{"id":3386,"date":"2024-03-25T10:51:46","date_gmt":"2024-03-25T03:51:46","guid":{"rendered":"https:\/\/dangngocduy.vn\/?p=3386"},"modified":"2024-03-25T10:51:46","modified_gmt":"2024-03-25T03:51:46","slug":"amd-announced-its-ai-chip-that-is-60-more-powerful-than-the-nvidia-h100","status":"publish","type":"post","link":"https:\/\/dangngocduy.vn\/vi\/amd-announced-its-ai-chip-that-is-60-more-powerful-than-the-nvidia-h100\/","title":{"rendered":"AMD c\u00f4ng b\u1ed1 chip AI m\u1ea1nh h\u01a1n 60% so v\u1edbi Nvidia H100"},"content":{"rendered":"
\n

\u0110\u1ea7u ti\u00ean \u0111\u01b0\u1ee3c gi\u1edbi thi\u1ec7u<\/a> v\u00e0o th\u00e1ng 6 v\u00e0 ch\u00ednh th\u1ee9c ra m\u1eaft v\u00e0o ng\u00e0y 6 th\u00e1ng 12 t\u1ea1i s\u1ef1 ki\u1ec7n Advancing AI \u1edf San Jose, California, Instinct MI300X nh\u1eafm \u0111\u1ebfn c\u00e1c h\u1ec7 th\u1ed1ng \u0111\u00e0o t\u1ea1o AI. Trong s\u1ef1 ki\u1ec7n, AMD \u0111\u00e3 g\u1ecdi m\u1eabu chip c\u1ee7a Nvidia l\u00e0 \u0111\u1ed1i th\u1ee7 c\u1ea1nh tranh ch\u00ednh.<\/p>\n

MI300X \u0111\u01b0\u1ee3c coi l\u00e0 \u0111\u1ec9nh cao trong ph\u01b0\u01a1ng ph\u00e1p thi\u1ebft k\u1ebf chiplet m\u1edbi nh\u1ea5t c\u1ee7a AMD d\u00e0nh cho c\u00e1c b\u1ed9 x\u1eed l\u00fd \u0111\u1ed3 h\u1ecda (GPU), k\u1ebft h\u1ee3p 8 ng\u0103n x\u1ebfp b\u1ed9 nh\u1edb HBM3 12Hi v\u1edbi 8 chiplet GPU 5nm CDNA 3 x\u1ebfp ch\u1ed3ng l\u00ean nhau. 3D, c\u00f3 t\u00ean l\u00e0 XCD. C\u00f4ng ngh\u1ec7 lai n\u00e0y \u0111\u01b0\u1ee3c AMD g\u1ecdi l\u00e0 3.5D, v\u1edbi s\u1ef1 k\u1ebft h\u1ee3p gi\u1eefa 3D v\u00e0 2.5D. K\u1ebft qu\u1ea3 l\u00e0 m\u1ed9t con chip m\u1edbi c\u00f3 c\u00f4ng su\u1ea5t 750 W, s\u1eed d\u1ee5ng 304 \u0111\u01a1n v\u1ecb t\u00ednh to\u00e1n, dung l\u01b0\u1ee3ng HBM3 l\u00e0 192 GB v\u00e0 \u0111\u1ea1t b\u0103ng th\u00f4ng 5,3 TB\/s.<\/p>\n

B\u00ean trong MI300X \u0111\u01b0\u1ee3c thi\u1ebft k\u1ebf \u0111\u1ec3 ho\u1ea1t \u0111\u1ed9ng theo nh\u00f3m v\u1edbi t\u1ed5ng c\u1ed9ng 8 ph\u00e2n v\u00f9ng x\u1eed l\u00fd ri\u00eang bi\u1ec7t. Th\u00f4ng qua k\u1ebft n\u1ed1i Infinity Fabric, c\u00e1c v\u00f9ng n\u00e0y s\u1ebd li\u00ean l\u1ea1c v\u1edbi nhau \u1edf b\u0103ng th\u00f4ng 896 GB\/s. To\u00e0n b\u1ed9 \u0111i\u1ec1u n\u00e0y mang l\u1ea1i hi\u1ec7u su\u1ea5t 10,4 Petaflops (tri\u1ec7u t\u1ef7 thao t\u00e1c m\u1ed7i gi\u00e2y). <\/p>\n

<\/p>\n
\"AMD<\/source><\/picture><\/div>
\n

Gi\u00e1m \u0111\u1ed1c \u0111i\u1ec1u h\u00e0nh AMD Lisa Su<\/a> c\u00f4ng b\u1ed1 th\u00f4ng s\u1ed1 k\u1ef9 thu\u1eadt c\u1ee7a MI300X v\u00e0 so s\u00e1nh n\u00f3 v\u1edbi Nvidia H100 tr\u00ean c\u00e1c m\u1eabu LLM h\u00e0ng \u0111\u1ea7u. H\u00ecnh ch\u1ee5p: wccftech<\/em><\/p>\n<\/figcaption><\/figure>\n

Theo c\u00f4ng b\u1ed1 c\u1ee7a AMD, MI300X c\u00f3 dung l\u01b0\u1ee3ng b\u1ed9 nh\u1edb l\u1edbn h\u01a1n 2,4 l\u1ea7n v\u00e0 l\u1ee3i th\u1ebf v\u1ec1 s\u1ee9c m\u1ea1nh t\u00ednh to\u00e1n g\u1ea5p 1,3 l\u1ea7n so v\u1edbi n\u1ec1n t\u1ea3ng H100 \u2013 chip AI th\u01b0\u1eddng \u0111\u01b0\u1ee3c s\u1eed d\u1ee5ng trong c\u00e1c \u1ee9ng d\u1ee5ng AI. \u0111\u01b0\u1ee3c t\u1ea1o g\u1ea7n \u0111\u00e2y, bao g\u1ed3m c\u1ea3 GPT-4 c\u1ee7a OpenAI. Dung l\u01b0\u1ee3ng b\u1ed9 nh\u1edb l\u1edbn s\u1ebd gi\u00fap t\u0103ng hi\u1ec7u su\u1ea5t t\u00ednh to\u00e1n nh\u1edd x\u1eed l\u00fd \u0111\u01b0\u1ee3c nhi\u1ec1u d\u1eef li\u1ec7u h\u01a1n c\u00f9ng l\u00fac.<\/p>\n

Khi th\u1eed nghi\u1ec7m v\u1edbi c\u00e1c m\u00f4 h\u00ecnh ng\u00f4n ng\u1eef l\u1edbn, MI300X v\u01b0\u1ee3t tr\u1ed9i so v\u1edbi \u0111\u1ed1i th\u1ee7 \u0111\u1ebfn t\u1eeb Nvidia. V\u1edbi Llama 2 70B (70 t\u1ef7 th\u00f4ng s\u1ed1) v\u00e0 FlashAttention 2, chip AMD nhanh h\u01a1n 20% khi so s\u00e1nh 1v1. Khi k\u1ebft h\u1ee3p v\u1edbi 8v8 (8v8 Server), MI300X m\u1ea1nh h\u01a1n H100 40% khi ch\u1ea1y tr\u00ean Llama 2 70B v\u00e0 l\u00ean t\u1edbi 60% khi ch\u1ea1y tr\u00ean Bloom 176B (176 t\u1ef7 th\u00f4ng s\u1ed1). <\/p>\n

<\/p>\n
\"AMD<\/source><\/picture><\/div>
\n

Chip AMD Instinct MI300X. H\u00ecnh ch\u1ee5p: AMD<\/em><\/p>\n<\/figcaption><\/figure>\n

AMD nh\u1ea5n m\u1ea1nh MI300X ngang b\u1eb1ng v\u1edbi H100 v\u1ec1 hi\u1ec7u n\u0103ng luy\u1ec7n t\u1eadp nh\u01b0ng c\u00f3 gi\u00e1 c\u1ea1nh tranh h\u01a1n khi so s\u00e1nh c\u00f9ng kh\u1ed1i l\u01b0\u1ee3ng c\u00f4ng vi\u1ec7c. Tuy nhi\u00ean, h\u00e3ng v\u1eabn ch\u01b0a c\u00f4ng b\u1ed1 gi\u00e1 b\u00e1n ch\u00ednh th\u1ee9c c\u1ee7a s\u1ea3n ph\u1ea9m.<\/p>\n

M\u1eb7c d\u00f9 c\u00f3 th\u00f4ng s\u1ed1 k\u1ef9 thu\u1eadt m\u1ea1nh m\u1ebd nh\u01b0ng MI300X kh\u00f4ng ph\u1ea3i l\u00e0 GPU m\u1ea1nh nh\u1ea5t. Gi\u1eefa th\u00e1ng 11, Nvidia c\u00f4ng b\u1ed1 H200 m\u1edbi nh\u1ea5t v\u1edbi hi\u1ec7u n\u0103ng g\u1ea7n g\u1ea5p \u0111\u00f4i c\u1ee7a H100<\/a> .<\/p>\n

B\u00ean c\u1ea1nh Instinct MI300X, AMD c\u0169ng gi\u1edbi thi\u1ec7u Instinct MI300A, b\u1ed9 x\u1eed l\u00fd APU (Accelerated Treatment Unit) d\u00e0nh cho trung t\u00e2m d\u1eef li\u1ec7u \u0111\u1ea7u ti\u00ean tr\u00ean th\u1ebf gi\u1edbi. Con chip n\u00e0y c\u0169ng s\u1eed d\u1ee5ng k\u1ef9 thu\u1eadt \u0111\u00f3ng g\u00f3i 3.5D do TSMC tri\u1ec3n khai t\u01b0\u01a1ng t\u1ef1 MI300X, k\u1ebft h\u1ee3p hai ti\u1ebfn tr\u00ecnh 5 nm v\u00e0 6 nm v\u1edbi 13 chiplets x\u1ebfp ch\u1ed3ng l\u00ean nhau, t\u1ea1o ra m\u1ed9t con chip c\u00f3 24 l\u00f5i ch\u1ee9a 53 t\u1ef7 b\u00f3ng b\u00e1n d\u1eabn. \u0110\u00e2y c\u0169ng l\u00e0 con chip l\u1edbn nh\u1ea5t m\u00e0 AMD t\u1eebng s\u1ea3n xu\u1ea5t. H\u00e3ng ch\u01b0a c\u00f4ng b\u1ed1 gi\u00e1 s\u1ea3n ph\u1ea9m.<\/p>\n

Gi\u00e1m \u0111\u1ed1c \u0111i\u1ec1u h\u00e0nh AMD Lisa Su cho bi\u1ebft: \u201cT\u1ea5t c\u1ea3 s\u1ef1 ch\u00fa \u00fd \u0111\u1ec1u t\u1eadp trung v\u00e0o chip AI\u201d. \u201cNh\u1eefng g\u00ec chip c\u1ee7a ch\u00fang t\u00f4i mang l\u1ea1i s\u1ebd tr\u1ef1c ti\u1ebfp mang l\u1ea1i tr\u1ea3i nghi\u1ec7m t\u1ed1t h\u01a1n cho ng\u01b0\u1eddi d\u00f9ng. Khi b\u1ea1n h\u1ecfi \u0111i\u1ec1u g\u00ec \u0111\u00f3, c\u00e2u tr\u1ea3 l\u1eddi s\u1ebd xu\u1ea5t hi\u1ec7n nhanh h\u01a1n, \u0111\u1eb7c bi\u1ec7t khi c\u00e2u tr\u1ea3 l\u1eddi tr\u1edf n\u00ean ph\u1ee9c t\u1ea1p h\u01a1n\u201d.<\/p>\n

Trong khi \u0111\u00f3, Meta v\u00e0 Microsoft cho bi\u1ebft h\u1ecd s\u1ebd mua Instinct MI300X \u0111\u1ec3 thay th\u1ebf chip Nvidia. \u0110\u1ea1i di\u1ec7n Meta cho bi\u1ebft h\u1ecd s\u1ebd trang b\u1ecb MI300X cho c\u00e1c h\u1ec7 th\u1ed1ng suy lu\u1eadn tr\u00ed tu\u1ec7 nh\u00e2n t\u1ea1o nh\u01b0 x\u1eed l\u00fd nh\u00e3n d\u00e1n AI, ch\u1ec9nh s\u1eeda h\u00ecnh \u1ea3nh v\u00e0 v\u1eadn h\u00e0nh tr\u1ee3 l\u00fd \u1ea3o. V\u00e0 Kevin Scott, CTO c\u1ee7a Microsoft, cho bi\u1ebft \u00f4ng s\u1ebd s\u1eed d\u1ee5ng chip AMD cho m\u00e1y ch\u1ee7 \u0111\u00e1m m\u00e2y Azure. OpenAI c\u0169ng s\u1ebd s\u1eed d\u1ee5ng MI300X cho h\u1ec7 th\u1ed1ng x\u1eed l\u00fd ph\u1ea7n m\u1ec1m c\u00f3 t\u00ean Triton, m\u1ed9t m\u00f4 h\u00ecnh \u0111\u01b0\u1ee3c s\u1eed d\u1ee5ng trong nghi\u00ean c\u1ee9u AI.<\/p>\n

AMD d\u1ef1 \u0111o\u00e1n t\u1ed5ng doanh thu GPU trung t\u00e2m d\u1eef li\u1ec7u c\u1ee7a h\u1ecd s\u1ebd \u0111\u1ea1t kho\u1ea3ng $2 t\u1ef7 v\u00e0o n\u0103m 2024. C\u00f4ng ty c\u0169ng \u01b0\u1edbc t\u00ednh t\u1ed5ng gi\u00e1 tr\u1ecb th\u1ecb tr\u01b0\u1eddng GPU AI c\u00f3 th\u1ec3 t\u0103ng l\u00ean $400 t\u1ef7 trong 4 n\u0103m t\u1edbi. B\u00e0 Su n\u00f3i th\u00eam: \u201cNvidia s\u1ebd chi\u1ebfm \u0111\u01b0\u1ee3c mi\u1ebfng b\u00e1nh l\u1edbn nh\u01b0ng ch\u00fang t\u00f4i c\u0169ng s\u1ebd nh\u1eadn \u0111\u01b0\u1ee3c ph\u1ea7n chia c\u00f4ng b\u1eb1ng.<\/p>\n

B\u1ea3o L\u00e2m<\/strong> (d\u1ef1a theo Wccftech, Toms Ph\u1ea7n c\u1ee9ng<\/em> ) <\/p>\n

<\/div>\n