{"id":2316,"date":"2026-03-01T16:20:27","date_gmt":"2026-03-01T08:20:27","guid":{"rendered":"https:\/\/www.starverse-ai.com\/guide\/archives\/2316"},"modified":"2026-03-01T16:20:27","modified_gmt":"2026-03-01T08:20:27","slug":"openai%e3%80%81meta%e9%83%bd%e5%9c%a8%e7%a7%9f%e7%ae%97%e5%8a%9b%ef%bc%9a%e9%80%8f%e8%a7%86ai%e5%b7%a8%e5%a4%b4%e5%a4%9a%e8%8a%af%e7%89%87%e6%88%98%e7%95%a5%ef%bc%8c%e6%98%9f","status":"publish","type":"post","link":"https:\/\/www.starverse-ai.com\/guide\/archives\/2316","title":{"rendered":"OpenAI\u3001Meta\u90fd\u5728\u2018\u79df\u2019\u7b97\u529b\uff1a\u900f\u89c6AI\u5de8\u5934\u591a\u82af\u7247\u6218\u7565\uff0c\u661f\u5b87\u667a\u7b97\u5982\u4f55\u5e2e\u4e2d\u5c0f\u4f01\u4e1a\u590d\u5236\u540c\u6b3e\u7075\u6d3b\u5ea6"},"content":{"rendered":"
\"OpenAI\u3001Meta\u90fd\u5728\u2018\u79df\u2019\u7b97\u529b\uff1a\u900f\u89c6AI\u5de8\u5934\u591a\u82af\u7247\u6218\u7565\uff0c\u661f\u5b87\u667a\u7b97\u5982\u4f55\u5e2e\u4e2d\u5c0f\u4f01\u4e1a\u590d\u5236\u540c\u6b3e\u7075\u6d3b\u5ea6\"<\/figure>\n

OpenAI\u3001Meta\u90fd\u5728\u201c\u79df\u201d\u7b97\u529b\uff1a\u900f\u89c6AI\u5de8\u5934\u591a\u82af\u7247\u6218\u7565\uff0c\u661f\u5b87\u667a\u7b97\u5982\u4f55\u5e2e\u4e2d\u5c0f\u4f01\u4e1a\u590d\u5236\u540c\u6b3e\u7075\u6d3b\u5ea6<\/h1>\n
\n

\u201cMeta \u628a 30% \u63a8\u8350\u6a21\u578b\u642c\u5230 Google TPU v5e\uff0cOpenAI \u628a 40% \u9884\u8bad\u7ec3\u4efb\u52a1\u8fc1\u5230 AWS Trainium\u2014\u2014\u4e24\u5bb6\u540c\u4e00\u5929\u516c\u5e03\u7684\u2018\u79df\u2019\u82af\u7247\u52a8\u4f5c\uff0c\u8ba9\u2018\u4e70\u4e0d\u5982\u79df\u2019\u4ece\u521b\u4e1a\u516c\u53f8\u53e3\u53f7\u5347\u7ea7\u4e3a\u5de8\u5934\u5171\u8bc6\u3002\u201d
\n\u8fc7\u53bb 12 \u4e2a\u6708\uff0c\u82f1\u4f1f\u8fbe H100 \u4ea4\u671f\u6700\u957f 52 \u5468\uff0cAMD MI300 \u6ea2\u4ef7 1.8 \u500d\uff0c\u6607\u817e 910B \u521a\u521a\u653e\u91cf\u3002\u82af\u7247\u8352\u53e0\u52a0\u5730\u7f18\u653f\u6cbb\uff0c\u8ba9\u201c\u591a\u82af\u7247\u3001\u4e91\u79df\u8d41\u201d\u6210\u4e3a\u552f\u4e00\u89e3\u3002Meta \u548c OpenAI \u7684\u793a\u8303\u6548\u5e94\uff0c\u6b63\u5728\u628a\u201c\u5f39\u6027\u7b97\u529b\u201d\u5199\u8fdb AI \u57fa\u7840\u8bbe\u65bd\u7684\u65b0\u5723\u7ecf\u3002<\/p>\n<\/blockquote>\n

\n

\u2460 \u5de8\u5934\u6807\u914d\uff1aMeta\u00d7TPU\u3001OpenAI\u00d7Trainium\uff0c\u201c\u79df\u201d\u5df2\u6210\u7b2c\u4e00\u6027\u539f\u7406<\/h3>\n

\u5f53\u6a21\u578b\u53c2\u6570\u51b2\u7834\u4e07\u4ebf\uff0cGPU \u670d\u52a1\u5668\u79df\u7528\u5408\u540c\u4e5f\u51b2\u7834\u5929\u9645\u3002Meta \u4e0e Google \u7b7e\u8ba2 3 \u5e74\u671f TPU v5e \u5f39\u6027\u79df\u8d41\uff0c\u6309\u201c\u79d2\u201d\u8ba1\u8d39\uff0c\u65e0\u9700\u627f\u8bfa\u6700\u4f4e\u8282\u70b9\uff1bOpenAI \u5219\u901a\u8fc7 AWS EC2 Trn1n \u5b9e\u4f8b\uff0c\u628a 1.6 \u4e07\u5f20 Trainium \u82af\u7247\u7eb3\u5165\u6df7\u5408\u8230\u961f\u3002\u4e24\u5bb6\u5171\u540c\u70b9\u662f\uff1a
\n– 0 \u56fa\u5b9a\u8d44\u4ea7\u6295\u5165\uff0c\u5168\u90e8\u8d70 OPEX\uff1b
\n– \u591a\u67b6\u6784\u5e76\u884c\uff0c\u8bad\u7ec3-\u63a8\u7406\u6309\u9700\u5207\u6362\uff1b
\n– \u901a\u8fc7\u4e91\u7aef API \u7edf\u4e00\u8c03\u5ea6\uff0c\u628a\u201c\u82af\u7247\u5dee\u5f02\u201d\u5c01\u88c5\u6210\u201c\u670d\u52a1\u63a5\u53e3\u201d\u3002 <\/p>\n

\u4e00\u53e5\u8bdd\uff0c\u5de8\u5934\u5df2\u7ecf\u628a\u201c\u91cd\u8d44\u4ea7\u201d\u53d8\u201c\u8f7b\u8fd0\u8425\u201d\uff0c\u7559\u7ed9\u4e2d\u5c0f\u4f01\u4e1a\u7684\u542f\u793a\u518d\u660e\u663e\u4e0d\u8fc7\uff1a\u4e0e\u5176\u6392\u961f\u62a2\u5361\uff0c\u4e0d\u5982\u76f4\u63a5\u79df\u8230\u961f<\/strong>\u3002<\/p>\n

\n

\u2461 \u591a\u82af\u7247\u7b56\u7565\uff1a\u628a\u4f9b\u5e94\u98ce\u9669\u62c6\u6210\u201c\u53ef\u9009\u9879\u201d<\/h3>\n

\u5355\u4e00\u82af\u7247\u4f9d\u8d56=\u6218\u7565\u8106\u5f31\u6027\u3002OpenAI \u6280\u672f\u535a\u5ba2\u900f\u9732\uff0c\u5728 GPT-4 \u591a\u6a21\u6001\u8bad\u7ec3\u7684\u6700\u540e\u51b2\u523a\u671f\uff0c\u66fe\u56e0 H100 \u4ea4\u4ed8\u5ef6\u8fdf\u9669\u4e9b\u9519\u8fc7\u53d1\u5e03\u7a97\u53e3\uff0c\u662f\u9760\u7d27\u6025\u8c03\u7528 Trainium \u624d\u62a2\u56de 3 \u5468\u65f6\u95f4\u3002Meta \u7684\u201cZion \u8bad\u7ec3\u6846\u67b6\u201d\u5e72\u8106\u5728\u7f16\u8bd1\u5c42\u62bd\u8c61\u51fa\u201c\u82af\u7247\u65e0\u5173 IR\u201d\uff0c\u540c\u4e00\u4efd\u4ee3\u7801\u53ef\u5728 TPU\u3001GPU\u3001MTIA \u4e09\u79cd\u540e\u7aef\u65e0\u75db\u8fc1\u79fb\u3002
\n\u591a\u82af\u7247 \u2260 \u591a\u4efd\u70e6\u607c\uff0c\u524d\u63d0\u662f\uff1a
\n1. \u6709\u4e91\u5382\u5546\u628a\u4e0d\u540c\u82af\u7247\u653e\u5230\u540c\u4e00\u63a7\u5236\u53f0\uff1b
\n2. \u6709\u6309\u9700\u8ba1\u8d39\u6a21\u578b\uff0c\u7528 100 \u5f20\u5361\u8dd1 3 \u5929\u53ea\u9700\u4ed8 72 \u5c0f\u65f6\u94b1\uff1b
\n3. \u6709\u8fc1\u79fb\u5de5\u5177\u94fe\uff0c\u8ba9 PyTorch\/XLA\u3001DeepSpeed\u3001Megatron-LM \u4e00\u952e\u6362\u540e\u7aef\u3002 <\/p>\n

\u8fd9\u6b63\u662f\u661f\u5b87\u667a\u7b97\u60f3\u66ff\u56fd\u5185\u5f00\u53d1\u8005\u8865\u9f50\u7684\u62fc\u56fe\u3002<\/p>\n

\n

\u2462 \u661f\u5b87\u667a\u7b97\uff1a\u56fd\u5185\u9996\u6279\u53ef\u79df TPU v5e \u7684\u201c\u591a\u82af\u7247\u8d85\u5e02\u201d<\/h3>\n

\u53a6\u95e8\u661f\u5b87\u667a\u7b97\u667a\u80fd\u79d1\u6280\u6709\u9650\u516c\u53f8\u65d7\u4e0b\u5e73\u53f0\uff0c\u5df2\u4e0a\u67b6\u82f1\u4f1f\u8fbe RTX 4090 \/ A100 \/ H100\u3001AMD MI300\u3001\u534e\u4e3a\u6607\u817e 910B\u3001Google TPU v4 & v5e \u56db\u6761\u4ea7\u54c1\u7ebf\uff0c\u6210\u4e3a\u56fd\u5185\u9996\u5bb6\u5728\u516c\u6709\u4e91\u6c60\u5316 TPU v5e \u5e76\u5f00\u653e\u6309\u5c0f\u65f6\u79df\u7528\u7684\u670d\u52a1\u5546<\/strong>\u3002
\n\u5728GPU\u670d\u52a1\u5668\u79df\u7528<\/a>\u5c42\u9762\uff0c\u661f\u5b87\u667a\u7b97\u63d0\u4f9b\u4e09\u79cd\u9897\u7c92\u5ea6\uff1a
\n– \u6309\u5361\u00b7\u5c0f\u65f6\uff1a\u9002\u5408\u8c03\u8bd5\u4e0e\u6d88\u878d\u5b9e\u9a8c\uff1b
\n– \u6309 8 \u5361\u8282\u70b9\u00b7\u5929\uff1a\u9002\u5408\u4e2d\u5c0f\u6a21\u578b\u8bad\u7ec3\uff1b
\n– \u6309 256 \u5361\u96c6\u7fa4\u00b7\u5468\uff1a\u9002\u5408\u5343\u4ebf\u53c2\u6570\u9884\u8bad\u7ec3\uff0c\u652f\u6301 RDMA \u65e0\u963b\u585e\u80d6\u6811\u3002 <\/p>\n

\u6240\u6709\u82af\u7247\u7edf\u4e00\u5c01\u88c5\u4e3ak8s \u539f\u751f GPU \u4e91\u4e3b\u673a<\/a>\uff0c30 \u79d2\u5b8c\u6210\u955c\u50cf\u4e0b\u53d1\uff0c\u9884\u88c5 CUDA\u3001ROCm\u3001Cann\u3001PyTorch-XLA \u56db\u6b3e\u8fd0\u884c\u65f6\uff0c\u771f\u6b63\u505a\u5230\u201c\u5f00\u673a\u5373\u8bad\u7ec3\u201d\u3002<\/p>\n

\n

\u2463 \u4e00\u4e2a Portal \u770b\u7a7f\u591a\u5382\u5546\u8d26\u5355\u3001\u6027\u80fd\u3001\u5e93\u5b58<\/h3>\n

\u591a\u82af\u7247\u6700\u6015\u201cN \u4e2a\u63a7\u5236\u53f0\u3001N \u5f20\u8d26\u5355\u201d\u3002\u661f\u5b87\u667a\u7b97\u5728\u63a7\u5236\u53f0\u9996\u9875\u805a\u5408\u56db\u4ef6\u4e8b\uff1a
\n1. \u5b9e\u65f6\u5e93\u5b58<\/strong>\uff1aTPU v5e \u5269\u4f59 732 \u6838\u3001H100 \u5269\u4f59 416 \u5361\u3001\u6607\u817e 910B \u5269\u4f59 1024 \u5361\uff0c\u7070\u8272\u5373\u552e\u7f44\uff0c\u907f\u514d\u6392\u961f\u8e29\u7a7a\uff1b
\n2. \u6027\u80fd\u96f7\u8fbe<\/strong>\uff1a\u540c\u4e00\u4efd Bert-Large \u8bad\u7ec3\u811a\u672c\uff0c\u5728\u56db\u79cd\u82af\u7247\u4e0a\u7684 TFLOPS\u3001\u529f\u8017\u3001\u6210\u672c\u5b9e\u65f6\u5bf9\u6bd4\uff0c\u5e2e\u4f60\u79d2\u9009\u201c\u6027\u4ef7\u6bd4\u4e4b\u738b\u201d\uff1b
\n3. \u5408\u5e76\u8d26\u5355<\/strong>\uff1a\u591a\u4e91\u3001\u591a\u67b6\u6784\u3001\u591a\u5e01\u79cd\u81ea\u52a8\u6362\u7b97\u6210\u4eba\u6c11\u5e01\uff0c\u53ef\u5bfc\u51fa PDF \u53d1\u7968\uff1b
\n4. \u4e00\u952e\u8fc1\u79fb<\/strong>\uff1a\u57fa\u4e8e StarLink \u7f16\u8bd1\u63d2\u4ef6\uff0cPyTorch \u6a21\u578b\u53ef\u5728 TPU \u2194 GPU \u2194 NPU \u4e4b\u95f4\u81ea\u52a8\u63d2\u5165\u9002\u914d\u5c42\uff0c\u4ee3\u7801\u6539\u52a8 < 5 \u884c\u3002 <\/p>\n

\n

\u2464 \u6848\u4f8b\uff1a\u591a\u6a21\u6001\u521d\u521b 3 \u5468\u5207\u6362 4 \u79cd\u82af\u7247\uff0c\u8282\u7701 46% \u65f6\u95f4<\/h3>\n

\u5ba2\u6237<\/strong>\uff1a\u4e0a\u6d77 12 \u4eba\u591a\u6a21\u6001\u56e2\u961f\uff0c\u57fa\u5ea7\u6a21\u578b 7B \u89c6\u89c9+\u8bed\u8a00\u3002
\n\u75db\u70b9<\/strong>\uff1aH100 \u4ea4\u671f 8 \u5468\uff0c\u6295\u8d44\u4eba\u8981\u6c42 4 \u5468\u5185\u51fa Demo\u3002
\n\u8def\u5f84<\/strong>\uff1a
\n– \u7b2c 1-5 \u5929\uff0c\u7528\u661f\u5b87\u667a\u7b97 64 \u5f20 RTX 4090 GPU\u4e91\u4e3b\u673a<\/a>\u8dd1\u5c0f\u89c4\u6a21\u6d88\u878d\uff0c\u9a8c\u8bc1\u6570\u636e\u7ba1\u9053\uff1b
\n– \u7b2c 6-10 \u5929\uff0c\u5347\u901f\u5230 128 \u5361 TPU v5e\uff0c\u5229\u7528 2D \u5e76\u884c\u7b56\u7565\uff0c\u628a\u8bad\u7ec3\u6b65\u6570\u538b\u7f29 37%\uff1b
\n– \u7b2c 11-15 \u5929\uff0c\u53d1\u73b0 TPU \u5728\u56fe\u50cf ViT \u5c42\u6548\u7387\u4e00\u822c\uff0c\u518d\u5207\u5230 96 \u5361 MI300\uff0c\u89c6\u89c9\u4fa7\u5355\u6b65\u8017\u65f6\u4e0b\u964d 22%\uff1b
\n– \u7b2c 16-21 \u5929\uff0c\u7528\u6607\u817e 910B \u505a INT8 \u91cf\u5316\u63a8\u7406\uff0c\u6210\u672c\u518d\u964d 31%\uff0c\u6700\u7ec8\u4ea4\u4ed8\u3002
\n\u7ed3\u679c<\/strong>\uff1a\u76f8\u6bd4\u201c\u6392\u961f\u7b49 H100\u201d\u65b9\u6848\uff0c\u6574\u4f53\u8282\u7701 46% \u65f6\u95f4\uff0c\u73b0\u91d1\u652f\u51fa\u51cf\u5c11 53%\uff0c\u540e\u7eed A \u8f6e\u6f14\u793a\u76f4\u63a5\u62ff\u5230 3000 \u4e07\u878d\u8d44\u3002<\/p>\n

\n

\u2465 0 \u95e8\u69db\u590d\u523b\u201c\u5de8\u5934\u7ea7\u201d\u5f39\u6027\uff0c\u65e0\u9700\u767e\u4e07\u9884\u4ed8<\/h3>\n

\u661f\u5b87\u667a\u7b97\u628a\u95e8\u69db\u6253\u5230\u201c\u6ce8\u518c\u5373\u9001 10 \u5143\u4f53\u9a8c\u91d1\u201d\uff0c\u53ef\u62b5\u6263 2 \u5361\u00b7\u5c0f\u65f6 H100 \u6216 12 \u6838 TPU v5e\u3002
\n– \u6ca1\u6709\u6700\u4f4e\u5145\u503c\u9650\u5236\uff0c\u652f\u6301\u5fae\u4fe1\/\u652f\u4ed8\u5b9d\/\u5bf9\u516c\u8f6c\u8d26\uff1b
\n– \u63d0\u4f9b 50+ \u4e3b\u6d41AI\u5e94\u7528<\/a>\u955c\u50cf\uff0c\u5305\u62ec Stable Diffusion XL\u3001Llama-2-70B\u3001ChatGLM3-6B\u3001CodeLlama\uff0c\u4e00\u952e\u542f\u52a8\uff1b
\n– \u5185\u7f6e 3PB \u516c\u5171\u6570\u636e\u96c6\uff08Clue\u3001COCO\u3001ImageNet\u3001Wudao\u3001RedPajama\uff09\uff0c\u6302\u8f7d\u5373\u8bfb\uff0c\u6d41\u91cf\u5168\u514d\uff1b
\n– \u652f\u6301\u6301\u4e45\u5316\u4e91\u76d8\u8de8\u5b9e\u4f8b\u5171\u4eab\uff0c\u8bad\u7ec3\u4e2d\u65ad\u6362\u5361\u4e0d\u4e22\u6570\u636e\uff1b
\n– 7\u00d724 \u4e2d\u6587\u6280\u672f\u793e\u7fa4\uff0c\u5e73\u5747 5 \u5206\u949f\u54cd\u5e94 CUDA\u3001DeepSpeed\u3001Megatron \u62a5\u9519\u3002 <\/p>\n

\u4ece\u201c\u4e00\u5361\u96be\u6c42\u201d\u5230\u201c\u591a\u82af\u7247\u4efb\u9009\u201d\uff0c\u4ece\u201c\u767e\u4e07\u9884\u4ed8\u201d\u5230\u201c\u6309\u5c0f\u65f6\u4ed8\u8d39\u201d\uff0c\u661f\u5b87\u667a\u7b97\u8ba9\u4e2d\u5c0f\u4f01\u4e1a\u7b2c\u4e00\u6b21\u7ad9\u5728\u4e0e Meta\u3001OpenAI \u540c\u4e00\u8d77\u8dd1\u7ebf\u3002\u73b0\u5728\u6ce8\u518c\uff0c\u5373\u53ef\u9886\u53d6 10 \u5143\u4f53\u9a8c\u91d1\uff0c\u76f4\u63a5\u6253\u5361 TPU v5e \u6216 RTX 4090\uff0c\u4eb2\u624b\u611f\u53d7\u201c\u5f39\u6027\u7b97\u529b\u201d\u5e26\u6765\u7684\u7248\u672c\u7b54\u6848\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

OpenAI\u3001Meta\u90fd\u5728\u201c\u79df\u201d\u7b97\u529b\uff1a\u900f\u89c6AI\u5de8\u5934\u591a\u82af\u7247\u6218\u7565…<\/p>\n","protected":false},"author":2,"featured_media":2315,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-2316","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-zixun"],"views":73,"_links":{"self":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts\/2316","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/comments?post=2316"}],"version-history":[{"count":0,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts\/2316\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/media\/2315"}],"wp:attachment":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/media?parent=2316"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/categories?post=2316"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/tags?post=2316"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}