Tr\u00ed tu\u1ec7 nh\u00e2n t\u1ea1o \u0111ang b\u01b0\u1edbc v\u00e0o m\u1ed9t giai \u0111o\u1ea1n ph\u00e1t tri\u1ec3n m\u1edbi, trong \u0111\u00f3 s\u1ee9c m\u1ea1nh t\u00ednh to\u00e1n \u0111\u00e3 tr\u1edf th\u00e0nh \u0111\u1ed9ng l\u1ef1c quy\u1ebft \u0111\u1ecbnh. D\u1eef li\u1ec7u v\u00e0 thu\u1eadt to\u00e1n cung c\u1ea5p nguy\u00ean li\u1ec7u th\u00f4 v\u00e0 c\u00e1c quy t\u1eafc, nh\u01b0ng chip m\u1edbi ch\u00ednh l\u00e0 \u0111\u1ed9ng c\u01a1 bi\u1ebfn nh\u1eefng ti\u1ec1m n\u0103ng \u0111\u00f3 th\u00e0nh hi\u1ec7n th\u1ef1c. Hi\u1ec7u n\u0103ng c\u1ee7a chip kh\u00f4ng ch\u1ec9 quy\u1ebft \u0111\u1ecbnh m\u1ee9c \u0111\u1ed9 ph\u00e1t tri\u1ec3n c\u1ee7a tr\u00ed tu\u1ec7 nh\u00e2n t\u1ea1o, m\u00e0 c\u00f2n \u0111\u1ecbnh h\u00ecnh li\u1ec7u c\u00e1c \u1ee9ng d\u1ee5ng c\u1ee7a n\u00f3 c\u00f3 th\u1ef1c s\u1ef1 c\u00f3 th\u1ec3 thay \u0111\u1ed5i di\u1ec7n m\u1ea1o c\u1ee7a c\u00e1c ng\u00e0nh c\u00f4ng nghi\u1ec7p v\u00e0 cu\u1ed9c s\u1ed1ng h\u00e0ng ng\u00e0y hay kh\u00f4ng.<\/p>\n\n\n\n
L\u1ea5y GPT-4 l\u00e0m v\u00ed d\u1ee5: qu\u00e1 tr\u00ecnh hu\u1ea5n luy\u1ec7n m\u00f4 h\u00ecnh n\u00e0y \u0111\u00e3 ti\u00eau th\u1ee5 kho\u1ea3ng 3.640 MWh \u0111i\u1ec7n \u2014 t\u01b0\u01a1ng \u0111\u01b0\u01a1ng v\u1edbi m\u1ee9c ti\u00eau th\u1ee5 h\u00e0ng n\u0103m c\u1ee7a 300 h\u1ed9 gia \u0111\u00ecnh \u2014 v\u00e0 ph\u1ee5 thu\u1ed9c v\u00e0o h\u00e0ng ch\u1ee5c ngh\u00ecn chip hi\u1ec7u su\u1ea5t cao ho\u1ea1t \u0111\u1ed9ng song song. Nhu c\u1ea7u n\u0103ng l\u01b0\u1ee3ng kh\u1ed5ng l\u1ed3 n\u00e0y nh\u1ea5n m\u1ea1nh m\u1ed1i li\u00ean h\u1ec7 kh\u00f4ng th\u1ec3 t\u00e1ch r\u1eddi gi\u1eefa AI v\u00e0 chip. Khi c\u00e1c thu\u1eadt to\u00e1n AI ng\u00e0y c\u00e0ng ph\u1ee9c t\u1ea1p, ki\u1ebfn tr\u00fac chip \u0111ang \u0111\u01b0\u1ee3c t\u1ed1i \u01b0u h\u00f3a cho c\u00e1c t\u00ednh n\u0103ng nh\u01b0 ph\u00e9p to\u00e1n ma tr\u1eadn v\u00e0 t\u00ednh to\u00e1n song song, t\u1ea1o th\u00e0nh m\u1ed9t l\u1ed9 tr\u00ecnh k\u1ef9 thu\u1eadt bao qu\u00e1t c\u1ea3 qu\u00e1 tr\u00ecnh hu\u1ea5n luy\u1ec7n v\u00e0 suy lu\u1eadn.<\/p>\n\n\n\n
Trong qu\u00e1 tr\u00ecnh hu\u1ea5n luy\u1ec7n, GPU v\u1eabn l\u00e0 x\u01b0\u01a1ng s\u1ed1ng c\u1ee7a h\u1ecdc s\u00e2u. C\u00e1c d\u00f2ng A100\/H100 c\u1ee7a NVIDIA, v\u1edbi h\u00e0ng ngh\u00ecn nh\u00e2n CUDA, cho ph\u00e9p hu\u1ea5n luy\u1ec7n m\u00f4 h\u00ecnh quy m\u00f4 kh\u1ed5ng l\u1ed3. TPU c\u1ee7a Google, \u0111\u01b0\u1ee3c t\u1ed1i \u01b0u h\u00f3a cho c\u00e1c ph\u00e9p to\u00e1n tensor, \u0111\u1ea1t c\u00f4ng su\u1ea5t l\u00ean t\u1edbi 100 PetaFLOPS \u1edf th\u1ebf h\u1ec7 th\u1ee9 t\u01b0, h\u1ed7 tr\u1ee3 c\u00e1c d\u1ecbch v\u1ee5 AI to\u00e0n c\u1ea7u tr\u00ean Google Cloud. Trong khi \u0111\u00f3, chip ASIC Cambricon c\u1ee7a Trung Qu\u1ed1c th\u1ec3 hi\u1ec7n m\u1ee9c hi\u1ec7u qu\u1ea3 cao h\u01a1n GPU g\u1ea5p h\u01a1n m\u01b0\u1eddi l\u1ea7n trong c\u00e1c thu\u1eadt to\u00e1n c\u1ee5 th\u1ec3, \u0111\u00e1nh d\u1ea5u m\u1ed9t b\u01b0\u1edbc \u0111\u1ed9t ph\u00e1 m\u1edbi trong l\u0129nh v\u1ef1c chip AI t\u00f9y ch\u1ec9nh.<\/p>\n\n\n\n
Ng\u01b0\u1ee3c l\u1ea1i, c\u00f4ng ngh\u1ec7 suy lu\u1eadn l\u1ea1i ch\u00fa tr\u1ecdng v\u00e0o \u0111\u1ed9 tr\u1ec5 th\u1ea5p v\u00e0 tri\u1ec3n khai t\u1ea1i thi\u1ebft b\u1ecb \u0111\u1ea7u cu\u1ed1i. C\u00e1c n\u1ec1n t\u1ea3ng FPGA nh\u01b0 Xilinx Versal mang l\u1ea1i th\u1eddi gian ph\u1ea3n h\u1ed3i \u1edf m\u1ee9c mili gi\u00e2y trong l\u0129nh v\u1ef1c l\u00e1i xe t\u1ef1 \u0111\u1ed9ng. Ki\u1ebfn tr\u00fac A17 Pro c\u1ee7a Apple v\u00e0 Kirin Da Vinci c\u1ee7a Huawei t\u00edch h\u1ee3p tr\u1ef1c ti\u1ebfp c\u00e1c \u0111\u01a1n v\u1ecb x\u1eed l\u00fd th\u1ea7n kinh (NPU) v\u00e0o thi\u1ebft b\u1ecb di \u0111\u1ed9ng, m\u1edf ra nh\u1eefng b\u01b0\u1edbc \u0111\u1ed9t ph\u00e1 trong c\u00e1c \u1ee9ng d\u1ee5ng x\u1eed l\u00fd h\u00ecnh \u1ea3nh th\u1eddi gian th\u1ef1c v\u00e0 th\u1ef1c t\u1ebf t\u0103ng c\u01b0\u1eddng (AR). Nh\u1eefng \u0111\u1ed5i m\u1edbi n\u00e0y \u0111ang \u0111\u01b0a tr\u00ed tu\u1ec7 nh\u00e2n t\u1ea1o (AI) v\u00e0o cu\u1ed9c s\u1ed1ng h\u00e0ng ng\u00e0y thay v\u00ec ch\u1ec9 gi\u1edbi h\u1ea1n n\u00f3 trong c\u00e1c trung t\u00e2m d\u1eef li\u1ec7u.<\/p>\n\n\n\n
Trong m\u1ecdi l\u0129nh v\u1ef1c, chip \u0111ang \u0111\u1ecbnh h\u00ecnh l\u1ea1i b\u1ee9c tranh t\u1ed5ng th\u1ec3. Trong l\u0129nh v\u1ef1c \u0111i\u1ec7n to\u00e1n \u0111\u00e1m m\u00e2y, c\u00e1c phi\u00ean b\u1ea3n P4d c\u1ee7a AWS \u0111\u01b0\u1ee3c trang b\u1ecb 8 GPU A100 cung c\u1ea5p h\u1ea1 t\u1ea7ng cho vi\u1ec7c \u0111\u00e0o t\u1ea1o m\u00f4 h\u00ecnh quy m\u00f4 doanh nghi\u1ec7p, trong khi chip Hanguang 800 c\u1ee7a Alibaba mang l\u1ea1i hi\u1ec7u su\u1ea5t g\u1ea5p b\u1ed1n l\u1ea7n so v\u1edbi GPU truy\u1ec1n th\u1ed1ng trong x\u1eed l\u00fd h\u00ecnh \u1ea3nh. Trong l\u0129nh v\u1ef1c l\u00e1i xe t\u1ef1 \u0111\u1ed9ng, NVIDIA Orin cung c\u1ea5p 254 TOPS \u0111\u1ec3 h\u1ed7 tr\u1ee3 nh\u1eadn th\u1ee9c v\u00e0 ra quy\u1ebft \u0111\u1ecbnh \u1edf c\u1ea5p \u0111\u1ed9 L4, trong khi Journey 6 c\u1ee7a Horizon Robotics c\u00e2n b\u1eb1ng gi\u1eefa hi\u1ec7u su\u1ea5t n\u0103ng l\u01b0\u1ee3ng v\u00e0 \u0111\u1ed9 tin c\u1eady. Trong l\u0129nh v\u1ef1c h\u00ecnh \u1ea3nh y t\u1ebf, c\u00e1c chip AI \u0111\u01b0\u1ee3c nh\u00fang v\u00e0o thi\u1ebft b\u1ecb ch\u1ea9n \u0111o\u00e1n \u0111\u00e3 \u0111\u1ea1t \u0111\u1ed9 ch\u00ednh x\u00e1c ph\u00e1t hi\u1ec7n kh\u1ed1i u 97% \u0111\u1ed3ng th\u1eddi b\u1ea3o v\u1ec7 quy\u1ec1n ri\u00eang t\u01b0 c\u1ee7a b\u1ec7nh nh\u00e2n th\u00f4ng qua m\u00e3 h\u00f3a c\u1ea5p \u0111\u1ed9 ph\u1ea7n c\u1ee9ng. Trong l\u0129nh v\u1ef1c \u0111i\u1ec7n t\u1eed ti\u00eau d\u00f9ng, Snapdragon 8 Gen3 cung c\u1ea5p d\u1ecbch thu\u1eadt th\u1eddi gian th\u1ef1c v\u00e0 t\u01b0\u01a1ng t\u00e1c gi\u1ecdng n\u00f3i th\u00f4ng minh \u0111\u1ed3ng th\u1eddi gi\u1ea3m ti\u00eau th\u1ee5 \u0111i\u1ec7n n\u0103ng 30% so v\u1edbi th\u1ebf h\u1ec7 tr\u01b0\u1edbc.<\/p>\n\n\n\n
Trong khi \u0111\u00f3, nh\u1eefng b\u01b0\u1edbc \u0111\u1ed9t ph\u00e1 \u0111ang m\u1edf r\u1ed9ng ranh gi\u1edbi c\u1ee7a l\u0129nh v\u1ef1c t\u00ednh to\u00e1n tr\u00ed tu\u1ec7 nh\u00e2n t\u1ea1o (AI). C\u00e1c ki\u1ebfn tr\u00fac d\u1ecb th\u1ec3 hi\u1ec7n nay cho ph\u00e9p CPU, GPU, TPU v\u00e0 DSP h\u1ee3p t\u00e1c th\u00f4ng qua c\u00e1c k\u1ebft n\u1ed1i t\u1ed1c \u0111\u1ed9 cao. C\u00e1c c\u00f4ng ngh\u1ec7 x\u1eed l\u00fd trong b\u1ed9 nh\u1edb (PIM) \u0111ang gi\u1ea3i quy\u1ebft v\u1ea5n \u0111\u1ec1 \u201cb\u1ee9c t\u01b0\u1eddng b\u1ed9 nh\u1edb\u201d t\u1ed3n t\u1ea1i t\u1eeb l\u00e2u, gi\u00fap gi\u1ea3m m\u1ee9c ti\u00eau th\u1ee5 n\u0103ng l\u01b0\u1ee3ng cho vi\u1ec7c suy lu\u1eadn xu\u1ed1ng t\u1edbi 90%. C\u00e1c chip l\u01b0\u1ee3ng t\u1eed, d\u00f9 v\u1eabn \u0111ang trong giai \u0111o\u1ea1n th\u1eed nghi\u1ec7m, cho th\u1ea5y ti\u1ec1m n\u0103ng \u0111\u1ed9t ph\u00e1 trong c\u00e1c t\u00e1c v\u1ee5 t\u1ed1i \u01b0u h\u00f3a v\u00e0 m\u00f4 ph\u1ecfng, h\u1ee9a h\u1eb9n m\u1ed9t s\u1ef1 chuy\u1ec3n \u0111\u1ed5i m\u00f4 h\u00ecnh t\u00ednh to\u00e1n trong t\u01b0\u01a1ng lai.<\/p>\n\n\n\n
Tuy nhi\u00ean, nh\u1eefng th\u00e1ch th\u1ee9c v\u1eabn c\u00f2n \u0111\u00f3. Chi ph\u00ed n\u0103ng l\u01b0\u1ee3ng \u0111\u1ec3 \u0111\u1ea1t \u0111\u01b0\u1ee3c hi\u1ec7u n\u0103ng \u0111ang t\u0103ng v\u1ecdt \u2014 GPU H100 c\u1ee7a NVIDIA ti\u00eau th\u1ee5 t\u1edbi 400W \u2014 bu\u1ed9c c\u00e1c trung t\u00e2m d\u1eef li\u1ec7u ph\u1ea3i \u00e1p d\u1ee5ng c\u00e1c k\u1ef9 thu\u1eadt l\u00e0m m\u00e1t ti\u00ean ti\u1ebfn nh\u01b0 ng\u00e2m ch\u1ea5t l\u1ecfng. C\u00e1c chip n\u1ed9i \u0111\u1ecba c\u1ee7a Trung Qu\u1ed1c, nh\u01b0 Ascend 910B c\u1ee7a Huawei v\u00e0 Kunlun 2 c\u1ee7a Baidu, \u0111\u00e3 \u0111\u1ea1t \u0111\u01b0\u1ee3c nh\u1eefng \u0111\u1ed9t ph\u00e1 \u201ct\u1eeb kh\u00f4ng \u0111\u1ebfn m\u1ed9t\u201d, nh\u01b0ng v\u1eabn ph\u1ea3i \u0111\u1ed1i m\u1eb7t v\u1edbi nh\u1eefng r\u00e0o c\u1ea3n v\u1ec1 kh\u1ea3 n\u0103ng t\u01b0\u01a1ng th\u00edch h\u1ec7 sinh th\u00e1i v\u00e0 tri\u1ec3n khai quy m\u00f4 l\u1edbn. T\u01b0\u01a1ng lai c\u00f3 th\u1ec3 n\u1eb1m \u1edf s\u1ef1 h\u1ee3p t\u00e1c gi\u1eefa \u0111\u00e1m m\u00e2y v\u00e0 thi\u1ebft b\u1ecb \u0111\u1ea7u cu\u1ed1i: m\u00e1y ch\u1ee7 \u0111\u00e1m m\u00e2y x\u1eed l\u00fd vi\u1ec7c hu\u1ea5n luy\u1ec7n quy m\u00f4 l\u1edbn, trong khi c\u00e1c thi\u1ebft b\u1ecb \u0111\u1ea7u cu\u1ed1i th\u1ef1c hi\u1ec7n suy lu\u1eadn nh\u1eb9 \u2014 c\u1ee5m hu\u1ea5n luy\u1ec7n tr\u00ean bo m\u1ea1ch d\u1ef1a tr\u00ean chip D1 c\u1ee7a Tesla l\u00e0 v\u00ed d\u1ee5 \u0111i\u1ec3n h\u00ecnh cho s\u1ef1 chuy\u1ec3n \u0111\u1ed5i n\u00e0y.<\/p>\n\n\n\n
T\u1eeb c\u00e1c GPU \u0111\u00e3 kh\u01a1i m\u00e0o cu\u1ed9c c\u00e1ch m\u1ea1ng t\u00ednh to\u00e1n AI cho \u0111\u1ebfn c\u00e1c ASIC v\u00e0 NPU chuy\u00ean d\u1ee5ng \u0111ang \u0111\u1ecbnh h\u00ecnh c\u00e1c ti\u00eau chu\u1ea9n m\u1edbi c\u1ee7a ng\u00e0nh, m\u1ed1i quan h\u1ec7 gi\u1eefa chip v\u00e0 AI \u0111\u00e3 ph\u00e1t tri\u1ec3n v\u01b0\u1ee3t ra kh\u1ecfi ph\u1ea1m vi \u201cc\u00f4ng c\u1ee5 v\u00e0 \u1ee9ng d\u1ee5ng\u201d \u0111\u1ec3 tr\u1edf th\u00e0nh m\u1ed9t h\u1ec7 sinh th\u00e1i c\u1ed9ng sinh. Khi m\u1ed9t NPU c\u1ee1 m\u00f3ng tay c\u00f3 th\u1ec3 ph\u00e2n lo\u1ea1i h\u00e0ng tri\u1ec7u h\u00ecnh \u1ea3nh trong th\u1eddi gian th\u1ef1c, v\u00e0 khi m\u1ed9t c\u1ee5m si\u00eau m\u00e1y t\u00ednh c\u00f3 th\u1ec3 m\u00f4 ph\u1ecfng qu\u00e1 tr\u00ecnh ti\u1ebfn h\u00f3a c\u1ee7a c\u00e1c thi\u00ean h\u00e0, chip kh\u00f4ng ch\u1ec9 l\u00e0 b\u1ed9 x\u01b0\u01a1ng c\u1ee7a AI \u2014 ch\u00fang \u0111ang m\u1edf r\u1ed9ng ranh gi\u1edbi c\u1ee7a ch\u00ednh nh\u1eadn th\u1ee9c con ng\u01b0\u1eddi. C\u00e2u h\u1ecfi c\u1ed1t l\u00f5i c\u1ee7a cu\u1ed9c \u0111ua t\u00ednh to\u00e1n n\u00e0y kh\u00f4ng ch\u1ec9 n\u1eb1m \u1edf s\u1ee9c m\u1ea1nh th\u00f4 s\u01a1, m\u00e0 c\u00f2n \u1edf vi\u1ec7c gi\u00fap AI ti\u1ebfn h\u00f3a t\u1eeb m\u1ed9t \u201cg\u00e3 kh\u1ed5ng l\u1ed3 ng\u1ed1n t\u00e0i nguy\u00ean\u201d th\u00e0nh m\u1ed9t \u201ctr\u00ed tu\u1ec7 ti\u1ebft ki\u1ec7m n\u0103ng l\u01b0\u1ee3ng\u201d. V\u00e0 chip ch\u00ednh l\u00e0 ch\u00eca kh\u00f3a \u0111\u1ec3 m\u1edf ra gi\u1ea3i ph\u00e1p t\u01b0\u01a1ng lai n\u00e0y.<\/p>","protected":false},"featured_media":939,"template":"","meta":{"_acf_changed":false,"footnotes":""},"categories":[28,18],"tags":[],"class_list":["post-938","application","type-application","status-publish","has-post-thumbnail","hentry","category-artificial-intelligence-high-performance-computing-hpc","category-application-industries"],"acf":[],"_links":{"self":[{"href":"https:\/\/www.wallisemi.com\/vi\/wp-json\/wp\/v2\/application\/938","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.wallisemi.com\/vi\/wp-json\/wp\/v2\/application"}],"about":[{"href":"https:\/\/www.wallisemi.com\/vi\/wp-json\/wp\/v2\/types\/application"}],"version-history":[{"count":1,"href":"https:\/\/www.wallisemi.com\/vi\/wp-json\/wp\/v2\/application\/938\/revisions"}],"predecessor-version":[{"id":940,"href":"https:\/\/www.wallisemi.com\/vi\/wp-json\/wp\/v2\/application\/938\/revisions\/940"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.wallisemi.com\/vi\/wp-json\/wp\/v2\/media\/939"}],"wp:attachment":[{"href":"https:\/\/www.wallisemi.com\/vi\/wp-json\/wp\/v2\/media?parent=938"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wallisemi.com\/vi\/wp-json\/wp\/v2\/categories?post=938"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wallisemi.com\/vi\/wp-json\/wp\/v2\/tags?post=938"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}