{"id":1825,"date":"2019-06-13T02:40:36","date_gmt":"2019-06-13T02:40:36","guid":{"rendered":"https:\/\/www.thegioimaychu.vn\/blog\/?p=1825"},"modified":"2024-08-16T10:01:48","modified_gmt":"2024-08-16T03:01:48","slug":"kinh-nghiem-lua-chon-gpu-card-chuyen-dung-cho-ai-deep-learning","status":"publish","type":"post","link":"https:\/\/thegioimaychu.vn\/blog\/tong-hop\/kinh-nghiem-lua-chon-gpu-card-chuyen-dung-cho-ai-deep-learning-p1825\/","title":{"rendered":"Kinh nghi\u1ec7m l\u1ef1a ch\u1ecdn GPU card chuy\u00ean d\u1ee5ng cho AI\/Deep Learning"},"content":{"rendered":"<style>.iblink{color:#975!important;}.iblink:hover{color:orange!important;}<\/style><p><em><a class=\"iblink\" href=\"\/blog\/tag\/deep-learning\/\">Deep Learning<\/a> l\u00e0 m\u1ed9t l\u0129nh v\u1ef1c c\u00f3 c\u00e1c y\u00eau c\u1ea7u x\u1eed l\u00fd \u0111\u1eb7c bi\u1ec7t v\u00e0 vi\u1ec7c l\u1ef1a ch\u1ecdn GPU c\u1ee7a b\u1ea1n s\u1ebd quy\u1ebft \u0111\u1ecbnh r\u1ea5t l\u1edbn \u0111\u1ebfn qu\u00e1 tr\u00ecnh tri\u1ec3n khai.\u00a0V\u1eady nh\u1eefng y\u1ebfu t\u1ed1 n\u00e0o l\u00e0 quan tr\u1ecdng khi b\u1ea1n quy\u1ebft \u0111\u1ecbnh mua GPU m\u1edbi cho m\u1ee5c \u0111\u00edch <a class=\"iblink\" href=\"\/blog\/tag\/deep-learning\/\">Deep Learning<\/a>?\u00a0RAM, GPU, Core, Tensor Core?\u00a0L\u1ef1a ch\u1ecdn nh\u01b0 th\u1ebf n\u00e0o \u0111\u1ec3 hi\u1ec7u qu\u1ea3 nh\u1ea5t v\u1ec1 chi ph\u00ed \u0111\u1ea7u t\u01b0?\u00a0B\u00e0i n\u00e0y s\u1ebd \u0111i s\u00e2u v\u00e0o nh\u1eefng c\u00e2u h\u1ecfi v\u00e0 s\u1ebd cho b\u1ea1n l\u1eddi khuy\u00ean gi\u00fap b\u1ea1n \u0111\u01b0a ra l\u1ef1a ch\u1ecdn ph\u00f9 h\u1ee3p.<\/em><\/p>\n<p>Vi\u1ec7c c\u00f3 trong tay GPU ch\u1ea1y nhanh l\u00e0 m\u1ed9t y\u1ebfu t\u1ed1 r\u1ea5t quan tr\u1ecdng khi b\u1eaft \u0111\u1ea7u d\u1ef1 \u00e1n Deep Learning v\u00ec \u0111i\u1ec1u n\u00e0y cho ph\u00e9p \u0111\u1ea1t \u0111\u01b0\u1ee3c k\u1ebft qu\u1ea3 nhanh ch\u00f3ng theo kinh nghi\u1ec7m th\u1ef1c t\u1ebf, \u0111\u00f3 l\u00e0 ch\u00eca kh\u00f3a \u0111\u1ec3 x\u00e2y d\u1ef1ng chuy\u00ean m\u00f4n m\u00e0 b\u1ea1n s\u1ebd c\u00f3 th\u1ec3 \u00e1p d\u1ee5ng Deep Learning v\u00e0o c\u00e1c v\u1ea5n \u0111\u1ec1 m\u1edbi.\u00a0Thi\u1ebfu thi\u1ebft b\u1ecb c\u00f3 s\u1ee9c m\u1ea1nh ph\u00f9 h\u1ee3p, b\u1ea1n s\u1ebd m\u1ea5t qu\u00e1 nhi\u1ec1u th\u1eddi gian \u0111\u1ec3 h\u1ecdc h\u1ecfi t\u1eeb nh\u1eefng sai l\u1ea7m c\u1ee7a ng\u01b0\u1eddi kh\u00e1c v\u00e0 c\u00f3 th\u1ec3 n\u1ea3n l\u00f2ng khi ti\u1ebfp t\u1ee5c tri\u1ec3n khai d\u1ef1 \u00e1n Deep Learning.\u00a0V\u1edbi GPU, t\u00f4i \u0111\u00e3 h\u1ecdc \u0111\u01b0\u1ee3c c\u00e1ch \u00e1p d\u1ee5ng Deep Learning v\u00e0o m\u1ed9t lo\u1ea1t c\u00e1c cu\u1ed9c thi Kaggle v\u00e0 t\u00f4i \u0111\u00e3 gi\u00e0nh \u0111\u01b0\u1ee3c v\u1ecb tr\u00ed th\u1ee9 hai trong Partly Sunny v\u1edbi cu\u1ed9c thi Chance of Hashtags Kaggle b\u1eb1ng c\u00e1ch s\u1eed d\u1ee5ng\u00a0ph\u01b0\u01a1ng ph\u00e1p Deep Learning, trong \u0111\u00f3 c\u00f3 m\u1ed9t nhi\u1ec7m v\u1ee5 x\u1ebfp h\u1ea1ng vi\u1ec7c d\u1ef1 b\u00e1o\u00a0 th\u1eddi ti\u1ebft.\u00a0Trong cu\u1ed9c thi, t\u00f4i \u0111\u00e3 s\u1eed d\u1ee5ng m\u1ed9t deep neural network (DNN) hai l\u1edbp c\u00f3 quy m\u00f4 l\u1edbn v\u1edbi c\u00e1c \u0111\u01a1n v\u1ecb tuy\u1ebfn t\u00ednh \u0111\u01b0\u1ee3c \u0111i\u1ec1u ch\u1ec9nh \u0111\u1ec3 chu\u1ea9n h\u00f3a v\u00e0 DNN n\u00e0y ch\u1ec9 v\u1eeba \u0111\u1ee7 trong b\u1ed9 nh\u1edb GPU 6GB c\u1ee7a t\u00f4i.\u00a0C\u00e1c GPU GTX Titan h\u1ed7 tr\u1ee3 t\u00f4i trong cu\u1ed9c thi l\u00e0 y\u1ebfu t\u1ed1 ch\u00ednh gi\u00fap t\u00f4i \u0111\u1ea1t v\u1ecb tr\u00ed th\u1ee9 2 trong cu\u1ed9c thi.<\/p>\r\n\t\t\t\t<div class=\"tgmcad hor\" style=\"text-align:center;margin:30px 0; padding: 10px 0; border-radius:5px; border-top:solid 1px #ccc; border-bottom:solid 1px #ccc; \">\r\n\t\t\t\t<a href=\"https:\/\/thegioimaychu.vn\/giai-phap\/deep-learning\/\"><img decoding=\"async\" alt=\"Gi\u1ea3i ph\u00e1p h\u1ea1 t\u1ea7ng Deep Learning, Tr\u00ed tu\u1ec7 Nh\u00e2n t\u1ea1o - AI\" src=\"https:\/\/thegioimaychu.vn\/blog\/wp-content\/uploads\/2019\/11\/tgmc-deep-learning.jpg\" \/><\/a>\r\n\t\t\t\t<\/div>\n<h2>T\u1ed5ng quan<\/h2>\n<p>B\u00e0i n\u00e0y \u0111\u01b0\u1ee3c c\u1ea5u tr\u00fac theo c\u00e1ch sau.\u00a0\u0110\u1ea7u ti\u00ean t\u00f4i b\u00e0n v\u1ec1 vi\u1ec7c c\u00f3 nhi\u1ec1u GPU s\u1ebd h\u1eefu \u00edch nh\u01b0 th\u1ebf n\u00e0o, sau \u0111\u00f3 t\u00f4i s\u1ebd th\u1ea3o lu\u1eadn v\u1ec1 t\u1ea5t c\u1ea3 c\u00e1c t\u00f9y ch\u1ecdn ph\u1ea7n c\u1ee9ng c\u00f3 li\u00ean quan nh\u01b0 GPU NVIDIA v\u00e0 AMD, Intel Xeon Phi, Google TPU v\u00e0 ph\u1ea7n c\u1ee9ng cho startup m\u1edbi.\u00a0Sau \u0111\u00f3, t\u00f4i th\u1ea3o lu\u1eadn v\u1ec1 th\u00f4ng s\u1ed1 k\u1ef9 thu\u1eadt GPU l\u00e0 ch\u1ec9 s\u1ed1 t\u1ed1t cho hi\u1ec7u su\u1ea5t c\u1ee7a Deep Learning.\u00a0Ph\u1ea7n ch\u00ednh th\u1ea3o lu\u1eadn v\u1ec1 ph\u00e2n t\u00edch hi\u1ec7u su\u1ea5t v\u00e0 hi\u1ec7u qu\u1ea3 chi ph\u00ed.\u00a0T\u00f4i k\u1ebft lu\u1eadn v\u1edbi c\u00e1c g\u1ee3i \u00fd GPU v\u1eeba t\u1ed5ng quan, v\u1eeba c\u1ee5 th\u1ec3 h\u01a1n.<\/p>\n<h2>Nhi\u1ec1u GPU l\u00e0m cho vi\u1ec7c training c\u1ee7a t\u00f4i nhanh h\u01a1n?<\/h2>\n<p>Khi b\u1eaft \u0111\u1ea7u s\u1eed d\u1ee5ng nhi\u1ec1u GPU, t\u00f4i r\u1ea5t h\u00e0o h\u1ee9ng v\u1edbi vi\u1ec7c thi\u1ebft l\u1eadp x\u1eed l\u00fd song song \u0111\u1ec3 c\u1ea3i thi\u1ec7n hi\u1ec7u n\u0103ng th\u1eddi gian ch\u1ea1y cho cu\u1ed9c thi Kaggle.\u00a0Tuy nhi\u00ean, t\u00f4i th\u1ea5y r\u1eb1ng r\u1ea5t kh\u00f3 \u0111\u1ec3 \u0111\u1ea1t \u0111\u01b0\u1ee3c t\u1ed1c \u0111\u1ed9 mong mu\u1ed1n ch\u1ec9 \u0111\u01a1n gi\u1ea3n b\u1eb1ng c\u00e1ch s\u1eed d\u1ee5ng nhi\u1ec1u GPU.\u00a0T\u00f2 m\u00f2 v\u1ec1 v\u1ea5n \u0111\u1ec1 n\u00e0y, t\u00f4i b\u1eaft \u0111\u1ea7u th\u1ef1c hi\u1ec7n nghi\u00ean c\u1ee9u v\u1ea5n \u0111\u1ec1 x\u1eed l\u00fd song song trong h\u1ecdc Deep Learning.\u00a0T\u00f4i \u0111\u00e3 ph\u00e2n t\u00edch vi\u1ec7c x\u1eed l\u00fd song song trong c\u00e1c ki\u1ebfn \u200b\u200btr\u00fac Deep Learning, ph\u00e1t tri\u1ec3n\u00a0k\u1ef9 thu\u1eadt l\u01b0\u1ee3ng t\u1eed h\u00f3a 8 bit\u00a0\u0111\u1ec3 t\u0103ng t\u1ed1c \u0111\u1ed9 trong c\u00e1c c\u1ee5m GPU t\u1eeb 23x l\u00ean 50x cho h\u1ec7 th\u1ed1ng 96 GPU v\u00e0 c\u00f4ng b\u1ed1 nghi\u00ean c\u1ee9u c\u1ee7a t\u00f4i t\u1ea1i ICLR 2016.<\/p>\n<p>\u0110i\u1ec3m quan tr\u1ecdng \u0111\u01b0\u1ee3c nh\u1eadn ra l\u00e0 c\u00e1c network\u00a0thu\u1ed9c ki\u1ec3u &#8220;convolution&#8221; v\u00e0 &#8220;recurrent&#8221; l\u1ea1i kh\u00e1 d\u1ec5 \u0111\u1ec3 song song h\u00f3a, \u0111\u1eb7c bi\u1ec7t n\u1ebfu b\u1ea1n ch\u1ec9 s\u1eed d\u1ee5ng m\u1ed9t m\u00e1y t\u00ednh ho\u1eb7c 4 GPU.\u00a0Tuy nhi\u00ean, c\u00e1c network \u0111\u01b0\u1ee3c k\u1ebft n\u1ed1i to\u00e0n ph\u1ea7n, bao g\u1ed3m c\u1ea3 c\u00e1c transformer, th\u00ec kh\u00f4ng \u0111\u01a1n gi\u1ea3n \u0111\u1ec3 song song v\u00e0 c\u1ea7n c\u00e1c thu\u1eadt to\u00e1n chuy\u00ean d\u1ee5ng \u0111\u1ec3 th\u1ef1c hi\u1ec7n t\u1ed1t.<\/p>\n<figure id=\"attachment_9\"><img loading=\"lazy\" decoding=\"async\" style=\"box-sizing: inherit; border: 0px; max-width: 100%; height: auto;\" src=\"https:\/\/thegioimaychu.vn\/blog\/wp-content\/uploads\/-000\/\/1\/gpu-pic.jpg\" alt=\"GPU pic\" width=\"680\" height=\"356\" data-attachment-id=\"9\" data-permalink=\"https:\/\/timdettmers.com\/2019\/04\/03\/which-gpu-for-deep-learning\/gpu-pic\/\" data-orig-file=\"https:\/\/i1.wp.com\/timdettmers.com\/wp-content\/uploads\/2014\/08\/gpu-pic.jpg?fit=2916%2C1524&amp;ssl=1\" data-orig-size=\"2916,1524\" data-comments-opened=\"1\" data-image-meta=\"{&quot;aperture&quot;:&quot;2.4&quot;,&quot;credit&quot;:&quot;&quot;,&quot;camera&quot;:&quot;U9200&quot;,&quot;caption&quot;:&quot;&quot;,&quot;created_timestamp&quot;:&quot;1407955348&quot;,&quot;copyright&quot;:&quot;&quot;,&quot;focal_length&quot;:&quot;4.13&quot;,&quot;iso&quot;:&quot;64&quot;,&quot;shutter_speed&quot;:&quot;0.033333&quot;,&quot;title&quot;:&quot;&quot;,&quot;orientation&quot;:&quot;1&quot;}\" data-image-title=\"GPU pic\" data-image-description=\"&lt;p&gt;Setup in my main computer: You can see three GXT Titan and an InfiniBand card. Is this a good setup for doing deep learning?&lt;\/p&gt; \" data-medium-file=\"https:\/\/i1.wp.com\/timdettmers.com\/wp-content\/uploads\/2014\/08\/gpu-pic.jpg?fit=300%2C157&amp;ssl=1\" data-large-file=\"https:\/\/i1.wp.com\/timdettmers.com\/wp-content\/uploads\/2014\/08\/gpu-pic.jpg?fit=1024%2C535&amp;ssl=1\" \/><figcaption><em>H\u00ecnh 1: Thi\u1ebft l\u1eadp trong m\u00e1y t\u00ednh ch\u00ednh c\u1ee7a t\u00f4i: B\u1ea1n c\u00f3 th\u1ec3 th\u1ea5y ba GPU v\u00e0 th\u1ebb InfiniBand.\u00a0\u0110\u00e2y c\u00f3 ph\u1ea3i l\u00e0 m\u1ed9t thi\u1ebft l\u1eadp t\u1ed1t \u0111\u1ec3 l\u00e0m h\u1ecdc t\u1eadp s\u00e2u?<\/em><\/figcaption><\/figure>\n<p>C\u00e1c th\u01b0 vi\u1ec7n hi\u1ec7n \u0111\u1ea1i nh\u01b0 TensorFlow v\u00e0 PyTorch r\u1ea5t ph\u00f9 h\u1ee3p \u0111\u1ec3 k\u1ebft n\u1ed1i song song c\u00e1c m\u1ea1ng l\u1eb7p v\u00e0 t\u00edch ch\u1eadp, v\u00e0 \u0111\u1ec3 t\u00edch ch\u1eadp, b\u1ea1n c\u00f3 th\u1ec3 mong \u0111\u1ee3i t\u1ed1c \u0111\u1ed9 t\u0103ng l\u00ean kho\u1ea3ng 1,9x \/ 2,8x \/ 3,5x cho 2\/3\/4 GPU.\u00a0\u0110\u1ed1i v\u1edbi c\u00e1c m\u1ea1ng l\u1eb7p l\u1ea1i, \u0111\u1ed9 d\u00e0i chu\u1ed7i l\u00e0 tham s\u1ed1 quan tr\u1ecdng nh\u1ea5t v\u00e0 \u0111\u1ed1i v\u1edbi c\u00e1c s\u1ef1 c\u1ed1 NLP ph\u1ed5 bi\u1ebfn, ng\u01b0\u1eddi ta c\u00f3 th\u1ec3 mong \u0111\u1ee3i c\u00e1c t\u1ed1c \u0111\u1ed9 t\u0103ng t\u1ed1c t\u01b0\u01a1ng t\u1ef1 ho\u1eb7c k\u00e9m h\u01a1n m\u1ed9t ch\u00fat so v\u1edbi c\u00e1c m\u1ea1ng ch\u1eadp.\u00a0C\u00e1c m\u1ea1ng k\u1ebft n\u1ed1i \u0111\u1ea7y \u0111\u1ee7, bao g\u1ed3m c\u1ea3 m\u00e1y bi\u1ebfn \u00e1p, tuy nhi\u00ean, th\u01b0\u1eddng c\u00f3 hi\u1ec7u su\u1ea5t k\u00e9m \u0111\u1ed1i v\u1edbi d\u1eef li\u1ec7u song song v\u00e0 c\u00e1c thu\u1eadt to\u00e1n ti\u00ean ti\u1ebfn h\u01a1n l\u00e0 c\u1ea7n thi\u1ebft \u0111\u1ec3 t\u0103ng t\u1ed1c c\u00e1c ph\u1ea7n n\u00e0y c\u1ee7a m\u1ea1ng.\u00a0N\u1ebfu b\u1ea1n ch\u1ea1y bi\u1ebfn \u00e1p tr\u00ean nhi\u1ec1u GPU, b\u1ea1n n\u00ean th\u1eed ch\u1ea1y n\u00f3 tr\u00ean 1 GPU v\u00e0 xem n\u00f3 c\u00f3 nhanh h\u01a1n hay kh\u00f4ng.<\/p>\n<h2>S\u1eed d\u1ee5ng nhi\u1ec1u GPU m\u00e0 kh\u00f4ng c\u1ea7n x\u1eed l\u00fd song song<\/h2>\n<p>M\u1ed9t l\u1ee3i th\u1ebf kh\u00e1c c\u1ee7a vi\u1ec7c s\u1eed d\u1ee5ng nhi\u1ec1u GPU, ngay c\u1ea3 khi b\u1ea1n kh\u00f4ng song song h\u00f3a c\u00e1c thu\u1eadt to\u00e1n, l\u00e0 b\u1ea1n c\u00f3 th\u1ec3 ch\u1ea1y nhi\u1ec1u thu\u1eadt to\u00e1n ho\u1eb7c th\u1eed nghi\u1ec7m ri\u00eang bi\u1ec7t tr\u00ean m\u1ed7i GPU.\u00a0T\u00ecm ki\u1ebfm hyperparameter m\u1ed9t c\u00e1ch hi\u1ec7u qu\u1ea3 l\u00e0 c\u00e1ch v\u1eadn d\u1ee5ng ph\u1ed5 bi\u1ebfn nh\u1ea5t c\u1ee7a thi\u1ebft l\u1eadp nhi\u1ec1u GPU.\u00a0B\u1ea1n kh\u00f4ng t\u0103ng t\u1ed1c, nh\u01b0ng b\u1ea1n c\u00f3 \u0111\u01b0\u1ee3c th\u00f4ng tin nhanh h\u01a1n v\u1ec1 hi\u1ec7u su\u1ea5t c\u1ee7a c\u00e1c thi\u1ebft l\u1eadp\u00a0hyperparameter kh\u00e1c nhau ho\u1eb7c ki\u1ebfn \u200b\u200btr\u00fac m\u1ea1ng kh\u00e1c nhau.\u00a0\u0110i\u1ec1u n\u00e0y c\u0169ng r\u1ea5t h\u1eefu \u00edch cho nh\u1eefng ng\u01b0\u1eddi m\u1edbi, v\u00ec b\u1ea1n c\u00f3 th\u1ec3 nhanh ch\u00f3ng hi\u1ec3u bi\u1ebft v\u00e0 tr\u1ea3i nghi\u1ec7m v\u1ec1 c\u00e1ch b\u1ea1n c\u00f3 th\u1ec3 training m\u1ed9t ki\u1ebfn \u200b\u200btr\u00fac Deep Learning kh\u00f4ng quen thu\u1ed9c.<\/p>\n<p>S\u1eed d\u1ee5ng nhi\u1ec1u GPU theo c\u00e1ch n\u00e0y th\u01b0\u1eddng h\u1eefu \u00edch h\u01a1n so v\u1edbi vi\u1ec7c ch\u1ea1y m\u1ed9t m\u1ea1ng tr\u00ean nhi\u1ec1u GPU th\u00f4ng qua song song d\u1eef li\u1ec7u.\u00a0B\u1ea1n n\u00ean ghi nh\u1edb \u0111i\u1ec1u n\u00e0y khi b\u1ea1n mua nhi\u1ec1u GPU: Ch\u1ea5t l\u01b0\u1ee3ng \u0111\u1ec3 c\u00f3 s\u1ef1 song song t\u1ed1t h\u01a1n nh\u01b0 s\u1ed1 l\u01b0\u1ee3ng l\u00e0n PCIe kh\u00f4ng quan tr\u1ecdng khi b\u1ea1n mua nhi\u1ec1u GPU.<\/p>\n<p>Ngo\u00e0i ra, l\u01b0u \u00fd r\u1eb1ng m\u1ed9t GPU duy nh\u1ea5t ph\u1ea3i \u0111\u1ee7 cho h\u1ea7u h\u1ebft m\u1ecdi t\u00e1c v\u1ee5.\u00a0Do \u0111\u00f3, ph\u1ea1m vi tr\u1ea3i nghi\u1ec7m m\u00e0 b\u1ea1n c\u00f3 th\u1ec3 c\u00f3 v\u1edbi 1 GPU s\u1ebd kh\u00f4ng kh\u00e1c so v\u1edbi khi b\u1ea1n c\u00f3 4 GPU.\u00a0S\u1ef1 kh\u00e1c bi\u1ec7t duy nh\u1ea5t l\u00e0 b\u1ea1n c\u00f3 th\u1ec3 ch\u1ea1y nhi\u1ec1u th\u1eed nghi\u1ec7m h\u01a1n trong m\u1ed9t th\u1eddi gian nh\u1ea5t \u0111\u1ecbnh v\u1edbi nhi\u1ec1u GPU.<\/p>\n<h2>T\u00f9y ch\u1ecdn c\u1ee7a b\u1ea1n: NVIDIA vs AMD vs Intel vs Google vs Amazon vs Microsoft vs Fancy Startup<\/h2>\n<h3>NVIDIA: Ng\u01b0\u1eddi d\u1eabn \u0111\u1ea7u<\/h3>\n<p>C\u00e1c th\u01b0 vi\u1ec7n ti\u00eau chu\u1ea9n c\u1ee7a NVIDIA gi\u00fap d\u1ec5 d\u00e0ng thi\u1ebft l\u1eadp c\u00e1c th\u01b0 vi\u1ec7n Deep Learning \u0111\u1ea7u ti\u00ean trong CUDA, trong khi kh\u00f4ng c\u00f3 th\u01b0 vi\u1ec7n ti\u00eau chu\u1ea9n m\u1ea1nh nh\u01b0 v\u1eady cho OpenCL c\u1ee7a AMD.\u00a0L\u1ee3i th\u1ebf ban \u0111\u1ea7u n\u00e0y k\u1ebft h\u1ee3p v\u1edbi s\u1ef1 h\u1ed7 tr\u1ee3 c\u1ed9ng \u0111\u1ed3ng m\u1ea1nh m\u1ebd t\u1eeb NVIDIA \u0111\u00e3 t\u0103ng quy m\u00f4 c\u1ee7a c\u1ed9ng \u0111\u1ed3ng CUDA m\u1ed9t c\u00e1ch nhanh ch\u00f3ng.\u00a0\u0110i\u1ec1u n\u00e0y c\u00f3 ngh\u0129a l\u00e0 n\u1ebfu b\u1ea1n s\u1eed d\u1ee5ng GPU NVIDIA, b\u1ea1n s\u1ebd d\u1ec5 d\u00e0ng t\u00ecm th\u1ea5y s\u1ef1 h\u1ed7 tr\u1ee3 n\u1ebfu c\u00f3 s\u1ef1 c\u1ed1 x\u1ea3y ra, b\u1ea1n s\u1ebd t\u00ecm th\u1ea5y s\u1ef1 h\u1ed7 tr\u1ee3 v\u00e0 l\u1eddi khuy\u00ean n\u1ebfu b\u1ea1n t\u1ef1 l\u1eadp tr\u00ecnh CUDA v\u00e0 b\u1ea1n s\u1ebd th\u1ea5y r\u1eb1ng h\u1ea7u h\u1ebft c\u00e1c th\u01b0 vi\u1ec7n Deep Learning \u0111\u1ec1u c\u00f3 h\u1ed7 tr\u1ee3 t\u1ed1t nh\u1ea5t cho GPU NVIDIA.\u00a0Trong nh\u1eefng th\u00e1ng cu\u1ed1i, NVIDIA v\u1eabn r\u00f3t th\u00eam nhi\u1ec1u t\u00e0i nguy\u00ean v\u00e0o\u00a0ph\u1ea7n m\u1ec1m\u00a0.\u00a0V\u00ed d\u1ee5:\u00a0th\u01b0 vi\u1ec7n Apexcung c\u1ea5p h\u1ed7 tr\u1ee3 \u0111\u1ec3 \u1ed5n \u0111\u1ecbnh \u0111\u1ed9 d\u1ed1c 16 bit trong PyTorch v\u00e0 c\u0169ng bao g\u1ed3m c\u00e1c tr\u00ecnh t\u1ed1i \u01b0u h\u00f3a nhanh \u0111\u01b0\u1ee3c h\u1ee3p nh\u1ea5t nh\u01b0 FuseAdam.\u00a0Nh\u00ecn chung, ph\u1ea7n m\u1ec1m l\u00e0 m\u1ed9t \u0111i\u1ec3m r\u1ea5t m\u1ea1nh \u0111\u1ed1i v\u1edbi GPU NVIDIA.<\/p>\n<p>M\u1eb7t kh\u00e1c, NVIDIA hi\u1ec7n c\u00f3 ch\u00ednh s\u00e1ch r\u1eb1ng vi\u1ec7c s\u1eed d\u1ee5ng CUDA trong c\u00e1c trung t\u00e2m d\u1eef li\u1ec7u ch\u1ec9 \u0111\u01b0\u1ee3c ph\u00e9p cho GPU Tesla ch\u1ee9 kh\u00f4ng ph\u1ea3i th\u1ebb GTX ho\u1eb7c RTX.\u00a0Kh\u00f4ng r\u00f5 \u00fd ngh\u0129a c\u1ee7a c\u00e1c trung t\u00e2m d\u1eef li\u1ec7u c\u1ee7a Cameron l\u00e0 g\u00ec nh\u01b0ng \u0111i\u1ec1u n\u00e0y c\u00f3 ngh\u0129a l\u00e0 c\u00e1c t\u1ed5 ch\u1ee9c v\u00e0 tr\u01b0\u1eddng \u0111\u1ea1i h\u1ecdc th\u01b0\u1eddng b\u1ecb bu\u1ed9c ph\u1ea3i mua GPU Tesla \u0111\u1eaft ti\u1ec1n v\u00e0 kh\u00f4ng hi\u1ec7u qu\u1ea3 do lo ng\u1ea1i c\u00e1c v\u1ea5n \u0111\u1ec1 ph\u00e1p l\u00fd.\u00a0Tuy nhi\u00ean, th\u1ebb Tesla kh\u00f4ng c\u00f3 l\u1ee3i th\u1ebf th\u1ef1c s\u1ef1 so v\u1edbi th\u1ebb GTX v\u00e0 RTX v\u00e0 c\u00f3 gi\u00e1 cao g\u1ea5p 10 l\u1ea7n.<\/p>\n<p>NVIDIA ch\u1ec9 c\u00f3 th\u1ec3 l\u00e0m \u0111i\u1ec1u n\u00e0y m\u00e0 kh\u00f4ng g\u1eb7p b\u1ea5t k\u1ef3 tr\u1edf ng\u1ea1i l\u1edbn n\u00e0o cho th\u1ea5y s\u1ee9c m\u1ea1nh c\u1ee7a s\u1ef1 \u0111\u1ed9c quy\u1ec1n c\u1ee7a h\u1ecd &#8211; h\u1ecd c\u00f3 th\u1ec3 l\u00e0m theo \u00fd m\u00ecnh v\u00e0 ch\u00fang t\u00f4i ph\u1ea3i ch\u1ea5p nh\u1eadn c\u00e1c \u0111i\u1ec1u kho\u1ea3n.\u00a0N\u1ebfu b\u1ea1n ch\u1ecdn nh\u1eefng l\u1ee3i th\u1ebf l\u1edbn m\u00e0 GPU NVIDIA c\u00f3 \u0111\u01b0\u1ee3c v\u1ec1 m\u1eb7t c\u1ed9ng \u0111\u1ed3ng v\u00e0 h\u1ed7 tr\u1ee3, b\u1ea1n c\u0169ng s\u1ebd c\u1ea7n ph\u1ea3i ch\u1ea5p nh\u1eadn r\u1eb1ng b\u1ea1n c\u00f3 th\u1ec3 \u0111\u01b0\u1ee3c \u0111\u1ea9y xung quanh theo \u00fd mu\u1ed1n.<\/p>\n<h3>AMD: H\u1ed7 tr\u1ee3 m\u1ea1nh m\u1ebd nh\u01b0ng thi\u1ebfu<\/h3>\n<p>HIP\u00a0th\u00f4ng qua\u00a0ROCm\u00a0h\u1ee3p nh\u1ea5t GPU NVIDIA v\u00e0 AMD theo ng\u00f4n ng\u1eef l\u1eadp tr\u00ecnh chung \u0111\u01b0\u1ee3c bi\u00ean d\u1ecbch th\u00e0nh ng\u00f4n ng\u1eef GPU t\u01b0\u01a1ng \u1ee9ng tr\u01b0\u1edbc khi \u0111\u01b0\u1ee3c bi\u00ean d\u1ecbch th\u00e0nh l\u1eafp r\u00e1p GPU.\u00a0N\u1ebfu ch\u00fang ta c\u00f3 t\u1ea5t c\u1ea3 m\u00e3 GPU c\u1ee7a m\u00ecnh trong HIP th\u00ec \u0111\u00e2y s\u1ebd l\u00e0 m\u1ed9t c\u1ed9t m\u1ed1c quan tr\u1ecdng, nh\u01b0ng \u0111i\u1ec1u n\u00e0y kh\u00e1 kh\u00f3 kh\u0103n v\u00ec r\u1ea5t kh\u00f3 \u0111\u1ec3 chuy\u1ec3n c\u00e1c c\u01a1 s\u1edf m\u00e3 TensorFlow v\u00e0 PyTorch.\u00a0TensorFlow v\u00e0 PyTorch c\u00f3 m\u1ed9t s\u1ed1 h\u1ed7 tr\u1ee3 cho GPU AMD v\u00e0 t\u1ea5t c\u1ea3 c\u00e1c m\u1ea1ng ch\u00ednh c\u00f3 th\u1ec3 ch\u1ea1y tr\u00ean GPU AMD, nh\u01b0ng n\u1ebfu b\u1ea1n mu\u1ed1n ph\u00e1t tri\u1ec3n m\u1ea1ng m\u1edbi, m\u1ed9t s\u1ed1 chi ti\u1ebft c\u00f3 th\u1ec3 b\u1ecb thi\u1ebfu khi\u1ebfn b\u1ea1n kh\u00f4ng th\u1ec3 th\u1ef1c hi\u1ec7n nh\u1eefng g\u00ec b\u1ea1n c\u1ea7n.\u00a0C\u1ed9ng \u0111\u1ed3ng ROCm c\u0169ng kh\u00f4ng qu\u00e1 l\u1edbn v\u00e0 do \u0111\u00f3 kh\u00f4ng \u0111\u01a1n gi\u1ea3n \u0111\u1ec3 kh\u1eafc ph\u1ee5c s\u1ef1 c\u1ed1 nhanh ch\u00f3ng.\u00a0AMD \u0111\u1ea7u t\u01b0 r\u1ea5t \u00edt v\u00e0o ph\u1ea7n m\u1ec1m Deep Learning c\u1ee7a h\u1ecd v\u00e0 v\u00ec th\u1ebf ng\u01b0\u1eddi ta kh\u00f4ng th\u1ec3 ng\u1edd r\u1eb1ng kho\u1ea3ng c\u00e1ch ph\u1ea7n m\u1ec1m gi\u1eefa NVIDIA v\u00e0 AMD s\u1ebd \u0111\u00f3ng l\u1ea1i.<\/p>\n<p>Hi\u1ec7n t\u1ea1i, hi\u1ec7u su\u1ea5t c\u1ee7a GPU AMD v\u1eabn \u1ed5n.\u00a0Hi\u1ec7n t\u1ea1i ch\u00fang c\u00f3 kh\u1ea3 n\u0103ng t\u00ednh to\u00e1n 16 bit, \u0111\u00e2y l\u00e0 m\u1ed9t c\u1ed9t m\u1ed1c quan tr\u1ecdng, tuy nhi\u00ean, c\u00e1c GPU c\u1ee7a Tensor c\u1ee7a NVIDIA cung c\u1ea5p hi\u1ec7u n\u0103ng t\u00ednh to\u00e1n v\u01b0\u1ee3t tr\u1ed9i h\u01a1n nhi\u1ec1u cho c\u00e1c m\u00e1y bi\u1ebfn \u00e1p v\u00e0 m\u1ea1ng ch\u1eadp (m\u1eb7c d\u00f9 kh\u00f4ng nhi\u1ec1u cho c\u00e1c m\u1ea1ng l\u1eb7p l\u1ea1i c\u1ea5p \u0111\u1ed9 t\u1eeb).<\/p>\n<p>Nh\u00ecn chung, t\u00f4i ngh\u0129 r\u1eb1ng t\u00f4i v\u1eabn kh\u00f4ng th\u1ec3 \u0111\u01b0a ra m\u1ed9t khuy\u1ebfn ngh\u1ecb r\u00f5 r\u00e0ng v\u1ec1 GPU AMD cho ng\u01b0\u1eddi d\u00f9ng th\u00f4ng th\u01b0\u1eddng ch\u1ec9 mu\u1ed1n GPU c\u1ee7a h\u1ecd ho\u1ea1t \u0111\u1ed9ng tr\u01a1n tru.\u00a0Ng\u01b0\u1eddi d\u00f9ng c\u00f3 nhi\u1ec1u kinh nghi\u1ec7m h\u01a1n s\u1ebd c\u00f3 \u00edt v\u1ea5n \u0111\u1ec1 h\u01a1n v\u00e0 b\u1eb1ng c\u00e1ch h\u1ed7 tr\u1ee3 GPU AMD v\u00e0 nh\u00e0 ph\u00e1t tri\u1ec3n ROCm \/ HIP, h\u1ecd g\u00f3p ph\u1ea7n v\u00e0o cu\u1ed9c chi\u1ebfn ch\u1ed1ng l\u1ea1i v\u1ecb th\u1ebf \u0111\u1ed9c quy\u1ec1n c\u1ee7a NVIDIA v\u00ec \u0111i\u1ec1u n\u00e0y s\u1ebd gi\u00fap \u00edch r\u1ea5t nhi\u1ec1u cho m\u1ecdi ng\u01b0\u1eddi v\u1ec1 l\u00e2u d\u00e0i.\u00a0N\u1ebfu b\u1ea1n l\u00e0 nh\u00e0 ph\u00e1t tri\u1ec3n GPU v\u00e0 mu\u1ed1n \u0111\u00f3ng g\u00f3p quan tr\u1ecdng cho \u0111i\u1ec7n to\u00e1n GPU, th\u00ec GPU AMD c\u00f3 th\u1ec3 l\u00e0 c\u00e1ch t\u1ed1t nh\u1ea5t \u0111\u1ec3 t\u1ea1o ra t\u00e1c \u0111\u1ed9ng t\u1ed1t trong d\u00e0i h\u1ea1n.\u00a0\u0110\u1ed1i v\u1edbi nh\u1eefng ng\u01b0\u1eddi kh\u00e1c, GPU NVIDIA c\u00f3 th\u1ec3 l\u00e0 l\u1ef1a ch\u1ecdn an to\u00e0n h\u01a1n.<\/p>\n<h3>Intel: S\u1ebd ph\u1ea3i c\u1ed1 g\u1eafng nhi\u1ec1u<\/h3>\n<p>Tr\u1ea3i nghi\u1ec7m c\u00e1 nh\u00e2n c\u1ee7a t\u00f4i v\u1edbi Intel Xeon Phi (d\u00f2ng chip n\u00e0y \u0111\u00e3 ng\u01b0ng ph\u00e1t tri\u1ec3n t\u1eeb n\u0103m 2020 v\u00e0 \u0111\u01b0\u1ee3c thay th\u1ebf b\u1eb1ng <a href=\"https:\/\/www.intel.com\/content\/www\/us\/en\/products\/details\/processors\/ai-accelerators\/gaudi.html\" target=\"_blank\" rel=\"noopener\">Intel Gaudi<\/a>) l\u00e0 r\u1ea5t \u0111\u00e1ng th\u1ea5t v\u1ecdng v\u00e0 t\u00f4i kh\u00f4ng th\u1ea5y h\u1ecd l\u00e0 \u0111\u1ed1i th\u1ee7 th\u1ef1c s\u1ef1 c\u1ee7a NVIDIA hay AMD v\u00e0 do \u0111\u00f3 t\u00f4i s\u1ebd n\u00f3i ng\u1eafn g\u1ecdn: N\u1ebfu b\u1ea1n quy\u1ebft \u0111\u1ecbnh \u0111i v\u1edbi Xeon Phi, h\u00e3y l\u01b0u \u00fd r\u1eb1ng b\u1ea1n c\u00f3 th\u1ec3 g\u1eb7p ph\u1ea3i ng\u01b0\u1eddi ngh\u00e8o h\u1ed7 tr\u1ee3, c\u00e1c v\u1ea5n \u0111\u1ec1 \u0111i\u1ec7n to\u00e1n l\u00e0m cho c\u00e1c ph\u1ea7n m\u00e3 ch\u1eadm h\u01a1n CPU, kh\u00f3 kh\u0103n trong vi\u1ec7c vi\u1ebft m\u00e3 \u0111\u01b0\u1ee3c t\u1ed1i \u01b0u h\u00f3a, kh\u00f4ng h\u1ed7 tr\u1ee3 \u0111\u1ea7y \u0111\u1ee7 c\u00e1c t\u00ednh n\u0103ng c\u1ee7a C ++ 11, kh\u00f4ng h\u1ed7 tr\u1ee3 m\u1ed9t s\u1ed1 m\u1eabu thi\u1ebft k\u1ebf GPU quan tr\u1ecdng, kh\u1ea3 n\u0103ng t\u01b0\u01a1ng th\u00edch k\u00e9m v\u1edbi c\u00e1c th\u01b0 vi\u1ec7n kh\u00e1c d\u1ef1a tr\u00ean c\u00e1c th\u00f3i quen BLAS (NumPy v\u00e0 SciPy) v\u00e0 c\u00f3 l\u1ebd nhi\u1ec1u n\u1ed7i th\u1ea5t v\u1ecdng kh\u00e1c m\u00e0 t\u00f4i kh\u00f4ng g\u1eb7p ph\u1ea3i.<\/p>\n<p>Ngo\u00e0i Xeon Phi, t\u00f4i th\u1ef1c s\u1ef1 mong \u0111\u1ee3i b\u1ed9 x\u1eed l\u00fd m\u1ea1ng th\u1ea7n kinh Intel Nervana (NNP) v\u00ec th\u00f4ng s\u1ed1 k\u1ef9 thu\u1eadt c\u1ee7a n\u00f3 c\u1ef1c k\u1ef3 m\u1ea1nh m\u1ebd trong tay m\u1ed9t nh\u00e0 ph\u00e1t tri\u1ec3n GPU v\u00e0 n\u00f3 s\u1ebd cho ph\u00e9p c\u00e1c thu\u1eadt to\u00e1n m\u1edbi c\u00f3 th\u1ec3 x\u00e1c \u0111\u1ecbnh l\u1ea1i c\u00e1ch s\u1eed d\u1ee5ng m\u1ea1ng th\u1ea7n kinh, nh\u01b0ng n\u00f3 \u0111\u00e3 b\u1ecb tr\u00ec ho\u00e3n v\u00f4 t\u1eadn v\u00e0 c\u00f3 tin \u0111\u1ed3n r\u1eb1ng m\u1ed9t ph\u1ea7n l\u1edbn c\u1ee7a s\u1ef1 ph\u00e1t tri\u1ec3n \u0111\u00e3 nh\u1ea3y thuy\u1ec1n.\u00a0NNP \u0111\u01b0\u1ee3c l\u00ean k\u1ebf ho\u1ea1ch cho qu\u00fd 3 n\u0103m 2019. N\u1ebfu b\u1ea1n mu\u1ed1n ch\u1edd \u0111\u1ee3i l\u00e2u nh\u01b0 v\u1eady, h\u00e3y nh\u1edb r\u1eb1ng m\u1ed9t ph\u1ea7n c\u1ee9ng t\u1ed1t kh\u00f4ng ph\u1ea3i l\u00e0 t\u1ea5t c\u1ea3 nh\u1eefng g\u00ec ch\u00fang ta c\u00f3 th\u1ec3 th\u1ea5y t\u1eeb Xeon Phi c\u1ee7a AMD v\u00e0 Intel.\u00a0N\u00f3 c\u00f3 th\u1ec3 l\u00e0 v\u00e0o n\u0103m 2020 ho\u1eb7c 2021 cho \u0111\u1ebfn khi NNP c\u1ea1nh tranh v\u1edbi GPU ho\u1eb7c TPU.<\/p>\n<h3>Google: X\u1eed l\u00fd theo y\u00eau c\u1ea7u m\u1ea1nh m\u1ebd, gi\u00e1 r\u1ebb<\/h3>\n<p>Google TPU \u0111\u00e3 ph\u00e1t tri\u1ec3n th\u00e0nh m\u1ed9t s\u1ea3n ph\u1ea9m d\u1ef1a tr\u00ean \u0111\u00e1m m\u00e2y r\u1ea5t tr\u01b0\u1edfng th\u00e0nh, ti\u1ebft ki\u1ec7m chi ph\u00ed.\u00a0C\u00e1ch d\u1ec5 nh\u1ea5t \u0111\u1ec3 hi\u1ec3u \u00fd ngh\u0129a c\u1ee7a TPU l\u00e0 xem n\u00f3 nh\u01b0 nhi\u1ec1u GPU chuy\u00ean d\u1ee5ng \u0111\u01b0\u1ee3c \u0111\u00f3ng g\u00f3i c\u00f9ng nhau ch\u1ec9 c\u00f3 m\u1ed9t m\u1ee5c \u0111\u00edch: Th\u1ef1c hi\u1ec7n ph\u00e9p nh\u00e2n ma tr\u1eadn nhanh.\u00a0N\u1ebfu ch\u00fang ta xem x\u00e9t c\u00e1c ph\u00e9p \u0111o hi\u1ec7u su\u1ea5t c\u1ee7a V100 h\u1ed7 tr\u1ee3 core-core so v\u1edbi TPUv2, ch\u00fang ta th\u1ea5y r\u1eb1ng c\u1ea3 hai h\u1ec7 th\u1ed1ng c\u00f3 hi\u1ec7u su\u1ea5t g\u1ea7n nh\u01b0 gi\u1ed1ng nhau cho ResNet50 [m\u1ea5t ngu\u1ed3n, kh\u00f4ng ph\u1ea3i tr\u00ean Wayback Machine].\u00a0Tuy nhi\u00ean, Google TPU ti\u1ebft ki\u1ec7m chi ph\u00ed h\u01a1n.\u00a0V\u00ec TPU c\u00f3 c\u01a1 s\u1edf h\u1ea1 t\u1ea7ng song song tinh vi, TPU s\u1ebd c\u00f3 l\u1ee3i \u00edch t\u1ed1c \u0111\u1ed9 l\u1edbn h\u01a1n GPU n\u1ebfu b\u1ea1n s\u1eed d\u1ee5ng nhi\u1ec1u h\u01a1n 1 \u0111\u00e1m m\u00e2y TPU (t\u01b0\u01a1ng \u0111\u01b0\u01a1ng v\u1edbi 4 GPU).<\/p>\n<p>M\u1eb7c d\u00f9 v\u1eabn c\u00f2n th\u1eed nghi\u1ec7m, PyTorch hi\u1ec7n c\u0169ng h\u1ed7 tr\u1ee3 TPU s\u1ebd gi\u00fap c\u1ee7ng c\u1ed1 c\u1ed9ng \u0111\u1ed3ng v\u00e0 h\u1ec7 sinh th\u00e1i TPU.<\/p>\n<p>TPU v\u1eabn c\u00f3 m\u1ed9t s\u1ed1 v\u1ea5n \u0111\u1ec1 \u1edf \u0111\u00e2y v\u00e0 \u1edf \u0111\u00f3, v\u00ed d\u1ee5, m\u1ed9t b\u00e1o c\u00e1o t\u1eeb th\u00e1ng 2 n\u0103m 2018 n\u00f3i r\u1eb1ng TPUv2 kh\u00f4ng h\u1ed9i t\u1ee5 khi s\u1eed d\u1ee5ng LSTM.\u00a0T\u00f4i kh\u00f4ng th\u1ec3 t\u00ecm th\u1ea5y ngu\u1ed3n n\u1ebfu s\u1ef1 c\u1ed1 \u0111\u00e3 \u0111\u01b0\u1ee3c kh\u1eafc ph\u1ee5c.<\/p>\n<p>M\u1eb7t kh\u00e1c, c\u00f3 m\u1ed9t c\u00e2u chuy\u1ec7n th\u00e0nh c\u00f4ng l\u1edbn cho vi\u1ec7c training c\u00e1c m\u00e1y bi\u1ebfn \u00e1p l\u1edbn tr\u00ean TPU.\u00a0C\u00e1c m\u00f4 h\u00ecnh d\u1ecbch thu\u1eadt GPT-2, BERT v\u00e0 m\u00e1y c\u00f3 th\u1ec3 \u0111\u01b0\u1ee3c training r\u1ea5t hi\u1ec7u qu\u1ea3 tr\u00ean TPU.\u00a0Theo \u01b0\u1edbc t\u00ednh c\u1ee7a t\u00f4i t\u1eeb\u00a0b\u00e0i \u0111\u0103ng tr\u00ean blog TPU v\u00e0 GPU c\u1ee7a t\u00f4i, TPU\u00a0nhanh h\u01a1n kho\u1ea3ng 56% so v\u1edbi GPU v\u00e0 nh\u1edd gi\u00e1 th\u1ea5p h\u01a1n so v\u1edbi GPU \u0111\u00e1m m\u00e2y, ch\u00fang l\u00e0 m\u1ed9t l\u1ef1a ch\u1ecdn tuy\u1ec7t v\u1eddi cho c\u00e1c d\u1ef1 \u00e1n bi\u1ebfn \u00e1p l\u1edbn.<\/p>\n<p>Tuy nhi\u00ean, m\u1ed9t v\u1ea5n \u0111\u1ec1 v\u1edbi vi\u1ec7c training c\u00e1c m\u00f4 h\u00ecnh l\u1edbn tr\u00ean TPU c\u00f3 th\u1ec3 l\u00e0 chi ph\u00ed t\u00edch l\u0169y.\u00a0TPU c\u00f3 hi\u1ec7u su\u1ea5t cao \u0111\u01b0\u1ee3c s\u1eed d\u1ee5ng t\u1ed1t nh\u1ea5t trong giai \u0111o\u1ea1n training.\u00a0Trong giai \u0111o\u1ea1n t\u1ea1o m\u1eabu v\u00e0 suy lu\u1eadn, b\u1ea1n n\u00ean d\u1ef1a v\u00e0o c\u00e1c t\u00f9y ch\u1ecdn kh\u00f4ng ph\u1ea3i \u0111\u00e1m m\u00e2y \u0111\u1ec3 gi\u1ea3m chi ph\u00ed.\u00a0Do \u0111\u00f3, training v\u1ec1 TPU, nh\u01b0ng t\u1ea1o m\u1eabu v\u00e0 suy lu\u1eadn v\u1ec1 GPU c\u00e1 nh\u00e2n c\u1ee7a b\u1ea1n l\u00e0 l\u1ef1a ch\u1ecdn t\u1ed1t nh\u1ea5t.<\/p>\n<p>\u0110\u1ec3 k\u1ebft lu\u1eadn, hi\u1ec7n t\u1ea1i, TPU d\u01b0\u1eddng nh\u01b0 \u0111\u01b0\u1ee3c s\u1eed d\u1ee5ng t\u1ed1t nh\u1ea5t \u0111\u1ec3 training m\u1ea1ng t\u00edch ch\u1eadp ho\u1eb7c m\u00e1y bi\u1ebfn \u00e1p l\u1edbn v\u00e0 n\u00ean \u0111\u01b0\u1ee3c b\u1ed5 sung b\u1eb1ng c\u00e1c t\u00e0i nguy\u00ean t\u00ednh to\u00e1n kh\u00e1c thay v\u00ec t\u00e0i nguy\u00ean Deep Learning ch\u00ednh.<\/p>\n<h3>Amazon AWS v\u00e0 Microsoft Azure: \u0110\u00e1ng tin c\u1eady nh\u01b0ng \u0111\u1eaft ti\u1ec1n<\/h3>\n<p>C\u00e1c phi\u00ean b\u1ea3n GPU t\u1eeb Amazon AWS v\u00e0 Microsoft Azure r\u1ea5t h\u1ea5p d\u1eabn b\u1edfi v\u00ec ng\u01b0\u1eddi ta c\u00f3 th\u1ec3 d\u1ec5 d\u00e0ng t\u0103ng quy m\u00f4 v\u00e0 thu nh\u1ecf d\u1ef1a tr\u00ean nhu c\u1ea7u.\u00a0\u0110i\u1ec1u n\u00e0y r\u1ea5t h\u1eefu \u00edch cho th\u1eddi h\u1ea1n gi\u1ea5y ho\u1eb7c cho c\u00e1c d\u1ef1 \u00e1n m\u1ed9t l\u1ea7n l\u1edbn h\u01a1n.\u00a0Tuy nhi\u00ean, t\u01b0\u01a1ng t\u1ef1 nh\u01b0 TPU, chi ph\u00ed th\u00f4 t\u0103ng l\u00ean nhanh ch\u00f3ng.\u00a0Hi\u1ec7n t\u1ea1i, c\u00e1c phi\u00ean b\u1ea3n \u0111\u00e1m m\u00e2y GPU qu\u00e1 \u0111\u1eaft \u0111\u1ec3 s\u1eed d\u1ee5ng ri\u00eang r\u1ebd v\u00e0 t\u00f4i khuy\u00ean b\u1ea1n n\u00ean c\u00f3 m\u1ed9t s\u1ed1 GPU gi\u00e1 r\u1ebb chuy\u00ean d\u1ee5ng \u0111\u1ec3 t\u1ea1o m\u1eabu tr\u01b0\u1edbc khi b\u1eaft \u0111\u1ea7u c\u00e1c c\u00f4ng vi\u1ec7c training cu\u1ed1i c\u00f9ng trong \u0111\u00e1m m\u00e2y.<\/p>\n<h3>Startup \u01b0a th\u00edch: Kh\u00e1i ni\u1ec7m ph\u1ea7n c\u1ee9ng mang t\u00ednh c\u00e1ch m\u1ea1ng kh\u00f4ng c\u00f3 ph\u1ea7n m\u1ec1m<\/h3>\n<p>C\u00f3 m\u1ed9t lo\u1ea1t c\u00e1c c\u00f4ng ty kh\u1edfi nghi\u1ec7p nh\u1eb1m t\u1ea1o ra th\u1ebf h\u1ec7 ph\u1ea7n c\u1ee9ng Deep Learning ti\u1ebfp theo.\u00a0C\u00e1c c\u00f4ng ty n\u00e0y th\u01b0\u1eddng c\u00f3 m\u1ed9t thi\u1ebft k\u1ebf l\u00fd thuy\u1ebft tuy\u1ec7t v\u1eddi, sau \u0111\u00f3 \u0111\u01b0\u1ee3c Google \/ Intel ho\u1eb7c nh\u1eefng ng\u01b0\u1eddi kh\u00e1c mua \u0111\u1ec3 c\u00f3 \u0111\u01b0\u1ee3c t\u00e0i tr\u1ee3 m\u00e0 h\u1ecd c\u1ea7n \u0111\u1ec3 t\u1ea1o ra m\u1ed9t thi\u1ebft k\u1ebf \u0111\u1ea7y \u0111\u1ee7 v\u00e0 s\u1ea3n xu\u1ea5t chip.\u00a0\u0110\u1ed1i v\u1edbi th\u1ebf h\u1ec7 chip ti\u1ebfp theo (3nm), chi ph\u00ed cho vi\u1ec7c n\u00e0y l\u00e0 kho\u1ea3ng 1 t\u1ef7 USD tr\u01b0\u1edbc khi m\u1ed9t con chip c\u00f3 th\u1ec3 \u0111\u01b0\u1ee3c s\u1ea3n xu\u1ea5t.\u00a0Khi giai \u0111o\u1ea1n n\u00e0y \u0111\u01b0\u1ee3c ho\u00e0n th\u00e0nh (ch\u01b0a c\u00f3 c\u00f4ng ty n\u00e0o qu\u1ea3n l\u00fd \u0111\u1ec3 th\u1ef1c hi\u1ec7n vi\u1ec7c n\u00e0y), v\u1ea5n \u0111\u1ec1 ch\u00ednh l\u00e0 ph\u1ea7n m\u1ec1m.\u00a0Kh\u00f4ng c\u00f3 c\u00f4ng ty n\u00e0o qu\u1ea3n l\u00fd \u0111\u1ec3 s\u1ea3n xu\u1ea5t ph\u1ea7n m\u1ec1m s\u1ebd ho\u1ea1t \u0111\u1ed9ng trong kho h\u1ecdc t\u1eadp s\u00e2u hi\u1ec7n t\u1ea1i.\u00a0M\u1ed9t b\u1ed9 ph\u1ea7n m\u1ec1m \u0111\u1ea7y \u0111\u1ee7 c\u1ea7n \u0111\u01b0\u1ee3c ph\u00e1t tri\u1ec3n \u0111\u1ec3 c\u1ea1nh tranh, \u0111i\u1ec1u n\u00e0y r\u00f5 r\u00e0ng t\u1eeb v\u00ed d\u1ee5 AMD vs NVIDIA: AMD c\u00f3 ph\u1ea7n c\u1ee9ng tuy\u1ec7t v\u1eddi nh\u01b0ng ch\u1ec9 c\u00f3 90% ph\u1ea7n m\u1ec1m &#8211; \u0111i\u1ec1u n\u00e0y kh\u00f4ng \u0111\u1ee7 \u0111\u1ec3 c\u1ea1nh tranh v\u1edbi NVIDIA.<\/p>\n<p>Hi\u1ec7n t\u1ea1i, kh\u00f4ng c\u00f3 c\u00f4ng ty n\u00e0o \u1edf g\u1ea7n ho\u00e0n th\u00e0nh c\u1ea3 hai b\u01b0\u1edbc ph\u1ea7n c\u1ee9ng v\u00e0 ph\u1ea7n m\u1ec1m.\u00a0Intel NNP c\u00f3 th\u1ec3 l\u00e0 g\u1ea7n nh\u1ea5t, nh\u01b0ng t\u1eeb t\u1ea5t c\u1ea3 nh\u1eefng th\u1ee9 n\u00e0y kh\u00f4ng th\u1ec3 mong \u0111\u1ee3i m\u1ed9t s\u1ea3n ph\u1ea9m c\u1ea1nh tranh tr\u01b0\u1edbc n\u0103m 2020 ho\u1eb7c 2021. V\u00ec v\u1eady, hi\u1ec7n t\u1ea1i ch\u00fang ta s\u1ebd c\u1ea7n ph\u1ea3i g\u1eafn b\u00f3 v\u1edbi GPU v\u00e0 TPU.<\/p>\n<p>Do \u0111\u00f3, ph\u1ea7n c\u1ee9ng m\u1edbi l\u1ea1 m\u1eaft t\u1eeb kh\u1edfi \u0111\u1ed9ng y\u00eau th\u00edch c\u1ee7a b\u1ea1n c\u00f3 th\u1ec3 \u0111\u01b0\u1ee3c b\u1ecf qua m\u1ed9t c\u00e1ch an to\u00e0n cho \u0111\u1ebfn b\u00e2y gi\u1edd.<\/p>\n<h2>\u0110i\u1ec1u g\u00ec l\u00e0m cho m\u1ed9t GPU nhanh h\u01a1n m\u1ed9t GPU kh\u00e1c?<\/h2>\n<p>C\u00e2u h\u1ecfi \u0111\u1ea7u ti\u00ean c\u1ee7a b\u1ea1n c\u00f3 th\u1ec3 l\u00e0 t\u00ednh n\u0103ng quan tr\u1ecdng nh\u1ea5t cho hi\u1ec7u n\u0103ng GPU nhanh \u0111\u1ec3 Deep Learning: \u0110\u00f3 c\u00f3 ph\u1ea3i l\u00e0 core CUDA kh\u00f4ng?\u00a0T\u1ed1c \u0111\u1ed9 \u0111\u1ed3ng h\u1ed3?\u00a0Dung l\u01b0\u1ee3ng RAM?<\/p>\n<p>N\u0103m 2019, vi\u1ec7c l\u1ef1a ch\u1ecdn GPU tr\u1edf n\u00ean kh\u00f3 hi\u1ec3u h\u01a1n bao gi\u1edd h\u1ebft: \u0111i\u1ec7n to\u00e1n 16 bit, core c\u0103ng, GPU 16 bit kh\u00f4ng c\u00f3 core k\u00e9o, nhi\u1ec1u th\u1ebf h\u1ec7 GPU v\u1eabn c\u00f2n kh\u1ea3 thi (Turn, Volta, Maxwell).\u00a0Nh\u01b0ng v\u1eabn c\u00f3 m\u1ed9t s\u1ed1 ch\u1ec9 s\u1ed1 hi\u1ec7u su\u1ea5t \u0111\u00e1ng tin c\u1eady m\u00e0 m\u1ecdi ng\u01b0\u1eddi c\u00f3 th\u1ec3 s\u1eed d\u1ee5ng nh\u01b0 m\u1ed9t quy t\u1eafc.\u00a0D\u01b0\u1edbi \u0111\u00e2y l\u00e0 m\u1ed9t s\u1ed1 h\u01b0\u1edbng d\u1eabn \u01b0u ti\u00ean cho c\u00e1c ki\u1ebfn \u200b\u200btr\u00fac Deep Learning kh\u00e1c nhau:<\/p>\n<p>M\u1ea1ng k\u1ebft h\u1ee3p v\u00e0 m\u00e1y bi\u1ebfn \u00e1p: core k\u00e9o&gt; FLOPs&gt; B\u0103ng th\u00f4ng b\u1ed9 nh\u1edb&gt; Kh\u1ea3 n\u0103ng 16 bit<br \/>\nM\u1ea1ng t\u00e1i ph\u00e1t: B\u0103ng th\u00f4ng b\u1ed9 nh\u1edb&gt; Kh\u1ea3 n\u0103ng 16 bit&gt; core k\u00e9o&gt; FLOP<\/p>\n<p>\u0110i\u1ec1u n\u00e0y c\u00f3 n\u1ed9i dung nh\u01b0 sau: N\u1ebfu t\u00f4i mu\u1ed1n s\u1eed d\u1ee5ng, v\u00ed d\u1ee5, m\u1ea1ng ch\u1eadp, tr\u01b0\u1edbc ti\u00ean t\u00f4i n\u00ean \u01b0u ti\u00ean GPU c\u00f3 core c\u0103ng, sau \u0111\u00f3 l\u00e0 s\u1ed1 FLOP cao, sau \u0111\u00f3 l\u00e0 b\u0103ng th\u00f4ng b\u1ed9 nh\u1edb cao v\u00e0 sau \u0111\u00f3 l\u00e0 GPU c\u00f3 kh\u1ea3 n\u0103ng 16 bit .\u00a0Trong khi \u01b0u ti\u00ean, \u0111i\u1ec1u quan tr\u1ecdng l\u00e0 ch\u1ecdn m\u1ed9t GPU c\u00f3 \u0111\u1ee7 b\u1ed9 nh\u1edb GPU \u0111\u1ec3 ch\u1ea1y c\u00e1c m\u00f4 h\u00ecnh m\u00e0 ng\u01b0\u1eddi ta quan t\u00e2m.<\/p>\n<h2>T\u1ea1i sao c\u1ea7n nh\u1eefng s\u1ef1 \u01b0u ti\u00ean n\u00e0y?<\/h2>\n<p>M\u1ed9t \u0111i\u1ec1u gi\u00fap t\u0103ng c\u01b0\u1eddng s\u1ef1 hi\u1ec3u bi\u1ebft c\u1ee7a b\u1ea1n \u0111\u1ec3 \u0111\u01b0a ra l\u1ef1a ch\u1ecdn s\u00e1ng su\u1ed1t l\u00e0 t\u00ecm hi\u1ec3u m\u1ed9t ch\u00fat v\u1ec1 ph\u1ea7n n\u00e0o c\u1ee7a ph\u1ea7n c\u1ee9ng gi\u00fap GPU nhanh cho hai ho\u1ea1t \u0111\u1ed9ng c\u0103ng th\u1eb3ng quan tr\u1ecdng nh\u1ea5t: Nh\u00e2n ma tr\u1eadn v\u00e0 t\u00edch ch\u1eadp.<\/p>\n<p>M\u1ed9t c\u00e1ch \u0111\u01a1n gi\u1ea3n v\u00e0 hi\u1ec7u qu\u1ea3 \u0111\u1ec3 suy ngh\u0129 v\u1ec1 ph\u00e9p nh\u00e2n ma tr\u1eadn A * B = C l\u00e0 gi\u1edbi h\u1ea1n b\u0103ng th\u00f4ng b\u1ed9 nh\u1edb: Sao ch\u00e9p b\u1ed9 nh\u1edb c\u1ee7a A, B cho \u0111\u1ebfn khi chip t\u1ed1n k\u00e9m h\u01a1n so v\u1edbi vi\u1ec7c t\u00ednh to\u00e1n A * B.\u00a0\u0110i\u1ec1u n\u00e0y c\u00f3 ngh\u0129a l\u00e0 b\u0103ng th\u00f4ng b\u1ed9 nh\u1edb l\u00e0 t\u00ednh n\u0103ng quan tr\u1ecdng nh\u1ea5t c\u1ee7a GPU n\u1ebfu b\u1ea1n mu\u1ed1n s\u1eed d\u1ee5ng c\u00e1c LSTM v\u00e0 c\u00e1c m\u1ea1ng l\u1eb7p l\u1ea1i kh\u00e1c th\u1ef1c hi\u1ec7n nhi\u1ec1u ph\u00e9p nh\u00e2n ma tr\u1eadn nh\u1ecf.\u00a0Ph\u00e9p nh\u00e2n ma tr\u1eadn c\u00e0ng nh\u1ecf, b\u0103ng th\u00f4ng b\u1ed9 nh\u1edb c\u00e0ng quan tr\u1ecdng.<\/p>\n<p>Ng\u01b0\u1ee3c l\u1ea1i, t\u00edch ch\u1eadp b\u1ecb r\u00e0ng bu\u1ed9c b\u1edfi t\u1ed1c \u0111\u1ed9 t\u00ednh to\u00e1n.\u00a0Do \u0111\u00f3, TFLOP tr\u00ean GPU l\u00e0 ch\u1ec9 s\u1ed1 t\u1ed1t nh\u1ea5t cho hi\u1ec7u su\u1ea5t c\u1ee7a ResNets v\u00e0 c\u00e1c ki\u1ebfn \u200b\u200btr\u00fac t\u00edch ch\u1eadp kh\u00e1c.\u00a0core k\u00e9o c\u00f3 th\u1ec3 t\u0103ng FLOP \u0111\u00e1ng k\u1ec3.<\/p>\n<p>Ph\u00e9p nh\u00e2n ma tr\u1eadn l\u1edbn nh\u01b0 \u0111\u01b0\u1ee3c s\u1eed d\u1ee5ng trong c\u00e1c m\u00e1y bi\u1ebfn \u00e1p n\u1eb1m gi\u1eefa t\u00edch ch\u1eadp v\u00e0 ph\u00e9p nh\u00e2n ma tr\u1eadn nh\u1ecf c\u1ee7a RNN.\u00a0Ph\u00e9p nh\u00e2n ma tr\u1eadn l\u1edbn \u0111\u01b0\u1ee3c h\u01b0\u1edfng l\u1ee3i r\u1ea5t nhi\u1ec1u t\u1eeb b\u1ed9 l\u01b0u tr\u1eef 16 bit, core k\u00e9o v\u00e0 FLOP nh\u01b0ng ch\u00fang v\u1eabn c\u1ea7n b\u0103ng th\u00f4ng b\u1ed9 nh\u1edb cao.<\/p>\n<p>L\u01b0u \u00fd r\u1eb1ng \u0111\u1ec3 s\u1eed d\u1ee5ng c\u00e1c l\u1ee3i \u00edch c\u1ee7a core k\u00e9o, b\u1ea1n n\u00ean s\u1eed d\u1ee5ng d\u1eef li\u1ec7u v\u00e0 tr\u1ecdng l\u01b0\u1ee3ng 16 bit &#8211; tr\u00e1nh s\u1eed d\u1ee5ng 32 bit v\u1edbi th\u1ebb RTX!\u00a0N\u1ebfu b\u1ea1n g\u1eb7p ph\u1ea3i v\u1ea5n \u0111\u1ec1 v\u1edbi training 16 bit khi s\u1eed d\u1ee5ng PyTorch, th\u00ec b\u1ea1n n\u00ean s\u1eed d\u1ee5ng t\u1ef7 l\u1ec7 m\u1ea5t \u0111\u1ed9ng nh\u01b0 \u0111\u01b0\u1ee3c cung c\u1ea5p b\u1edfi\u00a0th\u01b0 vi\u1ec7n Apex\u00a0.\u00a0N\u1ebfu b\u1ea1n s\u1eed d\u1ee5ng TensorFlow, b\u1ea1n c\u00f3 th\u1ec3 t\u1ef1 th\u1ef1c hi\u1ec7n chia t\u1ef7 l\u1ec7 t\u1ed5n th\u1ea5t: (1) nh\u00e2n s\u1ed1 l\u1ed7 c\u1ee7a b\u1ea1n v\u1edbi m\u1ed9t s\u1ed1 l\u1edbn, (2) t\u00ednh to\u00e1n \u0111\u1ed9 d\u1ed1c, (3) chia cho s\u1ed1 l\u1edbn, (4) c\u1eadp nh\u1eadt tr\u1ecdng s\u1ed1 c\u1ee7a b\u1ea1n.\u00a0Th\u00f4ng th\u01b0\u1eddng, training 16 bit s\u1ebd t\u1ed1t, nh\u01b0ng n\u1ebfu b\u1ea1n g\u1eb7p kh\u00f3 kh\u0103n trong vi\u1ec7c sao ch\u00e9p k\u1ebft qu\u1ea3 v\u1edbi t\u1ef7 l\u1ec7 m\u1ea5t 16 bit th\u01b0\u1eddng s\u1ebd gi\u1ea3i quy\u1ebft \u0111\u01b0\u1ee3c v\u1ea5n \u0111\u1ec1.<\/p>\n<figure id=\"attachment_763\"><img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/thegioimaychu.vn\/blog\/wp-content\/uploads\/-000\/\/1\/performance_RTX.png\" sizes=\"auto, (max-width: 1000px) 100vw, 1000px\" srcset=\"https:\/\/i1.wp.com\/timdettmers.com\/wp-content\/uploads\/2019\/04\/performance_RTX.png?w=1573&amp;ssl=1 1573w, https:\/\/i1.wp.com\/timdettmers.com\/wp-content\/uploads\/2019\/04\/performance_RTX.png?resize=300%2C175&amp;ssl=1 300w, https:\/\/i1.wp.com\/timdettmers.com\/wp-content\/uploads\/2019\/04\/performance_RTX.png?resize=768%2C448&amp;ssl=1 768w, https:\/\/i1.wp.com\/timdettmers.com\/wp-content\/uploads\/2019\/04\/performance_RTX.png?resize=1024%2C598&amp;ssl=1 1024w\" alt=\"H\u00ecnh 2: D\u1eef li\u1ec7u hi\u1ec7u su\u1ea5t \u0111\u01b0\u1ee3c chu\u1ea9n h\u00f3a c\u1ee7a GPU v\u00e0 TPU. Cao h\u01a1n th\u00ec t\u1ed1t h\u01a1n. Th\u1ebb RTX gi\u1ea3 \u0111\u1ecbnh t\u00ednh to\u00e1n 16 bit. C\u00e1c s\u1ed1 RNN t\u1eeb \u0111\u1ec1 c\u1eadp \u0111\u1ebfn hi\u1ec7u su\u1ea5t biLSTM cho c\u00e1c chu\u1ed7i ng\u1eafn c\u00f3 \u0111\u1ed9 d\u00e0i &lt;100. \u0110i\u1ec3m chu\u1ea9n \u0111\u01b0\u1ee3c th\u1ef1c hi\u1ec7n b\u1eb1ng PyTorch 1.0.1 v\u00e0 CUDA 10.\" width=\"680\" height=\"397\" data-attachment-id=\"763\" data-permalink=\"https:\/\/timdettmers.com\/performance_rtx\/\" data-orig-file=\"https:\/\/i1.wp.com\/timdettmers.com\/wp-content\/uploads\/2019\/04\/performance_RTX.png?fit=1573%2C918&amp;ssl=1\" data-orig-size=\"1573,918\" data-comments-opened=\"1\" data-image-meta=\"{&quot;aperture&quot;:&quot;0&quot;,&quot;credit&quot;:&quot;&quot;,&quot;camera&quot;:&quot;&quot;,&quot;caption&quot;:&quot;&quot;,&quot;created_timestamp&quot;:&quot;0&quot;,&quot;copyright&quot;:&quot;&quot;,&quot;focal_length&quot;:&quot;0&quot;,&quot;iso&quot;:&quot;0&quot;,&quot;shutter_speed&quot;:&quot;0&quot;,&quot;title&quot;:&quot;&quot;,&quot;orientation&quot;:&quot;0&quot;}\" data-image-title=\"performance_RTX\" data-image-description=\"\" data-medium-file=\"https:\/\/i1.wp.com\/timdettmers.com\/wp-content\/uploads\/2019\/04\/performance_RTX.png?fit=300%2C175&amp;ssl=1\" data-large-file=\"https:\/\/i1.wp.com\/timdettmers.com\/wp-content\/uploads\/2019\/04\/performance_RTX.png?fit=1024%2C598&amp;ssl=1\" \/><figcaption>H\u00ecnh 2: D\u1eef li\u1ec7u hi\u1ec7u su\u1ea5t \u0111\u01b0\u1ee3c chu\u1ea9n h\u00f3a c\u1ee7a GPU v\u00e0 TPU.\u00a0Cao h\u01a1n th\u00ec t\u1ed1t h\u01a1n.\u00a0Th\u1ebb RTX gi\u1ea3 \u0111\u1ecbnh t\u00ednh to\u00e1n 16 bit.\u00a0C\u00e1c s\u1ed1 RNN t\u1eeb \u0111\u1ec1 c\u1eadp \u0111\u1ebfn hi\u1ec7u su\u1ea5t biLSTM cho c\u00e1c chu\u1ed7i ng\u1eafn c\u00f3 \u0111\u1ed9 d\u00e0i &lt;100.\u00a0\u0110i\u1ec3m chu\u1ea9n \u0111\u01b0\u1ee3c th\u1ef1c hi\u1ec7n b\u1eb1ng PyTorch 1.0.1 v\u00e0 CUDA 10.<\/figcaption><\/figure>\n<h2>Ph\u00e2n t\u00edch hi\u1ec7u qu\u1ea3 chi ph\u00ed<\/h2>\n<p>Hi\u1ec7u qu\u1ea3 chi ph\u00ed c\u1ee7a GPU c\u00f3 l\u1ebd l\u00e0 ti\u00eau ch\u00ed quan tr\u1ecdng nh\u1ea5t \u0111\u1ec3 ch\u1ecdn GPU.\u00a0Ph\u00e2n t\u00edch hi\u1ec7u su\u1ea5t cho c\u1eadp nh\u1eadt b\u00e0i \u0111\u0103ng tr\u00ean blog n\u00e0y \u0111\u01b0\u1ee3c th\u1ef1c hi\u1ec7n nh\u01b0 sau:<br \/>\n(1) \u0110\u1ed1i v\u1edbi m\u00e1y bi\u1ebfn \u00e1p, t\u00f4i \u0111\u00e3 \u0111i\u1ec3m chu\u1ea9n Transformer-XL v\u00e0 BERT.<br \/>\n(2) \u0110\u1ed1i v\u1edbi RNN t\u1eeb v\u00e0 char t\u00f4i \u0111\u00e3 \u0111i\u1ec3m chu\u1ea9n c\u00e1c m\u00f4 h\u00ecnh biLSTM hi\u1ec7n \u0111\u1ea1i.<br \/>\n(3) Vi\u1ec7c \u0111o \u0111i\u1ec3m chu\u1ea9n trong (1) v\u00e0 (2) \u0111\u00e3 \u0111\u01b0\u1ee3c th\u1ef1c hi\u1ec7n cho Titan Xp, Titan RTX v\u00e0 RTX 2080 Ti.\u00a0\u0110\u1ed1i v\u1edbi c\u00e1c th\u1ebb kh\u00e1c, t\u00f4i chia t\u1ef7 l\u1ec7 ch\u00eanh l\u1ec7ch hi\u1ec7u su\u1ea5t m\u1ed9t c\u00e1ch tuy\u1ebfn t\u00ednh.<br \/>\n(4) T\u00f4i \u0111\u00e3 s\u1eed d\u1ee5ng \u0111i\u1ec3m chu\u1ea9n hi\u1ec7n c\u00f3 cho CNN: (\u00a01\u00a0,\u00a02\u00a0,\u00a03\u00a0,\u00a04\u00a0,\u00a05\u00a0,\u00a06\u00a0,\u00a07\u00a0).<br \/>\n(5) T\u00f4i \u0111\u00e3 s\u1eed d\u1ee5ng chi ph\u00ed trung b\u00ecnh c\u1ee7a Amazon v\u00e0 eBay l\u00e0m chi ph\u00ed tham chi\u1ebfu cho GPU.<\/p>\n<figure id=\"attachment_762\"><img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/thegioimaychu.vn\/blog\/wp-content\/uploads\/-000\/\/1\/performance_per_dollar_RTX.png\" sizes=\"auto, (max-width: 1000px) 100vw, 1000px\" srcset=\"https:\/\/i1.wp.com\/timdettmers.com\/wp-content\/uploads\/2019\/04\/performance_per_dollar_RTX.png?w=1573&amp;ssl=1 1573w, https:\/\/i1.wp.com\/timdettmers.com\/wp-content\/uploads\/2019\/04\/performance_per_dollar_RTX.png?resize=300%2C175&amp;ssl=1 300w, https:\/\/i1.wp.com\/timdettmers.com\/wp-content\/uploads\/2019\/04\/performance_per_dollar_RTX.png?resize=768%2C448&amp;ssl=1 768w, https:\/\/i1.wp.com\/timdettmers.com\/wp-content\/uploads\/2019\/04\/performance_per_dollar_RTX.png?resize=1024%2C598&amp;ssl=1 1024w\" alt=\"H\u00ecnh 3: S\u1ed1 hi\u1ec7u su\u1ea5t \/ chi ph\u00ed \u0111\u01b0\u1ee3c chu\u1ea9n h\u00f3a cho m\u1ea1ng t\u00edch ch\u1eadp (CNN), m\u1ea1ng l\u1eb7p l\u1ea1i (RNN) v\u00e0 m\u00e1y bi\u1ebfn \u00e1p. Cao h\u01a1n th\u00ec t\u1ed1t h\u01a1n. M\u1ed9t chi\u1ebfc RTX 2060 ti\u1ebft ki\u1ec7m chi ph\u00ed h\u01a1n 5 l\u1ea7n so v\u1edbi Tesla V100. C\u00e1c s\u1ed1 RNN t\u1eeb \u0111\u1ec1 c\u1eadp \u0111\u1ebfn hi\u1ec7u su\u1ea5t biLSTM cho c\u00e1c chu\u1ed7i ng\u1eafn c\u00f3 \u0111\u1ed9 d\u00e0i &lt;100. \u0110i\u1ec3m chu\u1ea9n \u0111\u01b0\u1ee3c th\u1ef1c hi\u1ec7n b\u1eb1ng PyTorch 1.0.1 v\u00e0 CUDA 10.\" width=\"680\" height=\"397\" data-attachment-id=\"762\" data-permalink=\"https:\/\/timdettmers.com\/performance_per_dollar_rtx\/\" data-orig-file=\"https:\/\/i1.wp.com\/timdettmers.com\/wp-content\/uploads\/2019\/04\/performance_per_dollar_RTX.png?fit=1573%2C918&amp;ssl=1\" data-orig-size=\"1573,918\" data-comments-opened=\"1\" data-image-meta=\"{&quot;aperture&quot;:&quot;0&quot;,&quot;credit&quot;:&quot;&quot;,&quot;camera&quot;:&quot;&quot;,&quot;caption&quot;:&quot;&quot;,&quot;created_timestamp&quot;:&quot;0&quot;,&quot;copyright&quot;:&quot;&quot;,&quot;focal_length&quot;:&quot;0&quot;,&quot;iso&quot;:&quot;0&quot;,&quot;shutter_speed&quot;:&quot;0&quot;,&quot;title&quot;:&quot;&quot;,&quot;orientation&quot;:&quot;0&quot;}\" data-image-title=\"performance_per_dollar_RTX\" data-image-description=\"\" data-medium-file=\"https:\/\/i1.wp.com\/timdettmers.com\/wp-content\/uploads\/2019\/04\/performance_per_dollar_RTX.png?fit=300%2C175&amp;ssl=1\" data-large-file=\"https:\/\/i1.wp.com\/timdettmers.com\/wp-content\/uploads\/2019\/04\/performance_per_dollar_RTX.png?fit=1024%2C598&amp;ssl=1\" \/><figcaption>H\u00ecnh 3: S\u1ed1 hi\u1ec7u su\u1ea5t \/ chi ph\u00ed \u0111\u01b0\u1ee3c chu\u1ea9n h\u00f3a cho m\u1ea1ng t\u00edch ch\u1eadp (CNN), m\u1ea1ng l\u1eb7p l\u1ea1i (RNN) v\u00e0 m\u00e1y bi\u1ebfn \u00e1p.\u00a0Cao h\u01a1n th\u00ec t\u1ed1t h\u01a1n.\u00a0M\u1ed9t chi\u1ebfc RTX 2060 ti\u1ebft ki\u1ec7m chi ph\u00ed h\u01a1n 5 l\u1ea7n so v\u1edbi Tesla V100.\u00a0C\u00e1c s\u1ed1 RNN t\u1eeb \u0111\u1ec1 c\u1eadp \u0111\u1ebfn hi\u1ec7u su\u1ea5t biLSTM cho c\u00e1c chu\u1ed7i ng\u1eafn c\u00f3 \u0111\u1ed9 d\u00e0i &lt;100.\u00a0\u0110i\u1ec3m chu\u1ea9n \u0111\u01b0\u1ee3c th\u1ef1c hi\u1ec7n b\u1eb1ng PyTorch 1.0.1 v\u00e0 CUDA 10.<\/figcaption><\/figure>\n<p>T\u1eeb d\u1eef li\u1ec7u n\u00e0y, ch\u00fang ta th\u1ea5y r\u1eb1ng RTX 2060 ti\u1ebft ki\u1ec7m chi ph\u00ed h\u01a1n RTX 2070, RTX 2080 ho\u1eb7c RTX 2080 Ti.\u00a0T\u1ea1i sao c\u00e1i n\u00e0y r\u1ea5t?\u00a0Kh\u1ea3 n\u0103ng th\u1ef1c hi\u1ec7n t\u00ednh to\u00e1n 16 bit v\u1edbi Tensor Cores c\u00f3 gi\u00e1 tr\u1ecb h\u01a1n nhi\u1ec1u so v\u1edbi vi\u1ec7c ch\u1ec9 c\u00f3 m\u1ed9t con t\u00e0u l\u1edbn h\u01a1n v\u1edbi nhi\u1ec1u core Tensor Cores.\u00a0V\u1edbi RTX 2060, b\u1ea1n c\u00f3 \u0111\u01b0\u1ee3c nh\u1eefng t\u00ednh n\u0103ng n\u00e0y v\u1edbi gi\u00e1 th\u1ea5p nh\u1ea5t.<\/p>\n<p>Tuy nhi\u00ean, ph\u00e2n t\u00edch n\u00e0y c\u00f3 nh\u1eefng th\u00e0nh ki\u1ebfn \u200b\u200bnh\u1ea5t \u0111\u1ecbnh c\u1ea7n \u0111\u01b0\u1ee3c t\u00ednh \u0111\u1ebfn:<br \/>\n(1) Ph\u00e2n t\u00edch n\u00e0y thi\u00ean v\u1ecb m\u1ea1nh m\u1ebd \u1ee7ng h\u1ed9 c\u00e1c th\u1ebb nh\u1ecf h\u01a1n.\u00a0GPU nh\u1ecf h\u01a1n, ti\u1ebft ki\u1ec7m chi ph\u00ed c\u00f3 th\u1ec3 kh\u00f4ng c\u00f3 \u0111\u1ee7 b\u1ed9 nh\u1edb \u0111\u1ec3 ch\u1ea1y c\u00e1c m\u00f4 h\u00ecnh m\u00e0 b\u1ea1n quan t\u00e2m!<br \/>\n(2) Gi\u00e1 qu\u00e1 cao th\u1ebb GTX 10xx: Hi\u1ec7n t\u1ea1i, th\u1ebb GTX 10XX d\u01b0\u1eddng nh\u01b0 \u0111\u01b0\u1ee3c \u0111\u1ecbnh gi\u00e1 qu\u00e1 cao do c\u00e1c game th\u1ee7 kh\u00f4ng th\u00edch th\u1ebb RTX.<br \/>\n(3) Xu h\u01b0\u1edbng GPU \u0111\u01a1n: M\u1ed9t m\u00e1y t\u00ednh c\u00f3 4 th\u1ebb kh\u00f4ng hi\u1ec7u qu\u1ea3 v\u1ec1 chi ph\u00ed (4x RTX 2080 Ti) ti\u1ebft ki\u1ec7m chi ph\u00ed h\u01a1n nhi\u1ec1u so v\u1edbi 2 m\u00e1y t\u00ednh c\u00f3 th\u1ebb chi ph\u00ed \/ hi\u1ec7u qu\u1ea3 nh\u1ea5t (8x RTX 2060).<\/p>\n<h2>C\u1ea3nh b\u00e1o: S\u1ef1 c\u1ed1 nhi\u1ec7t tr\u00ean RTX \u0111a GPU<\/h2>\n<p>C\u00f3 v\u1ea5n \u0111\u1ec1 v\u1edbi RTX 2080 Ti v\u00e0 c\u00e1c GPU RTX kh\u00e1c v\u1edbi qu\u1ea1t k\u00e9p ti\u00eau chu\u1ea9n n\u1ebfu b\u1ea1n s\u1eed d\u1ee5ng nhi\u1ec1u GPU ch\u1ea1y c\u1ea1nh nhau.\u00a0\u0110i\u1ec1u n\u00e0y \u0111\u1eb7c bi\u1ec7t nh\u01b0 v\u1eady \u0111\u1ed1i v\u1edbi nhi\u1ec1u RTX 2080 Ti trong m\u1ed9t m\u00e1y t\u00ednh nh\u01b0ng nhi\u1ec1u RTX 2080 v\u00e0 RTX 2070 c\u0169ng c\u00f3 th\u1ec3 b\u1ecb \u1ea3nh h\u01b0\u1edfng.\u00a0Qu\u1ea1t tr\u00ean m\u1ed9t s\u1ed1 th\u1ebb RTX l\u00e0 m\u1ed9t thi\u1ebft k\u1ebf m\u1edbi \u0111\u01b0\u1ee3c NVIDIA ph\u00e1t tri\u1ec3n \u0111\u1ec3 c\u1ea3i thi\u1ec7n tr\u1ea3i nghi\u1ec7m cho c\u00e1c game th\u1ee7 ch\u1ea1y m\u1ed9t GPU (im l\u1eb7ng, nhi\u1ec7t th\u1ea5p h\u01a1n cho m\u1ed9t GPU).\u00a0Tuy nhi\u00ean, thi\u1ebft k\u1ebf r\u1ea5t t\u1ec7 n\u1ebfu b\u1ea1n s\u1eed d\u1ee5ng nhi\u1ec1u GPU c\u00f3 thi\u1ebft k\u1ebf qu\u1ea1t k\u00e9p m\u1edf n\u00e0y.\u00a0N\u1ebfu b\u1ea1n mu\u1ed1n s\u1eed d\u1ee5ng nhi\u1ec1u th\u1ebb RTX ch\u1ea1y c\u1ea1nh nhau (tr\u1ef1c ti\u1ebfp trong khe PCIe ti\u1ebfp theo) th\u00ec b\u1ea1n n\u00ean l\u1ea5y phi\u00ean b\u1ea3n c\u00f3 thi\u1ebft k\u1ebf qu\u1ea1t \u0111\u01a1n ki\u1ec3u qu\u1ea1t gi\u00f3 ki\u1ec3u Drake.\u00a0\u0110i\u1ec1u n\u00e0y \u0111\u1eb7c bi\u1ec7t \u0111\u00fang v\u1edbi th\u1ebb RTX 2080 Ti.\u00a0<a class=\"iblink\" href=\"\/asus\/\">ASUS<\/a> v\u00e0 PNY hi\u1ec7n c\u00f3 c\u00e1c m\u1eabu RTX 2080 Ti tr\u00ean th\u1ecb tr\u01b0\u1eddng v\u1edbi qu\u1ea1t ki\u1ec3u qu\u1ea1t gi\u00f3.<\/p>\n<h2>K\u00edch th\u01b0\u1edbc b\u1ed9 nh\u1edb c\u1ea7n thi\u1ebft v\u00e0 training 16 bit<\/h2>\n<p>B\u1ed9 nh\u1edb tr\u00ean GPU c\u00f3 th\u1ec3 r\u1ea5t quan tr\u1ecdng \u0111\u1ed1i v\u1edbi m\u1ed9t s\u1ed1 \u1ee9ng d\u1ee5ng nh\u01b0 <a class=\"iblink\" href=\"\/blog\/tag\/computer-vision\/\">th\u1ecb gi\u00e1c m\u00e1y t\u00ednh<\/a>, d\u1ecbch m\u00e1y v\u00e0 m\u1ed9t s\u1ed1 \u1ee9ng d\u1ee5ng NLP kh\u00e1c v\u00e0 b\u1ea1n c\u00f3 th\u1ec3 ngh\u0129 r\u1eb1ng RTX 2070 ti\u1ebft ki\u1ec7m chi ph\u00ed, nh\u01b0ng b\u1ed9 nh\u1edb c\u1ee7a n\u00f3 qu\u00e1 nh\u1ecf v\u1edbi 8 GB.\u00a0Tuy nhi\u00ean, l\u01b0u \u00fd r\u1eb1ng th\u00f4ng qua training 16 bit, b\u1ea1n h\u1ea7u nh\u01b0 c\u00f3 b\u1ed9 nh\u1edb 16 GB v\u00e0 m\u1ecdi m\u00f4 h\u00ecnh ti\u00eau chu\u1ea9n s\u1ebd d\u1ec5 d\u00e0ng ph\u00f9 h\u1ee3p v\u1edbi RTX 2070 c\u1ee7a b\u1ea1n n\u1ebfu b\u1ea1n s\u1eed d\u1ee5ng 16 bit.\u00a0\u0110i\u1ec1u t\u01b0\u01a1ng t\u1ef1 c\u0169ng \u0111\u00fang v\u1edbi RTX 2080 v\u00e0 RTX 2080 Ti.\u00a0Tuy nhi\u00ean, xin l\u01b0u \u00fd r\u1eb1ng trong h\u1ea7u h\u1ebft c\u00e1c khung ph\u1ea7n m\u1ec1m, b\u1ea1n s\u1ebd kh\u00f4ng t\u1ef1 \u0111\u1ed9ng ti\u1ebft ki\u1ec7m m\u1ed9t n\u1eeda b\u1ed9 nh\u1edb b\u1eb1ng c\u00e1ch s\u1eed d\u1ee5ng 16 bit v\u00ec m\u1ed9t s\u1ed1 khung l\u01b0u tr\u1eef tr\u1ecdng l\u01b0\u1ee3ng trong 32 bit \u0111\u1ec3 th\u1ef1c hi\u1ec7n c\u1eadp nh\u1eadt \u0111\u1ed9 d\u1ed1c ch\u00ednh x\u00e1c h\u01a1n, v.v.\u00a0M\u1ed9t nguy\u00ean t\u1eafc nh\u1ecf l\u00e0 gi\u1ea3 s\u1eed th\u00eam 50% b\u1ed9 nh\u1edb v\u1edbi t\u00ednh to\u00e1n 16 bit.\u00a0V\u00ec v\u1eady, b\u1ed9 nh\u1edb 8GB 16 bit c\u00f3 k\u00edch th\u01b0\u1edbc t\u01b0\u01a1ng \u0111\u01b0\u01a1ng v\u1edbi b\u1ed9 nh\u1edb 12 GB 12 bit.<\/p>\n<h2>Khuy\u1ebfn ngh\u1ecb chung v\u1ec1 GPU<\/h2>\n<p>Hi\u1ec7n t\u1ea1i, khuy\u1ebfn ngh\u1ecb ch\u00ednh c\u1ee7a t\u00f4i l\u00e0 l\u1ea5y GPU RTX 2070 v\u00e0 s\u1eed d\u1ee5ng training 16 bit.\u00a0T\u00f4i s\u1ebd kh\u00f4ng bao gi\u1edd khuy\u00ean b\u1ea1n n\u00ean mua XP Titan, Titan V, b\u1ea5t k\u1ef3 th\u1ebb Quadro ho\u1eb7c b\u1ea5t k\u1ef3 GPU Phi\u00ean b\u1ea3n s\u00e1ng l\u1eadp n\u00e0o.\u00a0Tuy nhi\u00ean, c\u00f3 m\u1ed9t s\u1ed1 GPU c\u1ee5 th\u1ec3 c\u0169ng c\u00f3 v\u1ecb tr\u00ed c\u1ee7a ch\u00fang:<br \/>\n(1) \u0110\u1ec3 c\u00f3 th\u00eam b\u1ed9 nh\u1edb, t\u00f4i mu\u1ed1n gi\u1edbi thi\u1ec7u RTX 2080 Ti.\u00a0N\u1ebfu b\u1ea1n th\u1ef1c s\u1ef1 c\u1ea7n nhi\u1ec1u b\u1ed9 nh\u1edb b\u1ed5 sung, RTX Titan l\u00e0 l\u1ef1a ch\u1ecdn t\u1ed1t nh\u1ea5t &#8211; nh\u01b0ng h\u00e3y ch\u1eafc ch\u1eafn r\u1eb1ng b\u1ea1n th\u1ef1c s\u1ef1 c\u1ea7n b\u1ed9 nh\u1edb \u0111\u00f3!<br \/>\n(2) \u0110\u1ec3 c\u00f3 hi\u1ec7u su\u1ea5t cao h\u01a1n, t\u00f4i mu\u1ed1n gi\u1edbi thi\u1ec7u RTX 2080 Ti.<br \/>\n(3) N\u1ebfu b\u1ea1n thi\u1ebfu ti\u1ec1n, t\u00f4i s\u1ebd gi\u1edbi thi\u1ec7u b\u1ea5t k\u1ef3 th\u1ebb GTX 10XX gi\u00e1 r\u1ebb n\u00e0o t\u1eeb eBay (t\u00f9y thu\u1ed9c v\u00e0o dung l\u01b0\u1ee3ng b\u1ed9 nh\u1edb b\u1ea1n c\u1ea7n) ho\u1eb7c RTX 2060. N\u1ebfu qu\u00e1 \u0111\u1eaft, h\u00e3y xem\u00a0Colab\u00a0.<br \/>\n(4) N\u1ebfu b\u1ea1n ch\u1ec9 mu\u1ed1n b\u1eaft \u0111\u1ea7u v\u1edbi vi\u1ec7c t\u00ecm hi\u1ec3u s\u00e2u th\u00ec GTX 1060 (6GB) l\u00e0 m\u1ed9t l\u1ef1a ch\u1ecdn tuy\u1ec7t v\u1eddi.<br \/>\n(5) N\u1ebfu b\u1ea1n \u0111\u00e3 c\u00f3 GTX 1070 tr\u1edf l\u00ean: \u0110\u1ee3i n\u00f3 ra.\u00a0M\u1ed9t b\u1ea3n n\u00e2ng c\u1ea5p kh\u00f4ng \u0111\u00e1ng gi\u00e1 tr\u1eeb khi b\u1ea1n l\u00e0m vi\u1ec7c v\u1edbi c\u00e1c m\u00e1y bi\u1ebfn \u00e1p l\u1edbn.<br \/>\n(6) B\u1ea1n mu\u1ed1n h\u1ecdc nhanh c\u00e1ch Deep Learning: Nhi\u1ec1u GTX 1060 (6GB).<\/p>\n<h2>Deep Learning trong \u0111\u00e1m m\u00e2y<\/h2>\n<p>C\u1ea3 hai phi\u00ean b\u1ea3n GPU tr\u00ean AWS \/ Azure v\u00e0 TPU trong Google Cloud \u0111\u1ec1u l\u00e0 c\u00e1c t\u00f9y ch\u1ecdn kh\u1ea3 thi \u0111\u1ec3 Deep Learning.\u00a0M\u1eb7c d\u00f9 TPU r\u1ebb h\u01a1n m\u1ed9t ch\u00fat nh\u01b0ng n\u00f3 l\u1ea1i thi\u1ebfu t\u00ednh linh ho\u1ea1t v\u00e0 linh ho\u1ea1t c\u1ee7a GPU \u0111\u00e1m m\u00e2y.\u00a0TPU c\u00f3 th\u1ec3 l\u00e0 v\u0169 kh\u00ed \u0111\u01b0\u1ee3c l\u1ef1a ch\u1ecdn \u0111\u1ec3 training m\u00f4 h\u00ecnh nh\u1eadn d\u1ea1ng \u0111\u1ed1i t\u01b0\u1ee3ng ho\u1eb7c m\u00e1y bi\u1ebfn \u00e1p.\u00a0\u0110\u1ed1i v\u1edbi c\u00e1c GPU \u0111\u00e1m m\u00e2y t\u1ea3i c\u00f4ng vi\u1ec7c kh\u00e1c l\u00e0 \u0111\u1eb7t c\u01b0\u1ee3c an to\u00e0n h\u01a1n &#8211; \u0111i\u1ec1u t\u1ed1t v\u1ec1 c\u00e1c tr\u01b0\u1eddng h\u1ee3p \u0111\u00e1m m\u00e2y l\u00e0 b\u1ea1n c\u00f3 th\u1ec3 chuy\u1ec3n \u0111\u1ed5i gi\u1eefa GPU v\u00e0 TPU b\u1ea5t c\u1ee9 l\u00fac n\u00e0o ho\u1eb7c th\u1eadm ch\u00ed s\u1eed d\u1ee5ng c\u1ea3 hai c\u00f9ng m\u1ed9t l\u00fac.<\/p>\n<p>Tuy nhi\u00ean, h\u00e3y ch\u00fa \u00fd \u0111\u1ebfn chi ph\u00ed c\u01a1 h\u1ed9i \u1edf \u0111\u00e2y: N\u1ebfu b\u1ea1n h\u1ecdc c\u00e1c k\u1ef9 n\u0103ng \u0111\u1ec3 c\u00f3 m\u1ed9t lu\u1ed3ng c\u00f4ng vi\u1ec7c tr\u00f4i ch\u1ea3y v\u1edbi AWS \/ Azure, b\u1ea1n \u0111\u00e3 m\u1ea5t th\u1eddi gian c\u00f3 th\u1ec3 l\u00e0m vi\u1ec7c tr\u00ean GPU c\u00e1 nh\u00e2n v\u00e0 b\u1ea1n c\u0169ng s\u1ebd kh\u00f4ng c\u00f3 \u0111\u01b0\u1ee3c c\u00e1c k\u1ef9 n\u0103ng \u0111\u1ec3 s\u1eed d\u1ee5ng TPU.\u00a0N\u1ebfu b\u1ea1n s\u1eed d\u1ee5ng GPU c\u00e1 nh\u00e2n, b\u1ea1n s\u1ebd kh\u00f4ng c\u00f3 k\u1ef9 n\u0103ng m\u1edf r\u1ed9ng th\u00e0nh nhi\u1ec1u GPU \/ TPU h\u01a1n th\u00f4ng qua \u0111\u00e1m m\u00e2y.\u00a0N\u1ebfu b\u1ea1n s\u1eed d\u1ee5ng TPU, b\u1ea1n c\u00f3 th\u1ec3 b\u1ecb k\u1eb9t v\u1edbi TensorFlow trong m\u1ed9t th\u1eddi gian n\u1ebfu b\u1ea1n mu\u1ed1n c\u00e1c t\u00ednh n\u0103ng \u0111\u1ea7y \u0111\u1ee7 v\u00e0 s\u1ebd kh\u00f4ng \u0111\u01a1n gi\u1ea3n \u0111\u1ec3 chuy\u1ec3n c\u01a1 s\u1edf m\u00e3 c\u1ee7a b\u1ea1n sang PyTorch.\u00a0H\u1ecdc m\u1ed9t lu\u1ed3ng c\u00f4ng vi\u1ec7c GPU \/ TPU tr\u00ean \u0111\u00e1m m\u00e2y tr\u01a1n tru l\u00e0 m\u1ed9t chi ph\u00ed c\u01a1 h\u1ed9i \u0111\u1eaft \u0111\u1ecf v\u00e0 b\u1ea1n n\u00ean c\u00e2n nh\u1eafc chi ph\u00ed n\u00e0y n\u1ebfu b\u1ea1n l\u1ef1a ch\u1ecdn cho TPU, GPU \u0111\u00e1m m\u00e2y ho\u1eb7c GPU c\u00e1 nh\u00e2n.<\/p>\n<p>M\u1ed9t c\u00e2u h\u1ecfi kh\u00e1c c\u0169ng l\u00e0 v\u1ec1 khi n\u00e0o n\u00ean s\u1eed d\u1ee5ng d\u1ecbch v\u1ee5 \u0111\u00e1m m\u00e2y.\u00a0N\u1ebfu b\u1ea1n c\u1ed1 g\u1eafng Deep Learning ho\u1eb7c b\u1ea1n c\u1ea7n t\u1ea1o nguy\u00ean m\u1eabu th\u00ec GPU c\u00e1 nh\u00e2n c\u00f3 th\u1ec3 l\u00e0 l\u1ef1a ch\u1ecdn t\u1ed1t nh\u1ea5t v\u00ec c\u00e1c tr\u01b0\u1eddng h\u1ee3p \u0111\u00e1m m\u00e2y c\u00f3 th\u1ec3 t\u1ed1n k\u00e9m.\u00a0Tuy nhi\u00ean, m\u1ed9t khi b\u1ea1n \u0111\u00e3 t\u00ecm th\u1ea5y m\u1ed9t c\u1ea5u h\u00ecnh m\u1ea1ng s\u00e2u t\u1ed1t v\u00e0 b\u1ea1n ch\u1ec9 mu\u1ed1n training m\u1ed9t m\u00f4 h\u00ecnh b\u1eb1ng c\u00e1ch s\u1eed d\u1ee5ng song song d\u1eef li\u1ec7u th\u00ec s\u1eed d\u1ee5ng c\u00e1c tr\u01b0\u1eddng h\u1ee3p \u0111\u00e1m m\u00e2y l\u00e0 m\u1ed9t c\u00e1ch ti\u1ebfp c\u1eadn ch\u1eafc ch\u1eafn.\u00a0\u0110i\u1ec1u n\u00e0y c\u00f3 ngh\u0129a l\u00e0 m\u1ed9t GPU nh\u1ecf s\u1ebd \u0111\u1ee7 \u0111\u1ec3 t\u1ea1o m\u1eabu v\u00e0 ng\u01b0\u1eddi ta c\u00f3 th\u1ec3 d\u1ef1a v\u00e0o s\u1ee9c m\u1ea1nh c\u1ee7a <a class=\"iblink\" href=\"\/blog\/tag\/cloud-computing\/\">\u0111i\u1ec7n to\u00e1n \u0111\u00e1m m\u00e2y<\/a> \u0111\u1ec3 m\u1edf r\u1ed9ng quy m\u00f4 cho c\u00e1c th\u1eed nghi\u1ec7m l\u1edbn h\u01a1n.<\/p>\n<p>N\u1ebfu b\u1ea1n thi\u1ebfu ti\u1ec1n, c\u00e1c tr\u01b0\u1eddng h\u1ee3p <a class=\"iblink\" href=\"\/blog\/tag\/cloud-computing\/\">\u0111i\u1ec7n to\u00e1n \u0111\u00e1m m\u00e2y<\/a> c\u0169ng c\u00f3 th\u1ec3 l\u00e0 m\u1ed9t gi\u1ea3i ph\u00e1p t\u1ed1t: Nguy\u00ean m\u1eabu tr\u00ean CPU v\u00e0 sau \u0111\u00f3 tung ra c\u00e1c phi\u00ean b\u1ea3n GPU \/ TPU \u0111\u1ec3 ch\u1ea1y nhanh.\u00a0\u0110\u00e2y kh\u00f4ng ph\u1ea3i l\u00e0 lu\u1ed3ng c\u00f4ng vi\u1ec7c t\u1ed1t nh\u1ea5t v\u00ec vi\u1ec7c t\u1ea1o m\u1eabu tr\u00ean CPU c\u00f3 th\u1ec3 l\u00e0 m\u1ed9t n\u1ed7i \u0111au l\u1edbn, nh\u01b0ng n\u00f3 c\u00f3 th\u1ec3 l\u00e0 m\u1ed9t s\u1ef1 thay th\u1ebf hi\u1ec7u qu\u1ea3 v\u1ec1 chi ph\u00ed.<\/p>\n<h2>Ph\u1ea7n k\u1ebft lu\u1eadn<\/h2>\n<p>V\u1edbi th\u00f4ng tin trong b\u00e0i \u0111\u0103ng tr\u00ean blog n\u00e0y, b\u1ea1n s\u1ebd c\u00f3 th\u1ec3 suy lu\u1eadn GPU n\u00e0o ph\u00f9 h\u1ee3p v\u1edbi m\u00ecnh.\u00a0N\u00f3i chung, t\u00f4i th\u1ea5y ba chi\u1ebfn l\u01b0\u1ee3c ch\u00ednh (1) g\u1eafn v\u1edbi GPU GTX 1070 ho\u1eb7c t\u1ed1t h\u01a1n c\u1ee7a b\u1ea1n, (2) mua GPU RTX, (3) s\u1eed d\u1ee5ng m\u1ed9t s\u1ed1 lo\u1ea1i GPU \u0111\u1ec3 t\u1ea1o m\u1eabu v\u00e0 sau \u0111\u00f3 hu\u1ea5n luy\u1ec7n m\u00f4 h\u00ecnh c\u1ee7a b\u1ea1n tr\u00ean GPU TPU ho\u1eb7c GPU \u0111\u00e1m m\u00e2y.<\/p>\n<div>____<br \/><b>B\u00e0i vi\u1ebft li\u00ean quan<\/b><\/div>\n<ul>\n<li><a href=\"https:\/\/thegioimaychu.vn\/blog\/?post_type=post&#038;p=23415\">B\u01b0\u1edbc v\u00e0o k\u1ef7 nguy\u00ean tr\u1ee3 l\u00fd AI c\u00e1 nh\u00e2n &#8211; so s\u00e1nh chi ti\u1ebft NVIDIA RTX Spark v\u00e0 DGX Spark<\/a><\/li>\n<li><a href=\"https:\/\/thegioimaychu.vn\/blog\/?post_type=post&#038;p=23322\">NVIDIA RTX PRO 4500 Blackwell Server Edition: B\u01b0\u1edbc nh\u1ea3y v\u1ecdt t\u1eeb th\u1ebf h\u1ec7 L4 cho h\u1ea1 t\u1ea7ng AI suy lu\u1eadn hi\u1ec7u n\u0103ng cao<\/a><\/li>\n<li><a href=\"https:\/\/thegioimaychu.vn\/blog\/?post_type=post&#038;p=23119\">NVIDIA NGC\u2011ready Low\u2011latency Edge AI: Gi\u1ea3i ph\u00e1p cho Retail, Manufacturing v\u00e0 Smart Cities<\/a><\/li>\n<li><a href=\"https:\/\/thegioimaychu.vn\/blog\/?post_type=post&#038;p=23110\">So s\u00e1nh s\u1ee9c m\u1ea1nh c\u1ee7a m\u00e1y t\u00ednh AI DGX Spark v\u1edbi c\u00e1c card GPU m\u00e1y tr\u1ea1m chuy\u00ean nghi\u1ec7p c\u1ee7a NVIDIA<\/a><\/li>\n<li><a href=\"https:\/\/thegioimaychu.vn\/blog\/?post_type=post&#038;p=22969\">T\u1eeb Orin \u0111\u1ebfn Thor: B\u01b0\u1edbc nh\u1ea3y v\u1ecdt ki\u1ebfn tr\u00fac c\u1ee7a NVIDIA JetPack 7.0 v\u00e0 t\u01b0\u01a1ng lai c\u1ee7a robot h\u00ecnh ng\u01b0\u1eddi<\/a><\/li>\n<li><a href=\"https:\/\/thegioimaychu.vn\/blog\/?post_type=post&#038;p=22887\">T\u1ea1i sao g\u1ecdi NVIDIA DGX Spark l\u00e0 Si\u00eau m\u00e1y t\u00ednh AI c\u00e1 nh\u00e2n?<\/a><\/li>\n<\/ul>\n<\/p>","protected":false},"excerpt":{"rendered":"<p>Deep Learning l&agrave; m&#7897;t l&#297;nh v&#7921;c c&oacute; c&aacute;c y&ecirc;u c&#7847;u x&#7917; l&yacute; &#273;&#7863;c bi&#7879;t v&agrave; vi&#7879;c l&#7921;a ch&#7885;n GPU c&#7911;a b&#7841;n s&#7869; quy&#7871;t &#273;&#7883;nh r&#7845;t l&#7899;n &#273;&#7871;n qu&aacute; tr&igrave;nh tri&#7875;n khai.&nbsp;V&#7853;y nh&#7919;ng y&#7871;u t&#7889; n&agrave;o l&agrave; quan tr&#7885;ng khi b&#7841;n quy&#7871;t &#273;&#7883;nh mua GPU m&#7899;i cho m&#7909;c &#273;&iacute;ch Deep Learning?&nbsp;RAM, GPU, Core, Tensor Core?&nbsp;L&#7921;a ch&#7885;n&#8230;<\/p>\n","protected":false},"author":2,"featured_media":20282,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"inline_featured_image":false,"footnotes":""},"categories":[3,1],"tags":[7,12,671,1314],"class_list":["post-1825","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai-hpc","category-tong-hop","tag-ai","tag-deep-learning","tag-gpu","tag-nvidia"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v28.3 - https:\/\/yoast.com\/product\/yoast-seo-wordpress\/ -->\n<title>Kinh nghi\u1ec7m l\u1ef1a ch\u1ecdn GPU card chuy\u00ean d\u1ee5ng cho AI\/Deep Learning - Blog | TheGioiMayChu<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/thegioimaychu.vn\/blog\/tong-hop\/kinh-nghiem-lua-chon-gpu-card-chuyen-dung-cho-ai-deep-learning-p1825\/\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:title\" content=\"Kinh nghi\u1ec7m l\u1ef1a ch\u1ecdn GPU card chuy\u00ean d\u1ee5ng cho AI\/Deep Learning - Blog | TheGioiMayChu\" \/>\n<meta name=\"twitter:description\" content=\"Deep Learning l&agrave; m&#7897;t l&#297;nh v&#7921;c c&oacute; c&aacute;c y&ecirc;u c&#7847;u x&#7917; l&yacute; &#273;&#7863;c bi&#7879;t v&agrave; vi&#7879;c l&#7921;a ch&#7885;n GPU c&#7911;a b&#7841;n s&#7869; quy&#7871;t &#273;&#7883;nh r&#7845;t l&#7899;n &#273;&#7871;n qu&aacute; tr&igrave;nh tri&#7875;n khai.&nbsp;V&#7853;y nh&#7919;ng y&#7871;u t&#7889; n&agrave;o l&agrave; quan tr&#7885;ng khi b&#7841;n quy&#7871;t &#273;&#7883;nh mua GPU m&#7899;i cho m&#7909;c &#273;&iacute;ch Deep Learning?&nbsp;RAM, GPU, Core, Tensor Core?&nbsp;L&#7921;a ch&#7885;n...\" \/>\n<meta name=\"twitter:image\" content=\"https:\/\/thegioimaychu.vn\/blog\/wp-content\/uploads\/2019\/06\/TheGioiMayChu-Blog-multi-gpu-system.jpg\" \/>\n<meta name=\"twitter:label1\" content=\"Written by\" \/>\n\t<meta name=\"twitter:data1\" content=\"Vanito Hoang\" \/>\n\t<meta name=\"twitter:label2\" content=\"Est. reading time\" \/>\n\t<meta name=\"twitter:data2\" content=\"38 minutes\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\\\/\\\/schema.org\",\"@graph\":[{\"@type\":\"Article\",\"@id\":\"https:\\\/\\\/thegioimaychu.vn\\\/blog\\\/tong-hop\\\/kinh-nghiem-lua-chon-gpu-card-chuyen-dung-cho-ai-deep-learning-p1825\\\/#article\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/thegioimaychu.vn\\\/blog\\\/tong-hop\\\/kinh-nghiem-lua-chon-gpu-card-chuyen-dung-cho-ai-deep-learning-p1825\\\/\"},\"author\":{\"name\":\"Vanito Hoang\",\"@id\":\"https:\\\/\\\/thegioimaychu.vn\\\/blog\\\/#\\\/schema\\\/person\\\/feccec179330cfebfaa805639a357d18\"},\"headline\":\"Kinh nghi\u1ec7m l\u1ef1a ch\u1ecdn GPU card chuy\u00ean d\u1ee5ng cho AI\\\/Deep Learning\",\"datePublished\":\"2019-06-13T02:40:36+00:00\",\"dateModified\":\"2024-08-16T03:01:48+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\\\/\\\/thegioimaychu.vn\\\/blog\\\/tong-hop\\\/kinh-nghiem-lua-chon-gpu-card-chuyen-dung-cho-ai-deep-learning-p1825\\\/\"},\"wordCount\":7609,\"publisher\":{\"@id\":\"https:\\\/\\\/thegioimaychu.vn\\\/blog\\\/#organization\"},\"image\":{\"@id\":\"https:\\\/\\\/thegioimaychu.vn\\\/blog\\\/tong-hop\\\/kinh-nghiem-lua-chon-gpu-card-chuyen-dung-cho-ai-deep-learning-p1825\\\/#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/thegioimaychu.vn\\\/blog\\\/wp-content\\\/uploads\\\/2019\\\/06\\\/TheGioiMayChu-Blog-multi-gpu-system.jpg\",\"keywords\":[\"ai\",\"deep learning\",\"gpu\",\"NVIDIA\"],\"articleSection\":[\"AI - HPC\",\"T\u1ed5ng h\u1ee3p\"],\"inLanguage\":\"en-US\"},{\"@type\":\"WebPage\",\"@id\":\"https:\\\/\\\/thegioimaychu.vn\\\/blog\\\/tong-hop\\\/kinh-nghiem-lua-chon-gpu-card-chuyen-dung-cho-ai-deep-learning-p1825\\\/\",\"url\":\"https:\\\/\\\/thegioimaychu.vn\\\/blog\\\/tong-hop\\\/kinh-nghiem-lua-chon-gpu-card-chuyen-dung-cho-ai-deep-learning-p1825\\\/\",\"name\":\"Kinh nghi\u1ec7m l\u1ef1a ch\u1ecdn GPU card chuy\u00ean d\u1ee5ng cho AI\\\/Deep Learning - Blog | TheGioiMayChu\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/thegioimaychu.vn\\\/blog\\\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\\\/\\\/thegioimaychu.vn\\\/blog\\\/tong-hop\\\/kinh-nghiem-lua-chon-gpu-card-chuyen-dung-cho-ai-deep-learning-p1825\\\/#primaryimage\"},\"image\":{\"@id\":\"https:\\\/\\\/thegioimaychu.vn\\\/blog\\\/tong-hop\\\/kinh-nghiem-lua-chon-gpu-card-chuyen-dung-cho-ai-deep-learning-p1825\\\/#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/thegioimaychu.vn\\\/blog\\\/wp-content\\\/uploads\\\/2019\\\/06\\\/TheGioiMayChu-Blog-multi-gpu-system.jpg\",\"datePublished\":\"2019-06-13T02:40:36+00:00\",\"dateModified\":\"2024-08-16T03:01:48+00:00\",\"breadcrumb\":{\"@id\":\"https:\\\/\\\/thegioimaychu.vn\\\/blog\\\/tong-hop\\\/kinh-nghiem-lua-chon-gpu-card-chuyen-dung-cho-ai-deep-learning-p1825\\\/#breadcrumb\"},\"inLanguage\":\"en-US\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\\\/\\\/thegioimaychu.vn\\\/blog\\\/tong-hop\\\/kinh-nghiem-lua-chon-gpu-card-chuyen-dung-cho-ai-deep-learning-p1825\\\/\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"en-US\",\"@id\":\"https:\\\/\\\/thegioimaychu.vn\\\/blog\\\/tong-hop\\\/kinh-nghiem-lua-chon-gpu-card-chuyen-dung-cho-ai-deep-learning-p1825\\\/#primaryimage\",\"url\":\"https:\\\/\\\/thegioimaychu.vn\\\/blog\\\/wp-content\\\/uploads\\\/2019\\\/06\\\/TheGioiMayChu-Blog-multi-gpu-system.jpg\",\"contentUrl\":\"https:\\\/\\\/thegioimaychu.vn\\\/blog\\\/wp-content\\\/uploads\\\/2019\\\/06\\\/TheGioiMayChu-Blog-multi-gpu-system.jpg\",\"width\":1150,\"height\":746},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\\\/\\\/thegioimaychu.vn\\\/blog\\\/tong-hop\\\/kinh-nghiem-lua-chon-gpu-card-chuyen-dung-cho-ai-deep-learning-p1825\\\/#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"Trang ch\u1ee7\",\"item\":\"https:\\\/\\\/thegioimaychu.vn\\\/blog\\\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"T\u1ed5ng h\u1ee3p\",\"item\":\"https:\\\/\\\/thegioimaychu.vn\\\/blog\\\/.\\\/tong-hop\\\/\"},{\"@type\":\"ListItem\",\"position\":3,\"name\":\"Kinh nghi\u1ec7m l\u1ef1a ch\u1ecdn GPU card chuy\u00ean d\u1ee5ng cho AI\\\/Deep Learning\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\\\/\\\/thegioimaychu.vn\\\/blog\\\/#website\",\"url\":\"https:\\\/\\\/thegioimaychu.vn\\\/blog\\\/\",\"name\":\"Blog | TheGioiMayChu\",\"description\":\"C\u1eadp nh\u1eadt xu h\u01b0\u1edbng c\u00f4ng ngh\u1ec7 m\u00e1y ch\u1ee7\",\"publisher\":{\"@id\":\"https:\\\/\\\/thegioimaychu.vn\\\/blog\\\/#organization\"},\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\\\/\\\/thegioimaychu.vn\\\/blog\\\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"en-US\"},{\"@type\":\"Organization\",\"@id\":\"https:\\\/\\\/thegioimaychu.vn\\\/blog\\\/#organization\",\"name\":\"Blog | TheGioiMayChu\",\"url\":\"https:\\\/\\\/thegioimaychu.vn\\\/blog\\\/\",\"logo\":{\"@type\":\"ImageObject\",\"inLanguage\":\"en-US\",\"@id\":\"https:\\\/\\\/thegioimaychu.vn\\\/blog\\\/#\\\/schema\\\/logo\\\/image\\\/\",\"url\":\"https:\\\/\\\/thegioimaychu.vn\\\/blog\\\/wp-content\\\/uploads\\\/2018\\\/08\\\/cropped-ntcnlog-logo-2.png\",\"contentUrl\":\"https:\\\/\\\/thegioimaychu.vn\\\/blog\\\/wp-content\\\/uploads\\\/2018\\\/08\\\/cropped-ntcnlog-logo-2.png\",\"width\":500,\"height\":73,\"caption\":\"Blog | TheGioiMayChu\"},\"image\":{\"@id\":\"https:\\\/\\\/thegioimaychu.vn\\\/blog\\\/#\\\/schema\\\/logo\\\/image\\\/\"},\"sameAs\":[\"https:\\\/\\\/www.facebook.com\\\/thegioimaychu\\\/\"]},{\"@type\":\"Person\",\"@id\":\"https:\\\/\\\/thegioimaychu.vn\\\/blog\\\/#\\\/schema\\\/person\\\/feccec179330cfebfaa805639a357d18\",\"name\":\"Vanito Hoang\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"en-US\",\"@id\":\"https:\\\/\\\/secure.gravatar.com\\\/avatar\\\/e8187f37f72d74f9380a1367c8c530debee44f226a3301f0765c68f7cc1f3fa4?s=96&d=mm&r=g\",\"url\":\"https:\\\/\\\/secure.gravatar.com\\\/avatar\\\/e8187f37f72d74f9380a1367c8c530debee44f226a3301f0765c68f7cc1f3fa4?s=96&d=mm&r=g\",\"contentUrl\":\"https:\\\/\\\/secure.gravatar.com\\\/avatar\\\/e8187f37f72d74f9380a1367c8c530debee44f226a3301f0765c68f7cc1f3fa4?s=96&d=mm&r=g\",\"caption\":\"Vanito Hoang\"},\"url\":\"https:\\\/\\\/thegioimaychu.vn\\\/blog\\\/author\\\/vanito\\\/\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"Kinh nghi\u1ec7m l\u1ef1a ch\u1ecdn GPU card chuy\u00ean d\u1ee5ng cho AI\/Deep Learning - Blog | TheGioiMayChu","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/thegioimaychu.vn\/blog\/tong-hop\/kinh-nghiem-lua-chon-gpu-card-chuyen-dung-cho-ai-deep-learning-p1825\/","twitter_card":"summary_large_image","twitter_title":"Kinh nghi\u1ec7m l\u1ef1a ch\u1ecdn GPU card chuy\u00ean d\u1ee5ng cho AI\/Deep Learning - Blog | TheGioiMayChu","twitter_description":"Deep Learning l&agrave; m&#7897;t l&#297;nh v&#7921;c c&oacute; c&aacute;c y&ecirc;u c&#7847;u x&#7917; l&yacute; &#273;&#7863;c bi&#7879;t v&agrave; vi&#7879;c l&#7921;a ch&#7885;n GPU c&#7911;a b&#7841;n s&#7869; quy&#7871;t &#273;&#7883;nh r&#7845;t l&#7899;n &#273;&#7871;n qu&aacute; tr&igrave;nh tri&#7875;n khai.&nbsp;V&#7853;y nh&#7919;ng y&#7871;u t&#7889; n&agrave;o l&agrave; quan tr&#7885;ng khi b&#7841;n quy&#7871;t &#273;&#7883;nh mua GPU m&#7899;i cho m&#7909;c &#273;&iacute;ch Deep Learning?&nbsp;RAM, GPU, Core, Tensor Core?&nbsp;L&#7921;a ch&#7885;n...","twitter_image":"https:\/\/thegioimaychu.vn\/blog\/wp-content\/uploads\/2019\/06\/TheGioiMayChu-Blog-multi-gpu-system.jpg","twitter_misc":{"Written by":"Vanito Hoang","Est. reading time":"38 minutes"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"Article","@id":"https:\/\/thegioimaychu.vn\/blog\/tong-hop\/kinh-nghiem-lua-chon-gpu-card-chuyen-dung-cho-ai-deep-learning-p1825\/#article","isPartOf":{"@id":"https:\/\/thegioimaychu.vn\/blog\/tong-hop\/kinh-nghiem-lua-chon-gpu-card-chuyen-dung-cho-ai-deep-learning-p1825\/"},"author":{"name":"Vanito Hoang","@id":"https:\/\/thegioimaychu.vn\/blog\/#\/schema\/person\/feccec179330cfebfaa805639a357d18"},"headline":"Kinh nghi\u1ec7m l\u1ef1a ch\u1ecdn GPU card chuy\u00ean d\u1ee5ng cho AI\/Deep Learning","datePublished":"2019-06-13T02:40:36+00:00","dateModified":"2024-08-16T03:01:48+00:00","mainEntityOfPage":{"@id":"https:\/\/thegioimaychu.vn\/blog\/tong-hop\/kinh-nghiem-lua-chon-gpu-card-chuyen-dung-cho-ai-deep-learning-p1825\/"},"wordCount":7609,"publisher":{"@id":"https:\/\/thegioimaychu.vn\/blog\/#organization"},"image":{"@id":"https:\/\/thegioimaychu.vn\/blog\/tong-hop\/kinh-nghiem-lua-chon-gpu-card-chuyen-dung-cho-ai-deep-learning-p1825\/#primaryimage"},"thumbnailUrl":"https:\/\/thegioimaychu.vn\/blog\/wp-content\/uploads\/2019\/06\/TheGioiMayChu-Blog-multi-gpu-system.jpg","keywords":["ai","deep learning","gpu","NVIDIA"],"articleSection":["AI - HPC","T\u1ed5ng h\u1ee3p"],"inLanguage":"en-US"},{"@type":"WebPage","@id":"https:\/\/thegioimaychu.vn\/blog\/tong-hop\/kinh-nghiem-lua-chon-gpu-card-chuyen-dung-cho-ai-deep-learning-p1825\/","url":"https:\/\/thegioimaychu.vn\/blog\/tong-hop\/kinh-nghiem-lua-chon-gpu-card-chuyen-dung-cho-ai-deep-learning-p1825\/","name":"Kinh nghi\u1ec7m l\u1ef1a ch\u1ecdn GPU card chuy\u00ean d\u1ee5ng cho AI\/Deep Learning - Blog | TheGioiMayChu","isPartOf":{"@id":"https:\/\/thegioimaychu.vn\/blog\/#website"},"primaryImageOfPage":{"@id":"https:\/\/thegioimaychu.vn\/blog\/tong-hop\/kinh-nghiem-lua-chon-gpu-card-chuyen-dung-cho-ai-deep-learning-p1825\/#primaryimage"},"image":{"@id":"https:\/\/thegioimaychu.vn\/blog\/tong-hop\/kinh-nghiem-lua-chon-gpu-card-chuyen-dung-cho-ai-deep-learning-p1825\/#primaryimage"},"thumbnailUrl":"https:\/\/thegioimaychu.vn\/blog\/wp-content\/uploads\/2019\/06\/TheGioiMayChu-Blog-multi-gpu-system.jpg","datePublished":"2019-06-13T02:40:36+00:00","dateModified":"2024-08-16T03:01:48+00:00","breadcrumb":{"@id":"https:\/\/thegioimaychu.vn\/blog\/tong-hop\/kinh-nghiem-lua-chon-gpu-card-chuyen-dung-cho-ai-deep-learning-p1825\/#breadcrumb"},"inLanguage":"en-US","potentialAction":[{"@type":"ReadAction","target":["https:\/\/thegioimaychu.vn\/blog\/tong-hop\/kinh-nghiem-lua-chon-gpu-card-chuyen-dung-cho-ai-deep-learning-p1825\/"]}]},{"@type":"ImageObject","inLanguage":"en-US","@id":"https:\/\/thegioimaychu.vn\/blog\/tong-hop\/kinh-nghiem-lua-chon-gpu-card-chuyen-dung-cho-ai-deep-learning-p1825\/#primaryimage","url":"https:\/\/thegioimaychu.vn\/blog\/wp-content\/uploads\/2019\/06\/TheGioiMayChu-Blog-multi-gpu-system.jpg","contentUrl":"https:\/\/thegioimaychu.vn\/blog\/wp-content\/uploads\/2019\/06\/TheGioiMayChu-Blog-multi-gpu-system.jpg","width":1150,"height":746},{"@type":"BreadcrumbList","@id":"https:\/\/thegioimaychu.vn\/blog\/tong-hop\/kinh-nghiem-lua-chon-gpu-card-chuyen-dung-cho-ai-deep-learning-p1825\/#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Trang ch\u1ee7","item":"https:\/\/thegioimaychu.vn\/blog\/"},{"@type":"ListItem","position":2,"name":"T\u1ed5ng h\u1ee3p","item":"https:\/\/thegioimaychu.vn\/blog\/.\/tong-hop\/"},{"@type":"ListItem","position":3,"name":"Kinh nghi\u1ec7m l\u1ef1a ch\u1ecdn GPU card chuy\u00ean d\u1ee5ng cho AI\/Deep Learning"}]},{"@type":"WebSite","@id":"https:\/\/thegioimaychu.vn\/blog\/#website","url":"https:\/\/thegioimaychu.vn\/blog\/","name":"Blog | TheGioiMayChu","description":"C\u1eadp nh\u1eadt xu h\u01b0\u1edbng c\u00f4ng ngh\u1ec7 m\u00e1y ch\u1ee7","publisher":{"@id":"https:\/\/thegioimaychu.vn\/blog\/#organization"},"potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/thegioimaychu.vn\/blog\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"en-US"},{"@type":"Organization","@id":"https:\/\/thegioimaychu.vn\/blog\/#organization","name":"Blog | TheGioiMayChu","url":"https:\/\/thegioimaychu.vn\/blog\/","logo":{"@type":"ImageObject","inLanguage":"en-US","@id":"https:\/\/thegioimaychu.vn\/blog\/#\/schema\/logo\/image\/","url":"https:\/\/thegioimaychu.vn\/blog\/wp-content\/uploads\/2018\/08\/cropped-ntcnlog-logo-2.png","contentUrl":"https:\/\/thegioimaychu.vn\/blog\/wp-content\/uploads\/2018\/08\/cropped-ntcnlog-logo-2.png","width":500,"height":73,"caption":"Blog | TheGioiMayChu"},"image":{"@id":"https:\/\/thegioimaychu.vn\/blog\/#\/schema\/logo\/image\/"},"sameAs":["https:\/\/www.facebook.com\/thegioimaychu\/"]},{"@type":"Person","@id":"https:\/\/thegioimaychu.vn\/blog\/#\/schema\/person\/feccec179330cfebfaa805639a357d18","name":"Vanito Hoang","image":{"@type":"ImageObject","inLanguage":"en-US","@id":"https:\/\/secure.gravatar.com\/avatar\/e8187f37f72d74f9380a1367c8c530debee44f226a3301f0765c68f7cc1f3fa4?s=96&d=mm&r=g","url":"https:\/\/secure.gravatar.com\/avatar\/e8187f37f72d74f9380a1367c8c530debee44f226a3301f0765c68f7cc1f3fa4?s=96&d=mm&r=g","contentUrl":"https:\/\/secure.gravatar.com\/avatar\/e8187f37f72d74f9380a1367c8c530debee44f226a3301f0765c68f7cc1f3fa4?s=96&d=mm&r=g","caption":"Vanito Hoang"},"url":"https:\/\/thegioimaychu.vn\/blog\/author\/vanito\/"}]}},"_links":{"self":[{"href":"https:\/\/thegioimaychu.vn\/blog\/wp-json\/wp\/v2\/posts\/1825","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/thegioimaychu.vn\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/thegioimaychu.vn\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/thegioimaychu.vn\/blog\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/thegioimaychu.vn\/blog\/wp-json\/wp\/v2\/comments?post=1825"}],"version-history":[{"count":3,"href":"https:\/\/thegioimaychu.vn\/blog\/wp-json\/wp\/v2\/posts\/1825\/revisions"}],"predecessor-version":[{"id":20283,"href":"https:\/\/thegioimaychu.vn\/blog\/wp-json\/wp\/v2\/posts\/1825\/revisions\/20283"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/thegioimaychu.vn\/blog\/wp-json\/wp\/v2\/media\/20282"}],"wp:attachment":[{"href":"https:\/\/thegioimaychu.vn\/blog\/wp-json\/wp\/v2\/media?parent=1825"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/thegioimaychu.vn\/blog\/wp-json\/wp\/v2\/categories?post=1825"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/thegioimaychu.vn\/blog\/wp-json\/wp\/v2\/tags?post=1825"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}