{"id":260,"date":"2023-11-16T14:29:33","date_gmt":"2023-11-16T07:29:33","guid":{"rendered":"https:\/\/xuhuongai.com\/?p=260"},"modified":"2023-11-18T12:12:24","modified_gmt":"2023-11-18T05:12:24","slug":"nvidia-gioi-thieu-h200-vi-mach-cao-cap-moi-nhat-danh-cho-viec-huan-luyen-mo-hinh-tri-tue-nhan-tao","status":"publish","type":"post","link":"https:\/\/xuhuongai.com\/?p=260","title":{"rendered":"Nvidia gi\u1edbi thi\u1ec7u H200, vi m\u1ea1ch cao c\u1ea5p m\u1edbi nh\u1ea5t d\u00e0nh cho vi\u1ec7c hu\u1ea5n luy\u1ec7n m\u00f4 h\u00ecnh Tr\u00ed tu\u1ec7 Nh\u00e2n t\u1ea1o"},"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: Kif Leswing, “Nvidia unveils H200, its newest high-end chip for training AI models<\/a>“, CNBC, 13\/11\/2023.<\/cite><\/blockquote>\n\n\n

\n
\"\"
Jensen Huang, Ch\u1ee7 t\u1ecbch c\u1ee7a Nvidia, \u0111ang c\u1ea7m vi x\u1eed l\u00fd Grace Hopper si\u00eau chip CPU \u0111\u01b0\u1ee3c s\u1eed d\u1ee5ng cho AI t\u1ea1o sinh t\u1ea1i bu\u1ed5i thuy\u1ebft tr\u00ecnh ch\u00ednh c\u1ee7a Supermicro trong s\u1ef1 ki\u1ec7n Computex 2023.<\/figcaption><\/figure>\n<\/div>\n\n\n

Nvidia \u0111\u00e3 gi\u1edbi thi\u1ec7u chip \u0111\u1ed3 h\u1ecda H200, \u0111\u01b0\u1ee3c thi\u1ebft k\u1ebf cho hu\u1ea5n luy\u1ec7n v\u00e0 tri\u1ec3n khai m\u00f4 h\u00ecnh tr\u00ed tu\u1ec7 nh\u00e2n t\u1ea1o. \u0110\u00e2y l\u00e0 n\u00e2ng c\u1ea5p t\u1eeb H100, chip \u0111\u01b0\u1ee3c OpenAI s\u1eed d\u1ee5ng \u0111\u1ec3 hu\u1ea5n luy\u1ec7n m\u00f4 h\u00ecnh ng\u00f4n ng\u1eef GPT-4. Gi\u00e1 c\u1ee7a H100 dao \u0111\u1ed9ng t\u1eeb 25.000 \u0111\u1ebfn 40.000 \u0111\u00f4 la, v\u00e0 c\u1ea7n h\u00e0ng ngh\u00ecn chip n\u00e0y \u0111\u1ec3 t\u1ea1o ra nh\u1eefng m\u00f4 h\u00ecnh l\u1edbn nh\u1ea5t. GPU tr\u00ed tu\u1ec7 nh\u00e2n t\u1ea1o c\u1ee7a Nvidia \u0111\u00e3 t\u0103ng gi\u00e1 c\u1ed5 phi\u1ebfu c\u1ee7a c\u00f4ng ty l\u00ean h\u01a1n 230% trong n\u0103m 2023. H200 c\u00f3 b\u1ed9 nh\u1edb HBM3 dung l\u01b0\u1ee3ng 141GB, t\u1ea1o ra g\u1ea7n g\u1ea5p \u0111\u00f4i \u0111\u1ea7u ra so v\u1edbi H100. D\u1ef1 ki\u1ebfn H200 s\u1ebd c\u1ea1nh tranh v\u1edbi GPU MI300X c\u1ee7a AMD v\u00e0 t\u01b0\u01a1ng th\u00edch v\u1edbi H100. Nvidia c\u0169ng d\u1ef1 \u0111\u1ecbnh ph\u00e1t h\u00e0nh chip B100 v\u00e0o n\u0103m 2024.<\/p>\n\n\n\n

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

Nvidia has introduced the H200, a GPU designed for training and deploying AI models. It is an upgrade from the H100, which was used by OpenAI for training GPT-4. The H100 chips cost between $25,000 and $40,000, and thousands of them are needed for creating large models. Nvidia’s AI GPUs have driven up the company’s stock by over 230% in 2023. The H200 includes 141GB of next-generation memory and performs inference twice as fast as the H100. It will compete with AMD’s MI300X GPU and is compatible with the H100. Nvidia plans to release the B100 chip based on the Blackwell architecture in 2024.<\/p>\n<\/details>\n\n\n\n

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

Nvidia \u0111\u00e3 gi\u1edbi thi\u1ec7u m\u1eabu chip \u0111\u1ed3 h\u1ecda H200, \u0111\u01b0\u1ee3c thi\u1ebft k\u1ebf \u0111\u1ec3 hu\u1ea5n luy\u1ec7n v\u00e0 tri\u1ec3n khai c\u00e1c m\u00f4 h\u00ecnh tr\u00ed tu\u1ec7 nh\u00e2n t\u1ea1o \u0111ang th\u00fac \u0111\u1ea9y s\u1ef1 ph\u00e1t tri\u1ec3n v\u01b0\u1ee3t b\u1eadc c\u1ee7a tr\u00ed tu\u1ec7 nh\u00e2n t\u1ea1o s\u00e1ng t\u1ea1o.<\/p>\n\n\n\n

GPU m\u1edbi n\u00e0y l\u00e0 m\u1ed9t n\u00e2ng c\u1ea5p t\u1eeb H100, chip m\u00e0 OpenAI \u0111\u00e3 s\u1eed d\u1ee5ng \u0111\u1ec3 hu\u1ea5n luy\u1ec7n m\u00f4 h\u00ecnh ng\u00f4n ng\u1eef l\u1edbn ti\u00ean ti\u1ebfn nh\u1ea5t c\u1ee7a m\u00ecnh, GPT-4. C\u00e1c c\u00f4ng ty l\u1edbn, c\u00e1c startup v\u00e0 c\u00e1c c\u01a1 quan ch\u00ednh ph\u1ee7 \u0111ang c\u1ea1nh tranh \u0111\u1ec3 c\u00f3 \u0111\u01b0\u1ee3c s\u1ed1 l\u01b0\u1ee3ng chip h\u1ea1n ch\u1ebf.<\/p>\n\n\n\n

Theo \u01b0\u1edbc t\u00ednh t\u1eeb Raymond James, gi\u00e1 c\u1ee7a chip H100 dao \u0111\u1ed9ng t\u1eeb 25.000 \u0111\u1ebfn 40.000 \u0111\u00f4 la, v\u00e0 c\u1ea7n h\u00e0ng ngh\u00ecn chip n\u00e0y l\u00e0m vi\u1ec7c c\u00f9ng nhau \u0111\u1ec3 t\u1ea1o ra nh\u1eefng m\u00f4 h\u00ecnh l\u1edbn nh\u1ea5t trong qu\u00e1 tr\u00ecnh “hu\u1ea5n luy\u1ec7n.”<\/p>\n\n\n\n

S\u1ef1 ph\u1ea5n kh\u00edch v\u1edbi GPU tr\u00ed tu\u1ec7 nh\u00e2n t\u1ea1o c\u1ee7a Nvidia \u0111\u00e3 l\u00e0m t\u0103ng gi\u00e1 c\u1ed5 phi\u1ebfu c\u1ee7a c\u00f4ng ty n\u00e0y l\u00ean h\u01a1n 230% trong n\u0103m 2023. Nvidia d\u1ef1 ki\u1ebfn \u200b\u200bc\u00f3 kho\u1ea3ng 16 t\u1ef7 \u0111\u00f4 la doanh thu cho qu\u00fd ba c\u1ee7a n\u0103m t\u00e0i ch\u00ednh, t\u0103ng 170% so v\u1edbi c\u00f9ng k\u1ef3 n\u0103m tr\u01b0\u1edbc.<\/p>\n\n\n\n

C\u1ea3i ti\u1ebfn ch\u00ednh c\u1ee7a H200 l\u00e0 n\u00f3 bao g\u1ed3m b\u1ed9 nh\u1edb “HBM3” th\u1ebf h\u1ec7 ti\u1ebfp theo v\u1edbi dung l\u01b0\u1ee3ng 141GB s\u1ebd gi\u00fap chip th\u1ef1c hi\u1ec7n “suy lu\u1eadn,” t\u1ee9c l\u00e0 s\u1eed d\u1ee5ng m\u1ed9t m\u00f4 h\u00ecnh l\u1edbn sau khi n\u00f3 \u0111\u01b0\u1ee3c \u0111\u00e0o t\u1ea1o \u0111\u1ec3 t\u1ea1o ra v\u0103n b\u1ea3n, h\u00ecnh \u1ea3nh ho\u1eb7c d\u1ef1 \u0111o\u00e1n.<\/p>\n\n\n\n

Nvidia cho bi\u1ebft H200 s\u1ebd t\u1ea1o ra \u0111\u1ea7u ra g\u1ea7n g\u1ea5p \u0111\u00f4i so v\u1edbi H100. \u0110i\u1ec1u \u0111\u00f3 d\u1ef1a tr\u00ean m\u1ed9t b\u00e0i ki\u1ec3m tra s\u1eed d\u1ee5ng Llama 2 LLM c\u1ee7a Meta.<\/p>\n\n\n\n

D\u1ef1 ki\u1ebfn H200 s\u1ebd \u0111\u01b0\u1ee3c g\u1eedi v\u00e0o qu\u00fd hai n\u0103m 2024 v\u00e0 s\u1ebd c\u1ea1nh tranh v\u1edbi GPU MI300X c\u1ee7a AMD. Chip c\u1ee7a AMD, t\u01b0\u01a1ng t\u1ef1 nh\u01b0 H200, c\u00f3 b\u1ed9 nh\u1edb b\u1ed5 sung so v\u1edbi c\u00e1c phi\u00ean b\u1ea3n tr\u01b0\u1edbc, gi\u00fap \u0111\u01b0a m\u00f4 h\u00ecnh l\u1edbn v\u00e0o ph\u1ea7n c\u1ee9ng \u0111\u1ec3 ch\u1ea1y suy lu\u1eadn.<\/p>\n\n\n

\n
\"\"
Vi x\u1eed l\u00fd Nvidia H200 trong m\u1ed9t h\u1ec7 th\u1ed1ng Nvidia HGX v\u1edbi t\u00e1m GPU.<\/figcaption><\/figure>\n<\/div>\n\n\n

Nvidia cho bi\u1ebft H200 s\u1ebd t\u01b0\u01a1ng th\u00edch v\u1edbi H100, c\u00f3 ngh\u0129a l\u00e0 c\u00e1c c\u00f4ng ty tr\u00ed tu\u1ec7 nh\u00e2n t\u1ea1o \u0111ang \u0111\u00e0o t\u1ea1o v\u1edbi phi\u00ean b\u1ea3n tr\u01b0\u1edbc kh\u00f4ng c\u1ea7n thay \u0111\u1ed5i h\u1ec7 th\u1ed1ng m\u00e1y ch\u1ee7 ho\u1eb7c ph\u1ea7n m\u1ec1m \u0111\u1ec3 s\u1eed d\u1ee5ng phi\u00ean b\u1ea3n m\u1edbi.<\/p>\n\n\n\n

Nvidia cho bi\u1ebft n\u00f3 s\u1ebd c\u00f3 s\u1eb5n trong c\u1ea5u h\u00ecnh m\u00e1y ch\u1ee7 b\u1ed1n GPU ho\u1eb7c t\u00e1m GPU tr\u00ean c\u00e1c h\u1ec7 th\u1ed1ng ho\u00e0n ch\u1ec9nh HGX c\u1ee7a c\u00f4ng ty, c\u0169ng nh\u01b0 trong m\u1ed9t chip mang t\u00ean GH200, k\u1ebft h\u1ee3p GPU H200 v\u1edbi b\u1ed9 x\u1eed l\u00fd d\u1ef1a tr\u00ean Arm.
Tuy nhi\u00ean, H200 c\u00f3 th\u1ec3 kh\u00f4ng gi\u1eef \u0111\u01b0\u1ee3c v\u1ecb tr\u00ed chip tr\u00ed tu\u1ec7 nh\u00e2n t\u1ea1o Nvidia nhanh nh\u1ea5t trong th\u1eddi gian d\u00e0i.<\/p>\n\n\n\n

Trong khi c\u00e1c c\u00f4ng ty nh\u01b0 Nvidia cung c\u1ea5p nhi\u1ec1u c\u1ea5u h\u00ecnh kh\u00e1c nhau c\u1ee7a chip c\u1ee7a h\u1ecd, c\u00e1c b\u1ed9 vi x\u1eed l\u00fd m\u1edbi th\u01b0\u1eddng ti\u1ebfn b\u1ed9 \u0111\u00e1ng k\u1ec3 v\u00e0o kho\u1ea3ng hai n\u0103m, khi c\u00e1c nh\u00e0 s\u1ea3n xu\u1ea5t chuy\u1ec3n sang ki\u1ebfn tr\u00fac kh\u00e1c nhau m\u1edf kh\u00f3a nh\u1eefng c\u1ea3i ti\u1ebfn hi\u1ec7u n\u0103ng \u0111\u00e1ng k\u1ec3 h\u01a1n so v\u1edbi vi\u1ec7c th\u00eam b\u1ed9 nh\u1edb ho\u1eb7c t\u1ed1i \u01b0u nh\u1ecf h\u01a1n kh\u00e1c. C\u1ea3 H100 v\u00e0 H200 \u0111\u1ec1u d\u1ef1a tr\u00ean ki\u1ebfn tr\u00fac Hopper c\u1ee7a Nvidia.<\/p>\n\n\n\n

V\u00e0o th\u00e1ng 10, Nvidia \u0111\u00e3 th\u00f4ng b\u00e1o v\u1edbi c\u00e1c nh\u00e0 \u0111\u1ea7u t\u01b0 r\u1eb1ng h\u1ecd s\u1ebd chuy\u1ec3n t\u1eeb chu k\u1ef3 ki\u1ebfn tr\u00fac hai n\u0103m sang m\u1ed9t m\u00f4 h\u00ecnh ph\u00e1t h\u00e0nh m\u1ed9t n\u0103m do nhu c\u1ea7u cao v\u1ec1 GPU c\u1ee7a c\u00f4ng ty. C\u00f4ng ty \u0111\u00e3 tr\u00ecnh di\u1ec5n m\u1ed9t slide cho th\u1ea5y h\u1ecd s\u1ebd c\u00f4ng b\u1ed1 v\u00e0 ph\u00e1t h\u00e0nh chip B100 c\u1ee7a m\u00ecnh, d\u1ef1a tr\u00ean ki\u1ebfn tr\u00fac Blackwell s\u1eafp t\u1edbi, v\u00e0o n\u0103m 2024.<\/p>\n<\/details>\n","protected":false},"excerpt":{"rendered":"

Nvidia \u0111\u00e3 gi\u1edbi thi\u1ec7u chip \u0111\u1ed3 h\u1ecda H200, \u0111\u01b0\u1ee3c thi\u1ebft k\u1ebf cho hu\u1ea5n luy\u1ec7n v\u00e0 tri\u1ec3n khai m\u00f4 h\u00ecnh tr\u00ed tu\u1ec7 nh\u00e2n t\u1ea1o. \u0110\u00e2y l\u00e0 n\u00e2ng c\u1ea5p t\u1eeb H100, chip \u0111\u01b0\u1ee3c OpenAI s\u1eed d\u1ee5ng \u0111\u1ec3 hu\u1ea5n luy\u1ec7n m\u00f4 h\u00ecnh ng\u00f4n ng\u1eef GPT-4.<\/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":[66,67,48],"class_list":["post-260","post","type-post","status-publish","format-standard","hentry","category-ai-news","tag-chip-do-hoa","tag-huan-luyen-ai-tao-sinh","tag-nvidia"],"_links":{"self":[{"href":"https:\/\/xuhuongai.com\/index.php?rest_route=\/wp\/v2\/posts\/260","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=260"}],"version-history":[{"count":7,"href":"https:\/\/xuhuongai.com\/index.php?rest_route=\/wp\/v2\/posts\/260\/revisions"}],"predecessor-version":[{"id":436,"href":"https:\/\/xuhuongai.com\/index.php?rest_route=\/wp\/v2\/posts\/260\/revisions\/436"}],"wp:attachment":[{"href":"https:\/\/xuhuongai.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=260"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/xuhuongai.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=260"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/xuhuongai.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=260"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}