{"id":497,"date":"2023-11-18T16:37:44","date_gmt":"2023-11-18T09:37:44","guid":{"rendered":"https:\/\/xuhuongai.com\/?p=497"},"modified":"2023-11-27T22:25:41","modified_gmt":"2023-11-27T15:25:41","slug":"silicon-volley-cac-nha-thiet-ke-su-dung-ai-tao-sinh-de-ho-tro-thiet-ke-chip","status":"publish","type":"post","link":"https:\/\/xuhuongai.com\/?p=497","title":{"rendered":"C\u00e1c nh\u00e0 thi\u1ebft k\u1ebf s\u1eed d\u1ee5ng AI t\u1ea1o sinh \u0111\u1ec3 h\u1ed7 tr\u1ee3 thi\u1ebft k\u1ebf chip"},"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 \u2013 KamiMind<\/a>.<\/p>\nNgu\u1ed3n: Rick Merritt, “Silicon Volley: Designers Tap Generative AI for a Chip Assist<\/a>“, NVIDIA, 30\/10\/2023.<\/cite><\/blockquote>\n\n\n\n

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

M\u1ed9t nghi\u00ean c\u1ee9u \u0111\u01b0\u1ee3c c\u00f4ng b\u1ed1 nh\u1eb1m kh\u00e1m ph\u00e1 c\u00e1ch AI t\u1ea1o sinh c\u00f3 th\u1ec3 h\u1ed7 tr\u1ee3 trong m\u1ed9t trong nh\u1eefng n\u1ed7 l\u1ef1c k\u1ef9 thu\u1eadt ph\u1ee9c t\u1ea1p nh\u1ea5t: thi\u1ebft k\u1ebf b\u00e1n d\u1eabn. Nghi\u00ean c\u1ee9u cho th\u1ea5y c\u00e1ch c\u00e1c c\u00f4ng ty trong c\u00e1c l\u0129nh v\u1ef1c chuy\u00ean m\u00f4n cao c\u00f3 th\u1ec3 hu\u1ea5n luy\u1ec7n c\u00e1c m\u00f4 h\u00ecnh ng\u00f4n ng\u1eef l\u1edbn (LLMs) tr\u00ean d\u1eef li\u1ec7u n\u1ed9i b\u1ed9 c\u1ee7a h\u1ecd \u0111\u1ec3 x\u00e2y d\u1ef1ng c\u00e1c tr\u1ee3 l\u00fd nh\u1eb1m gi\u00fap t\u0103ng n\u0103ng su\u1ea5t.<\/p>\n\n\n\n

Thi\u1ebft k\u1ebf b\u00e1n d\u1eabn l\u00e0 m\u1ed9t nhi\u1ec7m v\u1ee5 th\u00e1ch th\u1ee9c, \u0111\u00f2i h\u1ecfi nhi\u1ec1u nh\u00f3m k\u1ef9 s\u01b0 ph\u1ed1i h\u1ee3p trong su\u1ed1t hai n\u0103m \u0111\u1ec3 x\u00e2y d\u1ef1ng m\u1ed9t chip hi\u1ec7n \u0111\u1ea1i. M\u1ed7i c\u00f4ng vi\u1ec7c \u0111\u1ec1u y\u00eau c\u1ea7u c\u00e1c ph\u01b0\u01a1ng ph\u00e1p, ch\u01b0\u01a1ng tr\u00ecnh ph\u1ea7n m\u1ec1m v\u00e0 ng\u00f4n ng\u1eef m\u00e1y t\u00ednh chuy\u00ean bi\u1ec7t.<\/p>\n\n\n\n

Mark Ren, m\u1ed9t gi\u00e1m \u0111\u1ed1c nghi\u00ean c\u1ee9u c\u1ee7a NVIDIA, tin r\u1eb1ng theo th\u1eddi gian, c\u00e1c m\u00f4 h\u00ecnh ng\u00f4n ng\u1eef l\u1edbn s\u1ebd h\u1ed7 tr\u1ee3 r\u1ea5t nhi\u1ec1u cho t\u1ea5t c\u1ea3 c\u00e1c qu\u00e1 tr\u00ecnh li\u00ean quan \u0111\u1ebfn vi\u1ec7c thi\u1ebft k\u1ebf b\u00e1n d\u1eabn. B\u00e0i b\u00e1o chi ti\u1ebft c\u00e1ch c\u00e1c k\u1ef9 s\u01b0 NVIDIA t\u1ea1o ra m\u1ed9t LLM t\u00f9y ch\u1ec9nh, \u0111\u01b0\u1ee3c g\u1ecdi l\u00e0 ChipNeMo, \u0111\u1ec3 s\u1eed d\u1ee5ng n\u1ed9i b\u1ed9. M\u00f4 h\u00ecnh n\u00e0y \u0111\u01b0\u1ee3c hu\u1ea5n luy\u1ec7n tr\u00ean d\u1eef li\u1ec7u n\u1ed9i b\u1ed9 c\u1ee7a c\u00f4ng ty \u0111\u1ec3 t\u1ea1o ra v\u00e0 t\u1ed1i \u01b0u h\u00f3a ph\u1ea7n m\u1ec1m v\u00e0 h\u1ed7 tr\u1ee3 c\u00e1c nh\u00e0 thi\u1ebft k\u1ebf con ng\u01b0\u1eddi.<\/p>\n\n\n\n

Nh\u00f3m nghi\u00ean c\u1ee9u \u0111\u00e3 kh\u1ea3o s\u00e1t c\u00e1c k\u1ef9 s\u01b0 NVIDIA v\u1ec1 c\u00e1c tr\u01b0\u1eddng h\u1ee3p s\u1eed d\u1ee5ng c\u00f3 th\u1ec3 v\u00e0 ch\u1ecdn ba c\u00e1i \u0111\u1ec3 b\u1eaft \u0111\u1ea7u: m\u1ed9t chatbot, m\u1ed9t tr\u00ecnh t\u1ea1o m\u00e3 v\u00e0 m\u1ed9t c\u00f4ng c\u1ee5 ph\u00e2n t\u00edch. C\u00f4ng c\u1ee5 ph\u00e2n t\u00edch, t\u1ef1 \u0111\u1ed9ng h\u00f3a nhi\u1ec7m v\u1ee5 duy tr\u00ec m\u00f4 t\u1ea3 c\u1ee7a c\u00e1c l\u1ed7i \u0111\u00e3 bi\u1ebft, \u0111\u00e3 \u0111\u01b0\u1ee3c \u0111\u00f3n nh\u1eadn nhi\u1ec1u nh\u1ea5t cho \u0111\u1ebfn nay.<\/p>\n\n\n\n

B\u00e0i b\u00e1o c\u0169ng th\u1ea3o lu\u1eadn v\u1ec1 c\u00e1ch t\u00f9y ch\u1ec9nh m\u00f4 h\u00ecnh AI v\u1edbi NVIDIA NeMo, m\u1ed9t framework \u0111\u1ec3 x\u00e2y d\u1ef1ng, t\u00f9y ch\u1ec9nh v\u00e0 tri\u1ec3n khai c\u00e1c m\u00f4 h\u00ecnh AI t\u1ea1o sinh. Nh\u00f3m \u0111\u00e3 s\u1eed d\u1ee5ng framework n\u00e0y \u0111\u1ec3 t\u00f9y ch\u1ec9nh m\u1ed9t m\u00f4 h\u00ecnh v\u1edbi 43 t\u1ef7 tham s\u1ed1, \u0111\u01b0\u1ee3c hu\u1ea5n luy\u1ec7n b\u1eb1ng h\u01a1n m\u1ed9t ngh\u00ecn t\u1ef7 tokens.<\/p>\n\n\n\n

Nghi\u00ean c\u1ee9u nh\u1ea5n m\u1ea1nh t\u1ea7m quan tr\u1ecdng c\u1ee7a vi\u1ec7c t\u00f9y ch\u1ec9nh m\u1ed9t LLM cho c\u00e1c nhi\u1ec7m v\u1ee5 thi\u1ebft k\u1ebf chip. C\u00e1c m\u00f4 h\u00ecnh t\u00f9y ch\u1ec9nh c\u00f3 th\u1ec3 \u0111\u1ea1t ho\u1eb7c v\u01b0\u1ee3t qu\u00e1 hi\u1ec7u su\u1ea5t c\u1ee7a c\u00e1c LLM \u0111a d\u1ee5ng l\u1edbn h\u01a1n trong m\u1ed9t s\u1ed1 tr\u01b0\u1eddng h\u1ee3p. N\u00f3 c\u0169ng n\u00eau r\u00f5 nhu c\u1ea7u thu th\u1eadp v\u00e0 l\u00e0m s\u1ea1ch d\u1eef li\u1ec7u \u0111\u1ec3 hu\u1ea5n luy\u1ec7n c\u1ea9n th\u1eadn, v\u00e0 c\u1eadp nh\u1eadt v\u1edbi nh\u1eefng c\u00f4ng c\u1ee5 m\u1edbi nh\u1ea5t c\u00f3 th\u1ec3 t\u0103ng t\u1ed1c v\u00e0 \u0111\u01a1n gi\u1ea3n h\u00f3a c\u00f4ng vi\u1ec7c.<\/p>\n\n\n\n

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

This article discusses how a research paper released today explores how generative AI can assist in one of the most complex engineering efforts: the design of semiconductors. The research demonstrates how companies in highly specialized fields can train large language models (LLMs) on their internal data to build assistants that increase productivity.<\/p>\n\n\n\n

Semiconductor design is a challenging task that involves multiple engineering teams coordinating for as long as two years to construct a state-of-the-art chip. Each job requires specialized methods, software programs and computer languages.<\/p>\n\n\n\n

Mark Ren, an NVIDIA Research director, believes that over time, large language models will greatly assist all processes involved in the design of semiconductors. The paper details how NVIDIA engineers created a custom LLM, known as ChipNeMo, for their internal use. This model is trained on the company’s internal data to generate and optimize software and assist human designers.<\/p>\n\n\n\n

The research team surveyed NVIDIA engineers for possible use cases and chose three to start with: a chatbot, a code generator, and an analysis tool. The analysis tool, which automates the task of maintaining descriptions of known bugs, has been the most well-received so far.<\/p>\n\n\n\n

The paper also discusses how to customize AI models with NVIDIA NeMo, a framework for building, customizing, and deploying generative AI models. The team used this framework to customize a model with 43 billion parameters, trained using more than a trillion tokens.<\/p>\n\n\n\n

The research emphasizes the importance of customizing an LLM for chip-design tasks. Custom models can match or exceed the performance of larger general-purpose LLMs in some cases. It also highlights the need to carefully collect and clean data for training and stay up to date with the latest tools that can speed up and simplify the work.<\/p>\n<\/details>\n\n\n\n

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

C\u00e1c k\u1ef9 s\u01b0 b\u00e1n d\u1eabn \u0111\u00e3 ch\u1ec9 ra c\u00e1ch m\u1ed9t ng\u00e0nh c\u00f4ng nghi\u1ec7p chuy\u00ean m\u00f4n c\u00f3 th\u1ec3 t\u00f9y ch\u1ec9nh c\u00e1c m\u00f4 h\u00ecnh ng\u00f4n ng\u1eef l\u1edbn \u0111\u1ec3 gia t\u0103ng hi\u1ec7u su\u1ea5t b\u1eb1ng c\u00e1ch s\u1eed d\u1ee5ng NVIDIA NeMo.<\/p>\n\n\n\n

M\u1ed9t b\u00e0i nghi\u00ean c\u1ee9u \u0111\u01b0\u1ee3c c\u00f4ng b\u1ed1 (h\u00f4m nay) m\u00f4 t\u1ea3 nh\u1eefng c\u00e1ch AI t\u1ea1o sinh c\u00f3 th\u1ec3 h\u1ed7 tr\u1ee3 m\u1ed9t trong nh\u1eefng n\u1ed7 l\u1ef1c k\u1ef9 thu\u1eadt ph\u1ee9c t\u1ea1p nh\u1ea5t: thi\u1ebft k\u1ebf b\u00e1n d\u1eabn.<\/p>\n\n\n\n

C\u00f4ng tr\u00ecnh n\u00e0y minh h\u1ecda c\u00e1ch c\u00e1c c\u00f4ng ty trong c\u00e1c l\u0129nh v\u1ef1c chuy\u00ean m\u00f4n cao c\u00f3 th\u1ec3 hu\u1ea5n luy\u1ec7n c\u00e1c m\u00f4 h\u00ecnh ng\u00f4n ng\u1eef l\u1edbn (LLMs) tr\u00ean d\u1eef li\u1ec7u n\u1ed9i b\u1ed9 c\u1ee7a h\u1ecd \u0111\u1ec3 x\u00e2y d\u1ef1ng tr\u1ee3 l\u00fd t\u0103ng n\u0103ng su\u1ea5t.<\/p>\n\n\n\n

\u00cdt l\u0129nh v\u1ef1c n\u00e0o th\u00e1ch th\u1ee9c nh\u01b0 thi\u1ebft k\u1ebf b\u00e1n d\u1eabn. D\u01b0\u1edbi k\u00ednh hi\u1ec3n vi, m\u1ed9t chip ti\u00ean ti\u1ebfn nh\u01b0 NVIDIA H100 Tensor Core GPU tr\u00f4ng gi\u1ed1ng nh\u01b0 m\u1ed9t th\u00e0nh ph\u1ed1 \u0111\u01b0\u1ee3c quy ho\u1ea1ch t\u1ed1t, \u0111\u01b0\u1ee3c x\u00e2y d\u1ef1ng v\u1edbi h\u00e0ng ch\u1ee5c t\u1ef7 transistor, \u0111\u01b0\u1ee3c k\u1ebft n\u1ed1i tr\u00ean nh\u1eefng con \u0111\u01b0\u1eddng m\u1ecfng h\u01a1n t\u00f3c ng\u01b0\u1eddi 10.000 l\u1ea7n.<\/p>\n\n\n\n

Nhi\u1ec1u nh\u00f3m k\u1ef9 s\u01b0 ph\u1ed1i h\u1ee3p trong v\u00f2ng hai n\u0103m \u0111\u1ec3 x\u00e2y d\u1ef1ng m\u1ed9t trong nh\u1eefng th\u00e0nh ph\u1ed1 k\u1ef9 thu\u1eadt s\u1ed1 l\u1edbn nh\u01b0 th\u1ebf n\u00e0y.<\/p>\n\n\n\n

M\u1ed9t s\u1ed1 nh\u00f3m \u0111\u1ecbnh r\u00f5 ki\u1ebfn tr\u00fac t\u1ed5ng th\u1ec3 c\u1ee7a chip, m\u1ed9t s\u1ed1 nh\u00f3m t\u1ea1o v\u00e0 \u0111\u1eb7t m\u1ed9t lo\u1ea1t m\u1ea1ch si\u00eau nh\u1ecf, v\u00e0 nh\u1eefng nh\u00f3m kh\u00e1c ki\u1ec3m tra c\u00f4ng vi\u1ec7c c\u1ee7a h\u1ecd. M\u1ed7i c\u00f4ng vi\u1ec7c \u0111\u1ec1u y\u00eau c\u1ea7u c\u00e1c ph\u01b0\u01a1ng ph\u00e1p chuy\u00ean m\u00f4n, c\u00e1c ch\u01b0\u01a1ng tr\u00ecnh ph\u1ea7n m\u1ec1m v\u00e0 ng\u00f4n ng\u1eef m\u00e1y t\u00ednh.<\/p>\n\n\n\n

T\u1ea7m nh\u00ecn r\u1ed9ng l\u1edbn cho LLMs<\/em><\/strong><\/p>\n\n\n\n

“T\u00f4i tin r\u1eb1ng theo th\u1eddi gian, c\u00e1c m\u00f4 h\u00ecnh ng\u00f4n ng\u1eef l\u1edbn s\u1ebd gi\u00fap cho m\u1ecdi quy tr\u00ecnh cho nhi\u1ec1u nh\u00f3m thi\u1ebft k\u1ebf,” Mark Ren, gi\u00e1m \u0111\u1ed1c nghi\u00ean c\u1ee9u NVIDIA v\u00e0 t\u00e1c gi\u1ea3 ch\u00ednh c\u1ee7a b\u00e0i vi\u1ebft, n\u00f3i.<\/p>\n\n\n\n

Bill Dally, gi\u00e1m \u0111\u1ed1c khoa h\u1ecdc c\u1ee7a NVIDIA, trong m\u1ed9t b\u00e0i ph\u00e1t bi\u1ec3u ch\u00ednh t\u1ea1i H\u1ed9i ngh\u1ecb Qu\u1ed1c t\u1ebf v\u1ec1 Thi\u1ebft k\u1ebf h\u1ed7 tr\u1ee3 b\u1edfi M\u00e1y t\u00ednh, m\u1ed9t cu\u1ed9c t\u1ee5 h\u1ecdp h\u00e0ng n\u0103m c\u1ee7a h\u00e0ng tr\u0103m k\u1ef9 s\u01b0 l\u00e0m vi\u1ec7c trong l\u0129nh v\u1ef1c \u0111\u01b0\u1ee3c g\u1ecdi l\u00e0 t\u1ef1 \u0111\u1ed9ng h\u00f3a thi\u1ebft k\u1ebf \u0111i\u1ec7n t\u1eed, ho\u1eb7c EDA.<\/p>\n\n\n\n

“\u0110\u00e2y l\u00e0 b\u01b0\u1edbc \u0111\u1ea7u ti\u00ean quan tr\u1ecdng trong vi\u1ec7c \u00e1p d\u1ee5ng LLMs v\u00e0o c\u00f4ng vi\u1ec7c ph\u1ee9c t\u1ea1p c\u1ee7a vi\u1ec7c thi\u1ebft k\u1ebf b\u00e1n d\u1eabn,” Dally n\u00f3i t\u1ea1i s\u1ef1 ki\u1ec7n t\u1ea1i San Francisco. “N\u00f3 cho th\u1ea5y c\u00e1ch ngay c\u1ea3 c\u00e1c l\u0129nh v\u1ef1c chuy\u00ean m\u00f4n cao c\u00f3 th\u1ec3 s\u1eed d\u1ee5ng d\u1eef li\u1ec7u n\u1ed9i b\u1ed9 c\u1ee7a h\u1ecd \u0111\u1ec3 hu\u1ea5n luy\u1ec7n c\u00e1c m\u00f4 h\u00ecnh AI t\u1ea1o sinh h\u1eefu \u00edch.”<\/p>\n\n\n\n

ChipNeMo xu\u1ea5t hi\u1ec7n<\/em><\/strong><\/p>\n\n\n\n

B\u00e0i nghi\u00ean c\u1ee9u chi ti\u1ebft c\u00e1ch c\u00e1c k\u1ef9 s\u01b0 NVIDIA t\u1ea1o ra cho vi\u1ec7c s\u1eed d\u1ee5ng n\u1ed9i b\u1ed9 m\u1ed9t LLM t\u00f9y ch\u1ec9nh, \u0111\u01b0\u1ee3c g\u1ecdi l\u00e0 ChipNeMo, \u0111\u01b0\u1ee3c hu\u1ea5n luy\u1ec7n tr\u00ean d\u1eef li\u1ec7u n\u1ed9i b\u1ed9 c\u1ee7a c\u00f4ng ty \u0111\u1ec3 t\u1ea1o ra v\u00e0 t\u1ed1i \u01b0u h\u00f3a ph\u1ea7n m\u1ec1m v\u00e0 h\u1ed7 tr\u1ee3 c\u00e1c nh\u00e0 thi\u1ebft k\u1ebf con ng\u01b0\u1eddi.<\/p>\n\n\n\n

D\u00e0i h\u1ea1n, c\u00e1c k\u1ef9 s\u01b0 hy v\u1ecdng s\u1ebd \u00e1p d\u1ee5ng AI t\u1ea1o sinh cho t\u1eebng giai \u0111o\u1ea1n c\u1ee7a thi\u1ebft k\u1ebf chip, c\u00f3 th\u1ec3 thu \u0111\u01b0\u1ee3c nh\u1eefng l\u1ee3i \u00edch \u0111\u00e1ng k\u1ec3 v\u1ec1 n\u0103ng su\u1ea5t t\u1ed5ng th\u1ec3, Ren n\u00f3i, ng\u01b0\u1eddi \u0111\u00e3 c\u00f3 s\u1ef1 nghi\u1ec7p k\u00e9o d\u00e0i h\u01a1n 20 n\u0103m trong EDA.<\/p>\n\n\n\n

Sau khi kh\u1ea3o s\u00e1t c\u00e1c k\u1ef9 s\u01b0 NVIDIA v\u1ec1 c\u00e1c t\u00ecnh hu\u1ed1ng \u1ee9ng d\u1ee5ng c\u00f3 th\u1ec3, nh\u00f3m nghi\u00ean c\u1ee9u \u0111\u00e3 ch\u1ecdn ba tr\u01b0\u1eddng h\u1ee3p \u0111\u1ec3 b\u1eaft \u0111\u1ea7u: m\u1ed9t chatbot, m\u1ed9t tr\u00ecnh t\u1ea1o m\u00e3 v\u00e0 m\u1ed9t c\u00f4ng c\u1ee5 ph\u00e2n t\u00edch.<\/p>\n\n\n\n

C\u00e1c t\u00ecnh hu\u1ed1ng \u1ee9ng d\u1ee5ng ban \u0111\u1ea7u<\/em><\/strong><\/p>\n\n\n\n

M\u1ed9t c\u00f4ng c\u1ee5 t\u1ef1 \u0111\u1ed9ng h\u00f3a c\u00e1c nhi\u1ec7m v\u1ee5 t\u1ed1n th\u1eddi gian c\u1ee7a vi\u1ec7c duy tr\u00ec c\u00e1c m\u00f4 t\u1ea3 c\u1eadp nh\u1eadt v\u1ec1 c\u00e1c l\u1ed7i \u0111\u00e3 bi\u1ebft, \u0111\u00e3 \u0111\u01b0\u1ee3c \u0111\u00f3n nh\u1eadn nhi\u1ec7t t\u00ecnh nh\u1ea5t cho \u0111\u1ebfn nay.<\/p>\n\n\n\n

M\u1ed9t chatbot prototype tr\u1ea3 l\u1eddi c\u00e1c c\u00e2u h\u1ecfi v\u1ec1 ki\u1ebfn tr\u00fac v\u00e0 thi\u1ebft k\u1ebf GPU \u0111\u00e3 gi\u00fap nhi\u1ec1u k\u1ef9 s\u01b0 nhanh ch\u00f3ng t\u00ecm th\u1ea5y c\u00e1c t\u00e0i li\u1ec7u k\u1ef9 thu\u1eadt trong c\u00e1c b\u00e0i ki\u1ec3m tra \u0111\u1ea7u ti\u00ean.<\/p>\n\n\n\n

\"\"\n\t\t\t\n\t\t\t\t\n\t\t\t<\/svg>\n\t\t<\/button>
Ph\u1ea7n m\u1ec1m t\u1ea1o m\u00e3 l\u1eadp tr\u00ecnh h\u1ed7 tr\u1ee3 \u0111\u1ed9i thi\u1ebft k\u1ebf vi\u1ebft ph\u1ea7n m\u1ec1m cho chip x\u1eed l\u00fd. Ngu\u1ed3n: NVIDIA<\/figcaption><\/figure>\n\n\n\n

M\u1ed9t tr\u00ecnh t\u1ea1o m\u00e3 \u0111ang \u0111\u01b0\u1ee3c ph\u00e1t tri\u1ec3n \u0111\u00e3 t\u1ea1o ra c\u00e1c \u0111o\u1ea1n m\u00e3 kho\u1ea3ng 10-20 d\u00f2ng ph\u1ea7n m\u1ec1m trong hai ng\u00f4n ng\u1eef chuy\u00ean m\u00f4n m\u00e0 c\u00e1c nh\u00e0 thi\u1ebft k\u1ebf chip s\u1eed d\u1ee5ng. N\u00f3 s\u1ebd \u0111\u01b0\u1ee3c t\u00edch h\u1ee3p v\u1edbi c\u00e1c c\u00f4ng c\u1ee5 hi\u1ec7n c\u00f3, v\u00ec v\u1eady c\u00e1c k\u1ef9 s\u01b0 c\u00f3 m\u1ed9t tr\u1ee3 l\u00fd h\u1eefu \u00edch cho c\u00e1c thi\u1ebft k\u1ebf \u0111ang ti\u1ebfn h\u00e0nh.<\/p>\n\n\n\n

T\u00f9y ch\u1ec9nh c\u00e1c m\u00f4 h\u00ecnh AI v\u1edbi NVIDIA NeMo<\/em><\/strong><\/p>\n\n\n\n

B\u00e0i vi\u1ebft ch\u1ee7 y\u1ebfu t\u1eadp trung v\u00e0o c\u00f4ng vi\u1ec7c c\u1ee7a nh\u00f3m thu th\u1eadp d\u1eef li\u1ec7u thi\u1ebft k\u1ebf c\u1ee7a m\u00ecnh v\u00e0 s\u1eed d\u1ee5ng n\u00f3 \u0111\u1ec3 t\u1ea1o ra m\u1ed9t m\u00f4 h\u00ecnh AI t\u1ea1o sinh chuy\u00ean m\u00f4n, m\u1ed9t qu\u00e1 tr\u00ecnh c\u00f3 th\u1ec3 chuy\u1ec3n \u0111\u1ed5i sang b\u1ea5t k\u1ef3 ng\u00e0nh c\u00f4ng nghi\u1ec7p n\u00e0o.<\/p>\n\n\n\n

L\u00e0m \u0111i\u1ec3m kh\u1edfi \u0111\u1ea7u, nh\u00f3m ch\u1ecdn m\u1ed9t m\u00f4 h\u00ecnh c\u01a1 s\u1edf v\u00e0 t\u00f9y ch\u1ec9nh n\u00f3 v\u1edbi NVIDIA NeMo, m\u1ed9t khung cho vi\u1ec7c x\u00e2y d\u1ef1ng, t\u00f9y ch\u1ec9nh v\u00e0 tri\u1ec3n khai c\u00e1c m\u00f4 h\u00ecnh AI t\u1ea1o sinh, \u0111\u01b0\u1ee3c bao g\u1ed3m trong n\u1ec1n t\u1ea3ng ph\u1ea7n m\u1ec1m NVIDIA AI Enterprise. M\u00f4 h\u00ecnh NeMo \u0111\u01b0\u1ee3c ch\u1ecdn c\u00f3 43 t\u1ef7 tham s\u1ed1, m\u1ed9t \u0111\u1ed9 \u0111o c\u1ee7a kh\u1ea3 n\u0103ng hi\u1ec3u c\u00e1c m\u00f4 h\u00ecnh c\u1ee7a n\u00f3. N\u00f3 \u0111\u01b0\u1ee3c hu\u1ea5n luy\u1ec7n b\u1eb1ng c\u00e1ch s\u1eed d\u1ee5ng h\u01a1n m\u1ed9t ngh\u00ecn t\u1ef7 token, nh\u1eefng t\u1eeb v\u1ef1ng v\u00e0 k\u00fd hi\u1ec7u trong v\u0103n b\u1ea3n v\u00e0 ph\u1ea7n m\u1ec1m.<\/p>\n\n\n\n

\"\"
Ki\u1ebfn tr\u00fac m\u00f4 h\u00ecnh ChipNeMo. Ngu\u1ed3n: NVIDIA<\/figcaption><\/figure>\n\n\n\n

Sau \u0111\u00f3, nh\u00f3m \u0111\u00e3 c\u1ea3i thi\u1ec7n m\u00f4 h\u00ecnh trong hai v\u00f2ng hu\u1ea5n luy\u1ec7n, v\u00f2ng \u0111\u1ea7u ti\u00ean s\u1eed d\u1ee5ng kho\u1ea3ng 24 t\u1ef7 token d\u1eef li\u1ec7u thi\u1ebft k\u1ebf n\u1ed9i b\u1ed9 c\u1ee7a n\u00f3 v\u00e0 v\u00f2ng th\u1ee9 hai tr\u00ean m\u1ed9t s\u1ef1 k\u1ebft h\u1ee3p c\u1ee7a kho\u1ea3ng 130.000 cu\u1ed9c tr\u00f2 chuy\u1ec7n v\u00e0 thi\u1ebft k\u1ebf m\u1eabu.<\/p>\n\n\n\n

C\u00f4ng vi\u1ec7c n\u00e0y l\u00e0 m\u1ed9t trong s\u1ed1 nhi\u1ec1u v\u00ed d\u1ee5 v\u1ec1 nghi\u00ean c\u1ee9u v\u00e0 minh ch\u1ee9ng cho kh\u00e1i ni\u1ec7m v\u1ec1 AI t\u1ea1o sinh trong ng\u00e0nh c\u00f4ng nghi\u1ec7p b\u00e1n d\u1eabn, m\u1edbi ch\u1ec9 b\u1eaft \u0111\u1ea7u xu\u1ea5t hi\u1ec7n t\u1eeb ph\u00f2ng th\u00ed nghi\u1ec7m.<\/p>\n\n\n\n

Chia s\u1ebb B\u00e0i h\u1ecdc<\/em><\/strong><\/p>\n\n\n\n

M\u1ed9t trong nh\u1eefng b\u00e0i h\u1ecdc quan tr\u1ecdng nh\u1ea5t m\u00e0 nh\u00f3m c\u1ee7a Ren \u0111\u00e3 h\u1ecdc \u0111\u01b0\u1ee3c l\u00e0 gi\u00e1 tr\u1ecb c\u1ee7a vi\u1ec7c t\u00f9y ch\u1ec9nh m\u1ed9t LLM.<\/p>\n\n\n\n

Tr\u00ean c\u00e1c nhi\u1ec7m v\u1ee5 thi\u1ebft k\u1ebf chip, c\u00e1c m\u00f4 h\u00ecnh ChipNeMo t\u00f9y ch\u1ec9nh v\u1edbi \u00edt nh\u1ea5t 13 t\u1ef7 tham s\u1ed1 ph\u00f9 h\u1ee3p ho\u1eb7c v\u01b0\u1ee3t tr\u1ed9i h\u01a1n hi\u1ec7u su\u1ea5t c\u1ee7a c\u1ea3 c\u00e1c LLMs \u0111a d\u1ee5ng v\u1edbi 70 t\u1ef7 tham s\u1ed1 nh\u01b0 LLaMA2. Trong m\u1ed9t s\u1ed1 t\u00ecnh hu\u1ed1ng \u1ee9ng d\u1ee5ng, c\u00e1c m\u00f4 h\u00ecnh ChipNeMo \u0111\u00e3 t\u1ed1t h\u01a1n r\u1ea5t nhi\u1ec1u.<\/p>\n\n\n\n

V\u1ec1 l\u00e2u d\u00e0i, ng\u01b0\u1eddi d\u00f9ng c\u1ea7n th\u1eadn tr\u1ecdng trong vi\u1ec7c thu th\u1eadp d\u1eef li\u1ec7u n\u00e0o v\u00e0 l\u00e0m c\u00e1ch n\u00e0o \u0111\u1ec3 l\u00e0m s\u1ea1ch n\u00f3 \u0111\u1ec3 s\u1eed d\u1ee5ng trong hu\u1ea5n luy\u1ec7n, \u00f4ng th\u00eam v\u00e0o.<\/p>\n\n\n\n

Cu\u1ed1i c\u00f9ng, Ren khuy\u00ean ng\u01b0\u1eddi d\u00f9ng n\u00ean c\u1eadp nh\u1eadt v\u1edbi nh\u1eefng c\u00f4ng c\u1ee5 m\u1edbi nh\u1ea5t c\u00f3 th\u1ec3 t\u0103ng t\u1ed1c v\u00e0 \u0111\u01a1n gi\u1ea3n h\u00f3a c\u00f4ng vi\u1ec7c.<\/p>\n\n\n\n

NVIDIA Research c\u00f3 h\u00e0ng tr\u0103m nh\u00e0 khoa h\u1ecdc v\u00e0 k\u1ef9 s\u01b0 tr\u00ean to\u00e0n th\u1ebf gi\u1edbi t\u1eadp trung v\u00e0o c\u00e1c ch\u1ee7 \u0111\u1ec1 nh\u01b0 AI, \u0111\u1ed3 h\u1ecda m\u00e1y t\u00ednh, th\u1ecb gi\u00e1c m\u00e1y t\u00ednh, xe t\u1ef1 l\u00e1i v\u00e0 robot. C\u00e1c d\u1ef1 \u00e1n g\u1ea7n \u0111\u00e2y kh\u00e1c trong l\u0129nh v\u1ef1c b\u00e1n d\u1eabn bao g\u1ed3m vi\u1ec7c s\u1eed d\u1ee5ng AI \u0111\u1ec3 thi\u1ebft k\u1ebf m\u1ea1ch nh\u1ecf h\u01a1n, nhanh h\u01a1n v\u00e0 t\u1ed1i \u01b0u h\u00f3a vi\u1ec7c \u0111\u1eb7t c\u00e1c kh\u1ed1i l\u1edbn.<\/p>\n<\/details>\n","protected":false},"excerpt":{"rendered":"

M\u1ed9t nghi\u00ean c\u1ee9u \u0111\u01b0\u1ee3c c\u00f4ng b\u1ed1 nh\u1eb1m kh\u00e1m ph\u00e1 c\u00e1ch AI t\u1ea1o sinh c\u00f3 th\u1ec3 h\u1ed7 tr\u1ee3 trong m\u1ed9t trong nh\u1eefng n\u1ed7 l\u1ef1c k\u1ef9 thu\u1eadt ph\u1ee9c t\u1ea1p nh\u1ea5t: thi\u1ebft k\u1ebf b\u00e1n d\u1eabn.<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[5,90],"tags":[54,98,53,48],"class_list":["post-497","post","type-post","status-publish","format-standard","hentry","category-ai-news","category-featured-news","tag-ai-tao-sinh","tag-chip-ban-dan","tag-genai","tag-nvidia"],"_links":{"self":[{"href":"https:\/\/xuhuongai.com\/index.php?rest_route=\/wp\/v2\/posts\/497","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=497"}],"version-history":[{"count":10,"href":"https:\/\/xuhuongai.com\/index.php?rest_route=\/wp\/v2\/posts\/497\/revisions"}],"predecessor-version":[{"id":600,"href":"https:\/\/xuhuongai.com\/index.php?rest_route=\/wp\/v2\/posts\/497\/revisions\/600"}],"wp:attachment":[{"href":"https:\/\/xuhuongai.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=497"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/xuhuongai.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=497"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/xuhuongai.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=497"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}