{"id":131,"date":"2023-11-15T15:21:08","date_gmt":"2023-11-15T08:21:08","guid":{"rendered":"https:\/\/xuhuongai.com\/?p=131"},"modified":"2023-11-18T12:22:24","modified_gmt":"2023-11-18T05:22:24","slug":"nvidia-dan-dau-intel-bam-sat-va-google-tut-lai-phia-sau-trong-cuoc-dua-ai-tao-sinh","status":"publish","type":"post","link":"https:\/\/xuhuongai.com\/?p=131","title":{"rendered":"Nvidia d\u1eabn \u0111\u1ea7u, Intel b\u00e1m s\u00e1t v\u00e0 Google t\u1ee5t l\u1ea1i ph\u00eda sau trong cu\u1ed9c \u0111ua AI t\u1ea1o sinh"},"content":{"rendered":"\n
\n

B\u1ea3n tin \u0111\u01b0\u1ee3c d\u1ecbch v\u00e0 t\u00f3m t\u1eaft b\u1edfi n\u1ec1n t\u1ea3ng t\u1ea1o tr\u1ee3 l\u00fd AI – KamiMind<\/a>.<\/p>\nNgu\u1ed3n: Samuel K. Moore, “Google, Intel, Nvidia Battle in Generative AI Training MLPerf training tests put Nvidia ahead, Intel close, and Google well behind<\/a>“, IEEE, 12\/11\/2023.<\/cite><\/blockquote>\n\n\n

\n
\"\"
\u1ea2nh minh h\u1ecda. Ngu\u1ed3n: NVIDIA<\/figcaption><\/figure>\n<\/div>\n\n\n

MLPerf, ti\u00eau chu\u1ea9n \u0111\u00e1nh gi\u00e1 h\u00e0ng \u0111\u1ea7u v\u1ec1 kh\u1ea3 n\u0103ng c\u1ee7a c\u00e1c h\u1ec7 th\u1ed1ng m\u00e1y t\u00ednh trong vi\u1ec7c hu\u1ea5n luy\u1ec7n m\u1ea1ng neural h\u1ecdc m\u00e1y, \u0111\u00e3 ch\u00ednh th\u1ee9c b\u01b0\u1edbc v\u00e0o th\u1eddi \u0111\u1ea1i AI t\u1ea1o sinh. Trong n\u0103m nay, MLPerf \u0111\u00e3 th\u00eam m\u1ed9t b\u00e0i ki\u1ec3m tra \u0111\u1ec3 hu\u1ea5n luy\u1ec7n c\u00e1c m\u00f4 h\u00ecnh ng\u00f4n ng\u1eef l\u1edbn (LLM) nh\u01b0 GPT-3 v\u00e0 g\u1ea7n \u0111\u00e2y \u0111\u00e3 b\u1ed5 sung Stable Diffusion, m\u1ed9t c\u00f4ng c\u1ee5 t\u1ea1o h\u00ecnh \u1ea3nh t\u1eeb v\u0103n b\u1ea3n. C\u00e1c m\u00e1y t\u00ednh s\u1eed d\u1ee5ng c\u00f4ng ngh\u1ec7 c\u1ee7a Intel v\u00e0 Nvidia \u0111\u00e3 tham gia v\u00e0o b\u00e0i ki\u1ec3m tra, v\u1edbi si\u00eau m\u00e1y t\u00ednh 10.000 GPU c\u1ee7a Nvidia tr\u1edf th\u00e0nh m\u00e1y t\u00ednh l\u1edbn nh\u1ea5t t\u1eebng \u0111\u01b0\u1ee3c ki\u1ec3m tra. T\u1ed5ng c\u1ed9ng, c\u00f3 19 c\u00f4ng ty v\u00e0 t\u1ed5 ch\u1ee9c \u0111\u00e3 n\u1ed9p h\u01a1n 200 k\u1ebft qu\u1ea3, cho th\u1ea5y t\u0103ng t\u1ed1c hi\u1ec7u su\u1ea5t l\u00ean g\u1ea5p 2,8 l\u1ea7n trong n\u0103m qua v\u00e0 g\u1ea5p 49 l\u1ea7n k\u1ec3 t\u1eeb khi MLPerf b\u1eaft \u0111\u1ea7u c\u00e1ch \u0111\u00e2y n\u0103m n\u0103m. Nvidia \u0111\u00e3 th\u1ed1ng tr\u1ecb c\u00e1c b\u00e0i ki\u1ec3m tra, nh\u01b0ng si\u00eau m\u00e1y t\u00ednh AI m\u1edbi c\u1ee7a h\u1ecd c\u00f3 t\u00ean Eos v\u1edbi 10.752 GPU \u0111\u00e1ng ch\u00fa \u00fd, ho\u00e0n th\u00e0nh b\u00e0i ki\u1ec3m tra hu\u1ea5n luy\u1ec7n GPT-3 trong ch\u01b0a \u0111\u1ea7y b\u1ed1n ph\u00fat. Intel c\u0169ng \u0111\u00e3 ti\u1ebfn b\u1ed9 b\u1eb1ng c\u00e1ch trang b\u1ecb chip gia t\u0103ng Gaudi 2 c\u1ee7a h\u1ecd v\u1edbi kh\u1ea3 n\u0103ng \u0111i\u1ec3m ph\u1ea9y 8-bit, gi\u00fap gi\u1ea3m th\u1eddi gian hu\u1ea5n luy\u1ec7n cho m\u1ed9t c\u1ee5m gia t\u0103ng 384 \u0111\u01a1n v\u1ecb \u0111i\u1ec3m ph\u1ea9y 103%. Nh\u1eefng ti\u1ebfn b\u1ed9 n\u00e0y trong l\u0129nh v\u1ef1c AI s\u00e1ng t\u1ea1o v\u00e0 th\u1eddi gian hu\u1ea5n luy\u1ec7n nhanh h\u01a1n l\u00e0 r\u1ea5t quan tr\u1ecdng \u0111\u1ec3 c\u1ea3i thi\u1ec7n li\u00ean t\u1ee5c c\u00e1c h\u1ec7 th\u1ed1ng AI.<\/p>\n\n\n\n

B\u1ea3n t\u00f3m t\u1eaft ti\u1ebfng Anh<\/summary>\n

MLPerf, the leading benchmark for computer systems’ ability to train machine-learning neural networks, has entered the generative AI era. This year, MLPerf added a test for training large language models (LLM) like GPT-3 and recently included Stable Diffusion, a text-to-image generator. Intel and Nvidia-powered computers participated in the benchmark, with Nvidia’s 10,000-GPU supercomputer being the largest ever tested. Overall, 19 companies and institutions submitted over 200 results, showing a 2.8-fold performance boost in the past five months and a 49-fold boost since MLPerf began five years ago. Nvidia dominated the benchmarks, but its new 10,752-GPU AI supercomputer called Eos was particularly notable, completing the GPT-3 training benchmark in under four minutes. Intel also made progress by enabling its Gaudi 2 accelerator chip with 8-bit floating-point capabilities, resulting in a 103% reduction in time-to-train for a 384-accelerator cluster. These advancements in generative AI and accelerated training times are crucial for the continued improvement of AI systems.<\/p>\n<\/details>\n\n\n\n

B\u1ea3n d\u1ecbch Anh – Vi\u1ec7t<\/summary>\n

B\u00e0i ki\u1ec3m tra h\u00e0ng \u0111\u1ea7u \u0111\u1ec3 \u0111\u00e1nh gi\u00e1 kh\u1ea3 n\u0103ng c\u1ee7a c\u00e1c h\u1ec7 th\u1ed1ng m\u00e1y t\u00ednh trong vi\u1ec7c hu\u1ea5n luy\u1ec7n m\u1ea1ng neural h\u1ecdc m\u00e1y \u0111\u00e3 ho\u00e0n to\u00e0n b\u01b0\u1edbc v\u00e0o th\u1eddi \u0111\u1ea1i AI t\u1ea1o sinh. Tr\u01b0\u1edbc \u0111\u00f3 trong n\u0103m nay, MLPerf \u0111\u00e3 th\u00eam m\u1ed9t b\u00e0i ki\u1ec3m tra \u0111\u1ec3 hu\u1ea5n luy\u1ec7n c\u00e1c m\u00f4 h\u00ecnh ng\u00f4n ng\u1eef l\u1edbn (LLM), \u0111\u1eb7c bi\u1ec7t l\u00e0 GPT-3. Th\u00e1ng n\u00e0y, h\u1ecd \u0111\u00e3 th\u00eam Stable Diffusion, m\u1ed9t b\u1ed9 sinh v\u0103n b\u1ea3n th\u00e0nh h\u00ecnh \u1ea3nh. C\u00e1c m\u00e1y t\u00ednh s\u1eed d\u1ee5ng c\u00f4ng ngh\u1ec7 Intel v\u00e0 Nvidia \u0111\u00e3 tham gia v\u00e0o b\u00e0i ki\u1ec3m tra m\u1edbi n\u00e0y. V\u00e0 cu\u1ed9c c\u1ea1nh tranh gi\u1eefa hai c\u00f4ng ty n\u00e0y ti\u1ebfp t\u1ee5c trong vi\u1ec7c hu\u1ea5n luy\u1ec7n GPT-3, v\u00e0 l\u1ea7n n\u00e0y h\u1ecd \u0111\u00e3 c\u00f3 s\u1ef1 tham gia c\u1ee7a Google.<\/p>\n\n\n\n

C\u1ea3 ba c\u00f4ng ty \u0111\u1ec1u \u0111\u00e3 s\u1eed d\u1ee5ng c\u00e1c h\u1ec7 th\u1ed1ng l\u1edbn cho nhi\u1ec7m v\u1ee5 n\u00e0y – si\u00eau m\u00e1y t\u00ednh 10.000 GPU c\u1ee7a Nvidia l\u00e0 m\u00e1y t\u00ednh l\u1edbn nh\u1ea5t t\u1eebng \u0111\u01b0\u1ee3c ki\u1ec3m tra – v\u00e0 k\u00edch th\u01b0\u1edbc n\u00e0y l\u00e0 c\u1ea7n thi\u1ebft trong AI s\u00e1ng t\u1ea1o. Ngay c\u1ea3 h\u1ec7 th\u1ed1ng l\u1edbn nh\u1ea5t c\u1ee7a Nvidia c\u0169ng s\u1ebd m\u1ea5t t\u00e1m ng\u00e0y l\u00e0m vi\u1ec7c \u0111\u1ec3 ho\u00e0n th\u00e0nh c\u00f4ng vi\u1ec7c LLM c\u1ee7a n\u00f3.<\/p>\n\n\n\n

T\u1ed5ng c\u1ed9ng, c\u00f3 19 c\u00f4ng ty v\u00e0 t\u1ed5 ch\u1ee9c \u0111\u00e3 n\u1ed9p h\u01a1n 200 k\u1ebft qu\u1ea3, cho th\u1ea5y t\u0103ng t\u1ed1c hi\u1ec7u su\u1ea5t l\u00ean g\u1ea5p 2,8 l\u1ea7n trong n\u0103m qua v\u00e0 g\u1ea5p 49 l\u1ea7n k\u1ec3 t\u1eeb khi MLPerf b\u1eaft \u0111\u1ea7u n\u0103m n\u0103m tr\u01b0\u1edbc.<\/p>\n\n\n\n

Nvidia v\u00e0 Microsoft th\u1eed nghi\u1ec7m “qu\u00e1i v\u1eadt” 10.752 GPU<\/strong><\/p>\n\n\n\n

Nvidia ti\u1ebfp t\u1ee5c l\u00e0m ch\u1ee7 c\u00e1c b\u00e0i ki\u1ec3m tra MLPerf v\u1edbi c\u00e1c h\u1ec7 th\u1ed1ng \u0111\u01b0\u1ee3c t\u1ea1o t\u1eeb GPU H100 c\u1ee7a h\u1ecd. Nh\u01b0ng k\u1ebft qu\u1ea3 t\u1eeb Eos, si\u00eau m\u00e1y t\u00ednh AI m\u1edbi c\u1ee7a c\u00f4ng ty v\u1edbi 10.752 GPU, l\u00e0 \u0111i\u1ec3m nh\u1ea5n. B\u1eb1ng vi\u1ec7c s\u1eed d\u1ee5ng t\u1ea5t c\u1ea3 c\u00e1c GPU \u0111\u00f3 cho c\u00f4ng vi\u1ec7c ki\u1ec3m tra hu\u1ea5n luy\u1ec7n GPT-3, Eos \u0111\u00e3 ho\u00e0n th\u00e0nh c\u00f4ng vi\u1ec7c trong ch\u01b0a \u0111\u1ea7y 4 ph\u00fat. Ph\u1ea7n m\u1ec1m \u0111i\u1ec7n to\u00e1n \u0111\u00e1m m\u00e2y Azure c\u1ee7a Microsoft \u0111\u00e3 th\u1eed nghi\u1ec7m m\u1ed9t h\u1ec7 th\u1ed1ng c\u00f3 k\u00edch th\u01b0\u1edbc t\u01b0\u01a1ng t\u1ef1 v\u00e0 ch\u1ec9 k\u00e9m Eos v\u00e0i gi\u00e2y. (Azure l\u00e0 n\u1ec1n t\u1ea3ng h\u1ed7 tr\u1ee3 l\u1eadp tr\u00ecnh c\u1ee7a GitHub CoPilot v\u00e0 ChatGPT c\u1ee7a OpenAI.)<\/p>\n\n\n\n

GPU c\u1ee7a Eos c\u00f3 kh\u1ea3 n\u0103ng th\u1ef1c hi\u1ec7n t\u1ed5ng c\u1ed9ng 42,6 t\u1ef7 t\u1ef7 ph\u00e9p t\u00ednh h\u00e0ng th\u1eadp ph\u00e2n m\u1ed7i gi\u00e2y (exaflops). V\u00e0 ch\u00fang \u0111\u01b0\u1ee3c k\u1ebft n\u1ed1i v\u1edbi nhau b\u1eb1ng c\u00e1ch s\u1eed d\u1ee5ng k\u1ebft n\u1ed1i m\u1ea1ng n\u1ed9i b\u1ed9 Quantum-2 Infiniband c\u1ee7a Nvidia, c\u00f3 kh\u1ea3 n\u0103ng truy\u1ec1n 1,1 tri\u1ec7u t\u1ef7 byte m\u1ed7i gi\u00e2y. “M\u1ed9t s\u1ed1 t\u1ed1c \u0111\u1ed9 v\u00e0 th\u00f4ng s\u1ed1 n\u00e0y th\u1eadt l\u00e0 \u0111\u00e1ng kinh ng\u1ea1c,” Dave Salvatore, Gi\u00e1m \u0111\u1ed1c ki\u1ec3m tra AI v\u00e0 \u0111i\u1ec7n to\u00e1n \u0111\u00e1m m\u00e2y c\u1ee7a Nvidia, n\u00f3i. “\u0110\u00e2y l\u00e0 m\u1ed9t m\u00e1y t\u00ednh c\u1ef1c k\u1ef3 m\u1ea1nh m\u1ebd.”<\/p>\n\n\n\n

Eos nh\u00e2n ba s\u1ed1 l\u01b0\u1ee3ng GPU H100 \u0111\u00e3 \u0111\u01b0\u1ee3c k\u1ebft h\u1ee3p th\u00e0nh m\u1ed9t m\u00e1y t\u00ednh duy nh\u1ea5t. S\u1ef1 gia t\u0103ng ba l\u1ea7n n\u00e0y \u0111\u00e3 \u0111\u1ea1t \u0111\u01b0\u1ee3c c\u1ea3i ti\u1ebfn hi\u1ec7u su\u1ea5t 2,8 l\u1ea7n, ho\u1eb7c \u0111\u1ea1t hi\u1ec7u qu\u1ea3 t\u1ec9 l\u1ec7 93%. S\u1ef1 gia t\u0103ng hi\u1ec7u su\u1ea5t hi\u1ec7u qu\u1ea3 l\u00e0 y\u1ebfu t\u1ed1 quan tr\u1ecdng \u0111\u1ec3 ti\u1ebfp t\u1ee5c c\u1ea3i thi\u1ec7n AI s\u00e1ng t\u1ea1o, m\u00e0 \u0111\u00e3 t\u0103ng g\u1ea5p m\u01b0\u1eddi l\u1ea7n m\u1ed7i n\u0103m.<\/p>\n\n\n\n

B\u00e0i ki\u1ec3m tra GPT-3 m\u00e0 Eos \u0111\u00e3 tham gia kh\u00f4ng ph\u1ea3i l\u00e0 m\u1ed9t qu\u00e1 tr\u00ecnh hu\u1ea5n luy\u1ec7n ho\u00e0n ch\u1ec9nh c\u1ee7a GPT-3, v\u00ec MLPerf mu\u1ed1n n\u00f3 trong t\u1ea7m tay c\u1ee7a nhi\u1ec1u c\u00f4ng ty. Thay v\u00e0o \u0111\u00f3, n\u00f3 li\u00ean quan \u0111\u1ebfn vi\u1ec7c hu\u1ea5n luy\u1ec7n h\u1ec7 th\u1ed1ng \u0111\u1ebfn m\u1ed9t \u0111i\u1ec3m ki\u1ec3m tra nh\u1ea5t \u0111\u1ecbnh ch\u1ee9ng minh r\u1eb1ng vi\u1ec7c hu\u1ea5n luy\u1ec7n s\u1ebd \u0111\u1ea1t \u0111\u1ed9 ch\u00ednh x\u00e1c c\u1ea7n thi\u1ebft trong th\u1eddi gian \u0111\u1ee7. V\u00e0 nh\u1eefng qu\u00e1 tr\u00ecnh hu\u1ea5n luy\u1ec7n n\u00e0y m\u1ea5t th\u1eddi gian. T\u1eeb vi\u1ec7c tham gia c\u1ee7a Eos trong 4 ph\u00fat, c\u00f3 th\u1ec3 suy ra r\u1eb1ng n\u00f3 s\u1ebd m\u1ea5t t\u00e1m ng\u00e0y \u0111\u1ec3 ho\u00e0n th\u00e0nh qu\u00e1 tr\u00ecnh hu\u1ea5n luy\u1ec7n, v\u00e0 \u0111\u00f3 l\u00e0 tr\u00ean m\u00e1y t\u00ednh AI si\u00eau m\u1ea1nh nh\u1ea5t c\u00f3 th\u1ec3 \u0111\u00e3 \u0111\u01b0\u1ee3c x\u00e2y d\u1ef1ng. M\u1ed9t m\u00e1y t\u00ednh c\u00f3 k\u00edch th\u01b0\u1edbc h\u1ee3p l\u00fd h\u01a1n – 512 H100 – s\u1ebd m\u1ea5t b\u1ed1n th\u00e1ng.<\/p>\n\n\n\n

Intel ti\u1ebfp t\u1ee5c ti\u1ebfn g\u1ea7n h\u01a1n<\/strong><\/p>\n\n\n\n

Intel \u0111\u00e3 n\u1ed9p k\u1ebft qu\u1ea3 cho c\u00e1c h\u1ec7 th\u1ed1ng s\u1eed d\u1ee5ng chip t\u0103ng t\u1ed1c Gaudi 2 v\u00e0 cho nh\u1eefng h\u1ec7 th\u1ed1ng kh\u00f4ng c\u00f3 b\u1ea5t k\u1ef3 b\u1ed9 t\u0103ng t\u1ed1c n\u00e0o, ch\u1ec9 d\u1ef1a v\u00e0o CPU Xeon th\u1ebf h\u1ec7 th\u1ee9 t\u01b0 c\u1ee7a h\u1ecd. Thay \u0111\u1ed5i l\u1edbn so v\u1edbi b\u1ed9 ki\u1ec3m tra hu\u1ea5n luy\u1ec7n tr\u01b0\u1edbc \u0111\u00f3 l\u00e0 c\u00f4ng ty \u0111\u00e3 k\u00edch ho\u1ea1t kh\u1ea3 n\u0103ng 8-bit floating-point (FP8) c\u1ee7a Gaudi 2. Vi\u1ec7c s\u1eed d\u1ee5ng s\u1ed1 h\u1ecdc ch\u00ednh x\u00e1c th\u1ea5p h\u01a1n, ch\u1eb3ng h\u1ea1n nh\u01b0 FP8, \u0111\u00e3 g\u00f3p ph\u1ea7n \u0111\u00e1ng k\u1ec3 v\u00e0o vi\u1ec7c c\u1ea3i thi\u1ec7n hi\u1ec7u su\u1ea5t GPU trong 10 n\u0103m qua. Vi\u1ec7c s\u1eed d\u1ee5ng FP8 trong c\u00e1c ph\u1ea7n c\u1ee7a GPT-3 v\u00e0 c\u00e1c m\u1ea1ng neural bi\u1ebfn \u0111\u1ed5i kh\u00e1c m\u00e0 \u0111\u1ed9 ch\u00ednh x\u00e1c th\u1ea5p c\u1ee7a ch\u00fang kh\u00f4ng \u1ea3nh h\u01b0\u1edfng \u0111\u00e3 ch\u1ee9ng minh gi\u00e1 tr\u1ecb c\u1ee7a n\u00f3 trong k\u1ebft qu\u1ea3 H100 c\u1ee7a Nvidia. B\u00e2y gi\u1edd Gaudi 2 \u0111ang nh\u1eadn \u0111\u01b0\u1ee3c s\u1ef1 th\u00fac \u0111\u1ea9y n\u00e0y.
“Ch\u00fang t\u00f4i \u0111\u00e3 d\u1ef1 t\u00ednh t\u0103ng 90% t\u1eeb vi\u1ec7c b\u1eadt FP8,” Eitan Medina, Gi\u00e1m \u0111\u1ed1c \u0111i\u1ec1u h\u00e0nh t\u1ea1i Habana Labs c\u1ee7a Intel n\u00f3i. “Ch\u00fang t\u00f4i \u0111\u00e3 cung c\u1ea5p h\u01a1n nh\u1eefng g\u00ec \u0111\u00e3 \u0111\u01b0\u1ee3c h\u1ee9a – gi\u1ea3m th\u1eddi gian hu\u1ea5n luy\u1ec7n 103% cho m\u1ed9t c\u1ee5m t\u0103ng t\u1ed1c 384.”<\/p>\n\n\n\n

K\u1ebft qu\u1ea3 m\u1edbi n\u00e0y \u0111\u01b0a h\u1ec7 th\u1ed1ng Gaudi 2 ch\u1ec9 s\u1eed d\u1ee5ng m\u1ed9t ph\u1ea7n nh\u1ecf h\u01a1n m\u1ed9t ph\u1ea7n ba t\u1ed1c \u0111\u1ed9 c\u1ee7a h\u1ec7 th\u1ed1ng Nvidia tr\u00ean c\u01a1 s\u1edf t\u1eebng chip v\u00e0 nhanh g\u1ea5p ba l\u1ea7n so v\u1edbi TPUv5e c\u1ee7a Google. Tr\u00ean b\u00e0i ki\u1ec3m tra t\u1ea1o h\u00ecnh \u1ea3nh m\u1edbi, Gaudi 2 c\u0169ng kho\u1ea3ng m\u1ed9t n\u1eeda t\u1ed1c \u0111\u1ed9 c\u1ee7a H100. GPT-3 l\u00e0 b\u00e0i ki\u1ec3m tra duy nh\u1ea5t \u0111\u01b0\u1ee3c k\u00edch ho\u1ea1t FP8 cho v\u00f2ng n\u00e0y, nh\u01b0ng Medina n\u00f3i r\u1eb1ng \u0111\u1ed9i c\u1ee7a \u00f4ng \u0111ang l\u00e0m vi\u1ec7c \u0111\u1ec3 k\u00edch ho\u1ea1t n\u00f3 cho c\u00e1c b\u00e0i ki\u1ec3m tra kh\u00e1c.<\/p>\n\n\n\n

Medina ti\u1ebfp t\u1ee5c th\u1ec3 hi\u1ec7n quan \u0111i\u1ec3m r\u1eb1ng Gaudi 2 c\u00f3 gi\u00e1 th\u00e0nh th\u1ea5p h\u01a1n \u0111\u00e1ng k\u1ec3 so v\u1edbi H100, v\u00e0 v\u00ec v\u1eady n\u00f3 c\u00f3 l\u1ee3i th\u1ebf v\u1ec1 m\u1ee9c \u0111\u1ed9 gi\u00e1 v\u00e0 hi\u1ec7u su\u1ea5t. Medina k\u1ef3 v\u1ecdng l\u1ee3i th\u1ebf s\u1ebd t\u0103ng l\u00ean v\u1edbi th\u1ebf h\u1ec7 ti\u1ebfp theo c\u1ee7a chip t\u0103ng t\u1ed1c Intel, Gaudi 3. Chip n\u00e0y s\u1ebd \u0111\u01b0\u1ee3c s\u1ea3n xu\u1ea5t h\u00e0ng lo\u1ea1t v\u00e0o n\u0103m 2024 v\u00e0 \u0111\u01b0\u1ee3c x\u00e2y d\u1ef1ng b\u1eb1ng c\u00f9ng quy tr\u00ecnh ch\u1ebf t\u1ea1o b\u00e1n d\u1eabn nh\u01b0 Nvidia H100.<\/p>\n\n\n\n

Ri\u00eang bi\u1ec7t, Intel \u0111\u00e3 n\u1ed9p k\u1ebft qu\u1ea3 cho c\u00e1c h\u1ec7 th\u1ed1ng ch\u1ec9 d\u1ef1a tr\u00ean CPU, m\u1ed9t l\u1ea7n n\u1eefa cho th\u1ea5y th\u1eddi gian hu\u1ea5n luy\u1ec7n t\u1eeb v\u00e0i ph\u00fat \u0111\u1ebfn v\u00e0i gi\u1edd cho m\u1ed9t s\u1ed1 b\u00e0i ki\u1ec3m tra. Ngo\u00e0i b\u00e0i ki\u1ec3m tra MLPerf, Intel c\u0169ng chia s\u1ebb m\u1ed9t s\u1ed1 d\u1eef li\u1ec7u cho th\u1ea5y m\u1ed9t h\u1ec7 th\u1ed1ng Xeon g\u1ed3m 4 n\u00fat, v\u1edbi c\u00e1c chip bao g\u1ed3m \u0111\u1ed9ng c\u01a1 ma tr\u1eadn AMX, c\u00f3 th\u1ec3 \u0111i\u1ec1u ch\u1ec9nh l\u1ea1i b\u1ed9 t\u1ea1o h\u00ecnh \u1ea3nh Stable Diffusion trong ch\u01b0a \u0111\u1ea7y 5 ph\u00fat. \u0110i\u1ec1u ch\u1ec9nh l\u1ea1i l\u00e0 qu\u00e1 tr\u00ecnh l\u00e0m cho m\u1ea1ng neural \u0111\u00e3 \u0111\u01b0\u1ee3c hu\u1ea5n luy\u1ec7n tr\u1edf n\u00ean chuy\u00ean m\u00f4n h\u01a1n \u0111\u1ed1i v\u1edbi m\u1ed9t nhi\u1ec7m v\u1ee5 nh\u1ea5t \u0111\u1ecbnh. V\u00ed d\u1ee5, AI thi\u1ebft k\u1ebf chip c\u1ee7a Nvidia l\u00e0 vi\u1ec7c \u0111i\u1ec1u ch\u1ec9nh l\u1ea1i m\u1ed9t m\u00f4 h\u00ecnh ng\u00f4n ng\u1eef l\u1edbn c\u00f3 t\u00ean l\u00e0 Nemo.<\/p>\n<\/details>\n","protected":false},"excerpt":{"rendered":"

MLPerf, ti\u00eau chu\u1ea9n \u0111\u00e1nh gi\u00e1 h\u00e0ng \u0111\u1ea7u v\u1ec1 kh\u1ea3 n\u0103ng c\u1ee7a c\u00e1c h\u1ec7 th\u1ed1ng m\u00e1y t\u00ednh trong vi\u1ec7c hu\u1ea5n luy\u1ec7n m\u1ea1ng neural h\u1ecdc m\u00e1y, \u0111\u00e3 ch\u00ednh th\u1ee9c b\u01b0\u1edbc v\u00e0o th\u1eddi \u0111\u1ea1i AI t\u1ea1o sinh.<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[5],"tags":[47,42,49,48],"class_list":["post-131","post","type-post","status-publish","format-standard","hentry","category-ai-news","tag-intel","tag-mlperf","tag-mo-hinh-ngon-ngu-lon-llm","tag-nvidia"],"_links":{"self":[{"href":"https:\/\/xuhuongai.com\/index.php?rest_route=\/wp\/v2\/posts\/131","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/xuhuongai.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/xuhuongai.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/xuhuongai.com\/index.php?rest_route=\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/xuhuongai.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=131"}],"version-history":[{"count":9,"href":"https:\/\/xuhuongai.com\/index.php?rest_route=\/wp\/v2\/posts\/131\/revisions"}],"predecessor-version":[{"id":452,"href":"https:\/\/xuhuongai.com\/index.php?rest_route=\/wp\/v2\/posts\/131\/revisions\/452"}],"wp:attachment":[{"href":"https:\/\/xuhuongai.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=131"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/xuhuongai.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=131"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/xuhuongai.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=131"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}