/td>\n

\n

\u043d\u0435\u0442<\/p>\n<\/td>\n

\n

2x RTX A5000<\/p>\n<\/td>\n

\n

\u0435\u0441\u0442\u044c<\/p>\n<\/td>\n<\/tr>\n

\n

\u041f\u043e\u0434\u0434\u0435\u0440\u0436\u043a\u0430 CUDA<\/strong><\/p>\n<\/td>\n

\n

11.6<\/p>\n<\/td>\n

\n

8.6<\/p>\n<\/td>\n

\n

8.6<\/p>\n<\/td>\n

\n

8.6<\/p>\n<\/td>\n<\/tr>\n

\n

\u041f\u043e\u0434\u0434\u0435\u0440\u0436\u043a\u0430 VULKAN<\/strong><\/p>\n<\/td>\n

\n

1.3<\/p>\n<\/td>\n

\n

\u0435\u0441\u0442\u044c<\/p>\n<\/td>\n

\n

\u0435\u0441\u0442\u044c<\/p>\n<\/td>\n

\n

\u0435\u0441\u0442\u044c, 1.2<\/p>\n<\/td>\n<\/tr>\n

\n

\u0426\u0435\u043d\u0430 (\u0440\u0443\u0431.)<\/strong><\/p>\n<\/td>\n

\n

100 000<\/p>\n<\/td>\n

\n

125 000\u00a0<\/p>\n<\/td>\n

\n

220 000\u00a0<\/p>\n<\/td>\n

\n

100 000<\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<\/div>\n

\u041e\u043f\u0438\u0441\u0430\u043d\u0438\u0435 \u0442\u0435\u0441\u0442\u043e\u0432\u043e\u0439 \u0441\u0440\u0435\u0434\u044b<\/strong><\/p>\n

\n
\n\n\n\n\n\n\n\n\n
\n

\n<\/td>\n

\n

RTX A4000 ADA<\/p>\n<\/td>\n

\n

RTX A4000<\/p>\n<\/td>\n<\/tr>\n

\n

\u041f\u0440\u043e\u0446\u0435\u0441\u0441\u043e\u0440<\/p>\n<\/td>\n

\n

AMD Ryzen 9 5950X 3.4GHz (16 cores)<\/p>\n<\/td>\n

\n

OctaCore Intel Xeon E-2288G, 3,5 GHz<\/p>\n<\/td>\n<\/tr>\n

\n

\u041e\u043f\u0435\u0440\u0430\u0442\u0438\u0432\u043d\u0430\u044f \u043f\u0430\u043c\u044f\u0442\u044c<\/p>\n<\/td>\n

\n

4x 32 Gb DDR4 ECC SO-DIMM<\/p>\n<\/td>\n

\n

2x 32 GB DDR4-3200 ECC DDR4 SDRAM 1600 \u041c\u0413\u0446<\/p>\n<\/td>\n<\/tr>\n

\n

\u041d\u0430\u043a\u043e\u043f\u0438\u0442\u0435\u043b\u044c<\/p>\n<\/td>\n

\n

1Tb NVMe SSD<\/p>\n<\/td>\n

\n

Samsung SSD 980 PRO 1TB<\/p>\n<\/td>\n<\/tr>\n

\n

\u041c\u0430\u0442\u0435\u0440\u0438\u043d\u0441\u043a\u0430\u044f \u043f\u043b\u0430\u0442\u0430<\/p>\n<\/td>\n

\n

ASRock X570D4I-2T<\/p>\n<\/td>\n

\n

Asus P11C-I Series<\/p>\n<\/td>\n<\/tr>\n

\n

\u041e\u043f\u0435\u0440\u0430\u0446\u0438\u043e\u043d\u043d\u0430\u044f \u0441\u0438\u0441\u0442\u0435\u043c\u0430\u00a0<\/p>\n<\/td>\n

\n

Microsoft Windows 10<\/p>\n<\/td>\n

\n

Microsoft Windows 10<\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<\/div>\n

\u0420\u0435\u0437\u0443\u043b\u044c\u0442\u0430\u0442\u044b \u0432 \u0442\u0435\u0441\u0442\u0430\u0445<\/h2>\n

V-Ray 5 Benchmark<\/strong><\/p>\n

\n
Points scored<\/figcaption><\/div>\n<\/figure>\n
\n
Points scored<\/figcaption><\/div>\n<\/figure>\n

\u0422\u0435\u0441\u0442\u044b V-Ray GPU CUDA \u0438 RTX \u043f\u043e\u0437\u0432\u043e\u043b\u044f\u044e\u0442 \u0438\u0437\u043c\u0435\u0440\u0438\u0442\u044c \u043e\u0442\u043d\u043e\u0441\u0438\u0442\u0435\u043b\u044c\u043d\u0443\u044e \u043f\u0440\u043e\u0438\u0437\u0432\u043e\u0434\u0438\u0442\u0435\u043b\u044c\u043d\u043e\u0441\u0442\u044c GPU \u043f\u0440\u0438 \u0440\u0435\u043d\u0434\u0435\u0440\u0438\u043d\u0433\u0435. GPU RTX A4000 \u043d\u0435\u0437\u043d\u0430\u0447\u0438\u0442\u0435\u043b\u044c\u043d\u043e \u0443\u0441\u0442\u0443\u043f\u0430\u0435\u0442 \u043f\u043e \u043f\u0440\u043e\u0438\u0437\u0432\u043e\u0434\u0438\u0442\u0435\u043b\u044c\u043d\u043e\u0441\u0442\u0438 RTX A4000 ADA (4% \u0438 11% \u0441\u043e\u043e\u0442\u0432\u0435\u0442\u0441\u0442\u0432\u0435\u043d\u043d\u043e).\u00a0<\/p>\n

\u041c\u0430\u0448\u0438\u043d\u043d\u043e\u0435 \u043e\u0431\u0443\u0447\u0435\u043d\u0438\u0435<\/h2>\n

\u00ab\u0421\u043e\u0431\u0430\u043a\u0438 \u043f\u0440\u043e\u0442\u0438\u0432 \u043a\u043e\u0448\u0435\u043a\u00bb<\/strong><\/p>\n

\u0414\u043b\u044f \u0441\u0440\u0430\u0432\u043d\u0435\u043d\u0438\u044f \u043f\u0440\u043e\u0438\u0437\u0432\u043e\u0434\u0438\u0442\u0435\u043b\u044c\u043d\u043e\u0441\u0442\u0438 GPU \u0434\u043b\u044f \u043d\u0435\u0439\u0440\u043e\u0441\u0435\u0442\u0435\u0439 \u043c\u044b \u0438\u0441\u043f\u043e\u043b\u044c\u0437\u0443\u0435\u043c \u043d\u0430\u0431\u043e\u0440 \u0434\u0430\u043d\u043d\u044b\u0445 \u00ab\u0421\u043e\u0431\u0430\u043a\u0438 \u043f\u0440\u043e\u0442\u0438\u0432 \u043a\u043e\u0448\u0435\u043a\u00bb \u2014 \u0442\u0435\u0441\u0442 \u0430\u043d\u0430\u043b\u0438\u0437\u0438\u0440\u0443\u0435\u0442 \u0441\u043e\u0434\u0435\u0440\u0436\u0438\u043c\u043e\u0435 \u0444\u043e\u0442\u043e\u0433\u0440\u0430\u0444\u0438\u0438 \u0438 \u0440\u0430\u0437\u043b\u0438\u0447\u0430\u0435\u0442, \u0438\u0437\u043e\u0431\u0440\u0430\u0436\u0435\u043d\u0430 \u043d\u0430 \u0444\u043e\u0442\u043e \u043a\u043e\u0448\u043a\u0430 \u0438\u043b\u0438 \u0441\u043e\u0431\u0430\u043a\u0430. \u0412\u0441\u0435 \u043d\u0435\u043e\u0431\u0445\u043e\u0434\u0438\u043c\u044b\u0435 \u0438\u0441\u0445\u043e\u0434\u043d\u044b\u0435 \u0434\u0430\u043d\u043d\u044b\u0435 \u043d\u0430\u0445\u043e\u0434\u044f\u0442\u0441\u044f \u0437\u0434\u0435\u0441\u044c<\/u><\/a>. \u041c\u044b \u0437\u0430\u043f\u0443\u0441\u043a\u0430\u043b\u0438 \u044d\u0442\u043e\u0442 \u0442\u0435\u0441\u0442 \u043d\u0430 \u0440\u0430\u0437\u043d\u044b\u0445 GPU \u0438 \u0432 \u0440\u0430\u0437\u043b\u0438\u0447\u043d\u044b\u0445 \u043e\u0431\u043b\u0430\u0447\u043d\u044b\u0445 \u0441\u0435\u0440\u0432\u0438\u0441\u0430\u0445 \u0438 \u043f\u043e\u043b\u0443\u0447\u0438\u043b\u0438 \u0441\u043b\u0435\u0434\u0443\u044e\u0449\u0438\u0435 \u0440\u0435\u0437\u0443\u043b\u044c\u0442\u0430\u0442\u044b:<\/p>\n

\n
Points scored<\/figcaption><\/div>\n<\/figure>\n

\u0412 \u044d\u0442\u043e\u043c \u0442\u0435\u0441\u0442\u0435 \u0440\u0435\u0437\u0443\u043b\u044c\u0442\u0430\u0442 RTX A4000 ADA \u043d\u0435\u0437\u043d\u0430\u0447\u0438\u0442\u0435\u043b\u044c\u043d\u043e \u043f\u0440\u0435\u0432\u0437\u043e\u0448\u0435\u043b RTX A4000 (9%), \u043d\u043e \u0441\u043b\u0435\u0434\u0443\u0435\u0442 \u043f\u043e\u043c\u043d\u0438\u0442\u044c \u043e \u043d\u0435\u0431\u043e\u043b\u044c\u0448\u043e\u043c \u0440\u0430\u0437\u043c\u0435\u0440\u0435 \u0438 \u043d\u0438\u0437\u043a\u043e\u043c \u044d\u043d\u0435\u0440\u0433\u043e\u043f\u043e\u0442\u0440\u0435\u0431\u043b\u0435\u043d\u0438\u0438 \u043d\u043e\u0432\u043e\u0433\u043e GPU.<\/p>\n

AI-Benchmark<\/u><\/a><\/p>\n

AI-Benchmark \u043f\u043e\u0437\u0432\u043e\u043b\u044f\u0435\u0442 \u0438\u0437\u043c\u0435\u0440\u0438\u0442\u044c \u043f\u0440\u043e\u0438\u0437\u0432\u043e\u0434\u0438\u0442\u0435\u043b\u044c\u043d\u043e\u0441\u0442\u044c \u0443\u0441\u0442\u0440\u043e\u0439\u0441\u0442\u0432\u0430 \u0432\u043e \u0432\u0440\u0435\u043c\u044f \u0432\u044b\u043f\u043e\u043b\u043d\u0435\u043d\u0438\u044f \u0437\u0430\u0434\u0430\u0447\u0438 \u0432\u044b\u0432\u043e\u0434\u0430 AI-\u043c\u043e\u0434\u0435\u043b\u0435\u0439. \u0415\u0434\u0438\u043d\u0438\u0446\u044b \u0438\u0437\u043c\u0435\u0440\u0435\u043d\u0438\u044f \u043c\u043e\u0433\u0443\u0442 \u0437\u0430\u0432\u0438\u0441\u0435\u0442\u044c \u043e\u0442 \u0442\u0435\u0441\u0442\u0430, \u043d\u043e \u043e\u0431\u044b\u0447\u043d\u043e \u044d\u0442\u043e \u043a\u043e\u043b\u0438\u0447\u0435\u0441\u0442\u0432\u043e \u043e\u043f\u0435\u0440\u0430\u0446\u0438\u0439 \u0432 \u0441\u0435\u043a\u0443\u043d\u0434\u0443 (OPS) \u0438\u043b\u0438 \u043a\u043e\u043b\u0438\u0447\u0435\u0441\u0442\u0432\u043e \u0438\u0437\u043e\u0431\u0440\u0430\u0436\u0435\u043d\u0438\u0439 \u0432 \u0441\u0435\u043a\u0443\u043d\u0434\u0443 (FPS).<\/p>\n

\n
Points scored<\/figcaption><\/div>\n<\/figure>\n
\n
\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
<\/td>\n\n

RTX A4000<\/strong><\/p>\n<\/td>\n

\n

RTX A4000 ADA<\/strong><\/p>\n<\/td>\n<\/tr>\n

\n

1\/19. MobileNet-V2<\/strong><\/p>\n<\/td>\n

\n

1.1 \u2014 inference | batch=50, size=224×224: 38.5 \u00b1 2.4 ms<\/p>\n

1.2 \u2014 training | batch=50, size=224×224: 109 \u00b1 4 ms<\/p>\n<\/td>\n

\n

1.1 \u2014 inference | batch=50, size=224×224: 53.5 \u00b1 0.7 ms<\/p>\n

1.2 \u2014 training | batch=50, size=224×224: 130.1 \u00b1 0.6 ms<\/p>\n<\/td>\n<\/tr>\n

\n

2\/19. Inception-V3<\/strong><\/p>\n<\/td>\n

\n

2.1 \u2014 inference | batch=20, size=346×346: 36.1 \u00b1 1.8 ms<\/p>\n

2.2 \u2014 training | batch=20, size=346×346: 137.4 \u00b1 0.6 ms<\/p>\n<\/td>\n

\n

2.1 \u2014 inference | batch=20, size=346×346: 36.8 \u00b1 1.1 ms<\/p>\n

2.2 \u2014 training | batch=20, size=346×346: 147.5 \u00b1 0.8 ms<\/p>\n<\/td>\n<\/tr>\n

\n

3\/19. Inception-V4<\/strong><\/p>\n<\/td>\n

\n

3.1 \u2014 inference | batch=10, size=346×346: 34.0 \u00b1 0.9 ms<\/p>\n

3.2 \u2014 training | batch=10, size=346×346: 139.4 \u00b1 1.0 ms<\/p>\n<\/td>\n

\n

3.1 \u2014 inference | batch=10, size=346×346: 33.0 \u00b1 0.8 ms<\/p>\n

3.2 \u2014 training | batch=10, size=346×346: 135.7 \u00b1 0.9 ms<\/p>\n<\/td>\n<\/tr>\n

\n

4\/19. Inception-ResNet-V2<\/strong><\/p>\n<\/td>\n

\n

4.1 \u2014 inference | batch=10, size=346×346: 45.7 \u00b1 0.6 ms<\/p>\n

4.2 \u2014 training | batch=8, size=346×346: 153.4 \u00b1 0.8 ms<\/p>\n<\/td>\n

\n

4.1 \u2014 inference batch=10, size=346×346: 33.6 \u00b1 0.7 ms<\/p>\n

4.2 \u2014 training batch=8, size=346×346: 132 \u00b1 1 ms<\/p>\n<\/td>\n<\/tr>\n

\n

5\/19. ResNet-V2-50<\/strong><\/p>\n<\/td>\n

\n

5.1 \u2014 inference | batch=10, size=346×346: 25.3 \u00b1 0.5 ms<\/p>\n

5.2 \u2014 training | batch=10, size=346×346: 91.1 \u00b1 0.8 ms<\/p>\n<\/td>\n

\n

5.1 \u2014 inference | batch=10, size=346×346: 26.1 \u00b1 0.5 ms<\/p>\n

5.2 \u2014 training | batch=10, size=346×346: 92.3 \u00b1 0.6 ms<\/p>\n<\/td>\n<\/tr>\n

\n

6\/19. ResNet-V2-152<\/strong><\/p>\n<\/td>\n

\n

6.1 \u2014 inference | batch=10, size=256×256: 32.4 \u00b1 0.5 ms<\/p>\n

6.2 \u2014 training | batch=10, size=256×256: 131.4 \u00b1 0.7 ms<\/p>\n<\/td>\n

\n

6.1 \u2014 inference | batch=10, size=256×256: 23.7 \u00b1 0.6 ms<\/p>\n

6.2 \u2014 training | batch=10, size=256×256: 107.1 \u00b1 0.9 ms<\/p>\n<\/td>\n<\/tr>\n

\n

7\/19. VGG-16<\/strong><\/p>\n<\/td>\n

\n

7.1 \u2014 inference | batch=20, size=224×224: 54.9 \u00b1 0.9 ms<\/p>\n

7.2 \u2014 training | batch=2, size=224×224: 83.6 \u00b1 0.7 ms<\/p>\n<\/td>\n

\n

7.1 \u2014 inference | batch=20, size=224×224: 66.3 \u00b1 0.9 ms<\/p>\n

7.2 \u2014 training | batch=2, size=224×224: 109.3 \u00b1 0.8 ms<\/p>\n<\/td>\n<\/tr>\n

\n

8\/19. SRCNN 9-5-5<\/strong><\/p>\n<\/td>\n

\n

8.1 \u2014 inference | batch=10, size=512×512: 51.5 \u00b1 0.9 ms<\/p>\n

8.2 \u2014 inference | batch=1, size=1536×1536: 45.7 \u00b1 0.9 ms<\/p>\n

8.3 \u2014 training | batch=10, size=512×512: 183 \u00b1 1 ms<\/p>\n<\/td>\n

\n

8.1 \u2014 inference | batch=10, size=512×512: 59.9 \u00b1 1.6 ms<\/p>\n

8.2 \u2014 inference | batch=1, size=1536×1536: 53.1 \u00b1 0.7 ms<\/p>\n

8.3 \u2014 training | batch=10, size=512×512: 176 \u00b1 2 ms<\/p>\n<\/td>\n<\/tr>\n

\n

9\/19. VGG-19 Super-Res<\/strong><\/p>\n<\/td>\n

\n

9.1 \u2014 inference | batch=10, size=256×256: 99.5 \u00b1 0.8 ms<\/p>\n

9.2 \u2014 inference | batch=1, size=1024×1024: 162 \u00b1 1 ms<\/p>\n

9.3 \u2014 training | batch=10, size=224×224: 204 \u00b1 2 ms<\/p>\n<\/td>\n

\n

\n<\/td>\n<\/tr>\n

\n

10\/19. ResNet-SRGAN<\/strong><\/p>\n<\/td>\n

\n

10.1 \u2014 inference | batch=10, size=512×512: 85.8 \u00b1 0.6 ms<\/p>\n

10.2 \u2014 inference | batch=1, size=1536×1536: 82.4 \u00b1 1.9 ms<\/p>\n

10.3 \u2014 training | batch=5, size=512×512: 133 \u00b1 1 ms<\/p>\n<\/td>\n

\n

10.1 \u2014 inference | batch=10, size=512×512: 98.9 \u00b1 0.8 ms<\/p>\n

10.2 \u2014 inference | batch=1, size=1536×1536: 86.1 \u00b1 0.6 ms<\/p>\n

10.3 \u2014 training | batch=5, size=512×512: 130.9 \u00b1 0.6 ms<\/p>\n<\/td>\n<\/tr>\n

\n

11\/19. ResNet-DPED<\/strong><\/p>\n<\/td>\n

\n

11.1 \u2014 inference | batch=10, size=256×256: 114.9 \u00b1 0.6 ms<\/p>\n

11.2 \u2014 inference | batch=1, size=1024×1024: 182 \u00b1 2 ms<\/p>\n

11.3 \u2014 training | batch=15, size=128×128: 178.1 \u00b1 0.8 ms<\/p>\n<\/td>\n

\n

11.1 \u2014 inference | batch=10, size=256×256: 146.4 \u00b1 0.5 ms<\/p>\n

11.2 \u2014 inference | batch=1, size=1024×1024: 234.3 \u00b1 0.5 ms<\/p>\n

11.3 \u2014 training | batch=15, size=128×128: 234.7 \u00b1 0.6 ms<\/p>\n<\/td>\n<\/tr>\n

\n

12\/19. U-Net<\/strong><\/p>\n<\/td>\n

\n

12.1 \u2014 inference | batch=4, size=512×512: 180.8 \u00b1 0.7 ms<\/p>\n

12.2 \u2014 inference | batch=1, size=1024×1024: 177.0 \u00b1 0.4 ms<\/p>\n

12.3 \u2014 training | batch=4, size=256×256: 198.6 \u00b1 0.5 ms<\/p>\n<\/td>\n

\n

12.1 \u2014 inference | batch=4, size=512×512: 222.9 \u00b1 0.5 ms<\/p>\n

12.2 \u2014 inference | batch=1, size=1024×1024: 220.4 \u00b1 0.6 ms<\/p>\n

12.3 \u2014 training | batch=4, size=256×256: 229.1 \u00b1 0.7 ms<\/p>\n<\/td>\n<\/tr>\n

\n

13\/19. Nvidia-SPADE<\/strong><\/p>\n<\/td>\n

\n

13.1 \u2014 inference | batch=5, size=128×128: 54.5 \u00b1 0.5 ms<\/p>\n

13.2 \u2014 training | batch=1, size=128×128: 103.6 \u00b1 0.6 ms<\/p>\n<\/td>\n

\n

13.1 \u2014 inference | batch=5, size=128×128: 59.6 \u00b1 0.6 ms<\/p>\n

13.2 \u2014 training | batch=1, size=128×128: 94.6 \u00b1 0.6 ms<\/p>\n<\/td>\n<\/tr>\n

\n

14\/19. ICNet<\/strong><\/p>\n<\/td>\n

\n

14.1 \u2014 inference | batch=5, size=1024×1536: 126.3 \u00b1 0.8 ms<\/p>\n

14.2 \u2014 training | batch=10, size=1024×1536: 426 \u00b1 9 ms<\/p>\n<\/td>\n

\n

14.1 \u2014 inference | batch=5, size=1024×1536: 144 \u00b1 4 ms<\/p>\n

14.2 \u2014 training | batch=10, size=1024×1536: 475 \u00b1 17 ms<\/p>\n<\/td>\n<\/tr>\n

\n

15\/19. PSPNet<\/strong><\/p>\n<\/td>\n

\n

15.1 \u2014 inference | batch=5, size=720×720: 249 \u00b1 12 ms<\/p>\n

15.2 \u2014 training | batch=1, size=512×512: 104.6 \u00b1 0.6 ms<\/p>\n<\/td>\n

\n

15.1 \u2014 inference | batch=5, size=720×720: 291.4 \u00b1 0.5 ms<\/p>\n

15.2 \u2014 training | batch=1, size=512×512: 99.8 \u00b1 0.9 ms<\/p>\n<\/td>\n<\/tr>\n

\n

16\/19. DeepLab<\/strong><\/p>\n<\/td>\n

\n

16.1 \u2014 inference | batch=2, size=512×512: 71.7 \u00b1 0.6 ms<\/p>\n

16.2 \u2014 training | batch=1, size=384×384: 84.9 \u00b1 0.5 ms<\/p>\n<\/td>\n

\n

16.1 \u2014 inference | batch=2, size=512×512: 71.5 \u00b1 0.7 ms<\/p>\n

16.2 \u2014 training | batch=1, size=384×384: 69.4 \u00b1 0.6 ms<\/p>\n<\/td>\n<\/tr>\n

\n

17\/19. Pixel-RNN<\/strong><\/p>\n<\/td>\n

\n

17.1 \u2014 inference | batch=50, size=64×64: 299 \u00b1 14 ms<\/p>\n

17.2 \u2014 training | batch=10, size=64×64: 1258 \u00b1 64 ms<\/p>\n<\/td>\n

\n

17.1 \u2014 inference | batch=50, size=64×64: 321 \u00b1 30 ms<\/p>\n

17.2 \u2014 training | batch=10, size=64×64: 1278 \u00b1 74 ms<\/p>\n<\/td>\n<\/tr>\n

\n

18\/19. LSTM-Sentiment<\/strong><\/p>\n<\/td>\n

\n

18.1 \u2014 inference | batch=100, size=1024×300: 395 \u00b1 11 ms<\/p>\n

18.2 \u2014 training | batch=10, size=1024×300: 676 \u00b1 15 ms<\/p>\n<\/td>\n

\n

18.1 \u2014 inference | batch=100, size=1024×300: 345 \u00b1 10 ms<\/p>\n

18.2 \u2014 training | batch=10, size=1024×300: 774 \u00b1 17 ms<\/p>\n<\/td>\n<\/tr>\n

\n

19\/19. GNMT-Translation<\/strong><\/p>\n<\/td>\n

\n

19.1 \u2014 inference | batch=1, size=1×20: 119 \u00b1 2 ms<\/p>\n<\/td>\n

\n

19.1 \u2014 inference | batch=1, size=1×20: 156 \u00b1 1 ms<\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<\/div>\n

\u0420\u0435\u0437\u0443\u043b\u044c\u0442\u0430\u0442\u044b \u044d\u0442\u043e\u0433\u043e \u0442\u0435\u0441\u0442\u0430 \u043f\u043e\u043a\u0430\u0437\u044b\u0432\u0430\u044e\u0442, \u0447\u0442\u043e \u043f\u0440\u043e\u0438\u0437\u0432\u043e\u0434\u0438\u0442\u0435\u043b\u044c\u043d\u043e\u0441\u0442\u044c RTX A4000 \u043d\u0435\u0437\u043d\u0430\u0447\u0438\u0442\u0435\u043b\u044c\u043d\u043e (\u043d\u0430 6%) \u0432\u044b\u0448\u0435, \u0447\u0435\u043c \u0443 RTX A4000 ADA. \u041e\u0434\u043d\u0430\u043a\u043e, \u0441\u0442\u043e\u0438\u0442 \u043e\u0442\u043c\u0435\u0442\u0438\u0442\u044c, \u0447\u0442\u043e \u0440\u0435\u0437\u0443\u043b\u044c\u0442\u0430\u0442\u044b \u0442\u0435\u0441\u0442\u043e\u0432 \u043c\u043e\u0433\u0443\u0442 \u0440\u0430\u0437\u043b\u0438\u0447\u0430\u0442\u044c\u0441\u044f \u0432 \u0437\u0430\u0432\u0438\u0441\u0438\u043c\u043e\u0441\u0442\u0438 \u043e\u0442 \u043a\u043e\u043d\u043a\u0440\u0435\u0442\u043d\u044b\u0445 \u0437\u0430\u0434\u0430\u0447 \u0438 \u0443\u0441\u043b\u043e\u0432\u0438\u0439 \u0440\u0430\u0431\u043e\u0442\u044b.<\/p>\n

PyTorch<\/u><\/a><\/p>\n

\n\u0420\u0435\u0437\u0443\u043b\u044c\u0442\u0430\u0442\u044b \u0441 RTX A 4000<\/summary>\n
\n
\n
\n\n\n\n\n\n
\n

Benchmarking<\/strong><\/p>\n<\/td>\n

\n

Model average train time (ms)<\/strong><\/p>\n<\/td>\n<\/tr>\n

\n

Training double precision type mnasnet0_5<\/p>\n<\/td>\n

\n

62.995805740356445<\/p>\n<\/td>\n<\/tr>\n

\n

Tra<\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<\/div>\n<\/div>\n<\/details>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[],"tags":[],"class_list":["post-349067","post","type-post","status-publish","format-standard","hentry"],"_links":{"self":[{"href":"https:\/\/savepearlharbor.com\/index.php?rest_route=\/wp\/v2\/posts\/349067","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=349067"}],"version-history":[{"count":0,"href":"https:\/\/savepearlharbor.com\/index.php?rest_route=\/wp\/v2\/posts\/349067\/revisions"}],"wp:attachment":[{"href":"https:\/\/savepearlharbor.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=349067"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/savepearlharbor.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=349067"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/savepearlharbor.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=349067"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}