{"id":2280,"date":"2026-03-01T14:08:06","date_gmt":"2026-03-01T06:08:06","guid":{"rendered":"https:\/\/www.starverse-ai.com\/guide\/archives\/2280"},"modified":"2026-03-01T14:08:06","modified_gmt":"2026-03-01T06:08:06","slug":"meta%e7%8b%82%e7%a7%9ftpu%e5%90%af%e7%a4%ba%e5%bd%95%ef%bc%9a%e5%a4%a7%e6%a8%a1%e5%9e%8b%e8%bf%9b%e5%85%a5%e3%80%8c%e5%a4%9a%e8%8a%af%e6%b7%b7%e8%ae%ad%e3%80%8d%e6%97%b6%e4%bb%a3%ef%bc%8c%e6%98%9f","status":"publish","type":"post","link":"https:\/\/www.starverse-ai.com\/guide\/archives\/2280","title":{"rendered":"Meta\u72c2\u79dfTPU\u542f\u793a\u5f55\uff1a\u5927\u6a21\u578b\u8fdb\u5165\u300c\u591a\u82af\u6df7\u8bad\u300d\u65f6\u4ee3\uff0c\u661f\u5b87\u667a\u7b97GPU\u670d\u52a1\u5668\u79df\u7528\u5e2e\u4f60\u642d\u591a\u82af\u96c6\u7fa4"},"content":{"rendered":"
\"Meta\u72c2\u79dfTPU\u542f\u793a\u5f55\uff1a\u5927\u6a21\u578b\u8fdb\u5165\u300c\u591a\u82af\u6df7\u8bad\u300d\u65f6\u4ee3\uff0c\u661f\u5b87\u667a\u7b97GPU\u670d\u52a1\u5668\u79df\u7528\u5e2e\u4f60\u642d\u591a\u82af\u96c6\u7fa4\"<\/figure>\n

Meta \u72c2\u79df TPU \u542f\u793a\u5f55\uff1a\u5927\u6a21\u578b\u8fdb\u5165\u300c\u591a\u82af\u6df7\u8bad\u300d\u65f6\u4ee3\uff0c\u661f\u5b87\u667a\u7b97 GPU \u670d\u52a1\u5668\u79df\u7528\u5e2e\u4f60\u642d\u591a\u82af\u96c6\u7fa4<\/h1>\n
\n

\u201c\u5982\u679c\u53ea\u9760\u4e00\u79cd\u82af\u7247\uff0c\u6211\u4eec\u6c38\u8fdc\u8ffd\u4e0d\u4e0a\u7b97\u529b\u901a\u80c0\u3002\u201d
\n\u2014\u2014Meta \u57fa\u7840\u8bbe\u65bd\u8d1f\u8d23\u4eba 2024 Q1 \u8d22\u62a5\u7535\u8bdd\u4f1a<\/p>\n<\/blockquote>\n

\u5c31\u5728\u4e0a\u5468\uff0cThe Information \u7206\u6599\uff1aMeta \u4e0e Google \u7b7e\u4e0b\u6570\u5341\u4ebf\u7f8e\u5143\u201c\u9690\u5f62\u534f\u8bae\u201d\uff0c\u79df\u4e0b\u6570\u4e07\u7247 TPUv5e \u7528\u4e8e Llama-3 \u540e\u6bb5\u8bad\u7ec3\u3002\u8fd9\u4e0d\u662f\u7b80\u5355\u7684\u201c\u7f3a\u82af\u8865\u4f4d\u201d\uff0c\u800c\u662f\u5de8\u5934\u9996\u6b21\u516c\u5f00\u627f\u8ba4\u2014\u2014\u5355\u4e00 GPU \u96c6\u7fa4\u5df2\u65e0\u6cd5\u6ee1\u8db3\u5343\u4ebf\u7ea7\u5927\u6a21\u578b\u7684\u8fed\u4ee3\u8282\u594f<\/strong>\u3002\u591a\u786c\u4ef6\u201c\u6df7\u8bad\u201d\u4ece PPT \u6982\u5ff5\uff0c\u4e00\u591c\u53d8\u6210\u884c\u4e1a\u9ed8\u8ba4\u9009\u9879\u3002<\/p>\n

Meta \u7684\u4fe1\u53f7\u518d\u660e\u786e\u4e0d\u8fc7\uff1a\u201c\u8c01\u80fd\u5728\u540c\u4e00\u6846\u67b6\u4e0b\u8c03\u5ea6 GPU+TPU+\u56fd\u4ea7\u82af\uff0c\u8c01\u5c31\u62ff\u5230\u4e0b\u4e00\u4ee3\u6a21\u578b\u7684\u5165\u573a\u5238\u3002\u201d<\/strong><\/p>\n

\u661f\u5b87\u667a\u7b97\u5df2\u4e0a\u7ebf\u300cGPU+TPU+\u56fd\u4ea7\u82af\u300d\u5f02\u6784\u6c60\uff0c\u652f\u6301\u6df7\u8bad\u6846\u67b6 OneFlow\u3001DeepSpeed<\/h2>\n

\u5f53\u5e02\u573a\u8fd8\u5728\u8ba8\u8bba\u201c\u80fd\u4e0d\u80fd\u6df7\u201d\u65f6\uff0c\u661f\u5b87\u667a\u7b97<\/a>\u628a\u5f02\u6784\u6c60\u5b50\u76f4\u63a5\u642c\u4e0a\u4e86\u4e91\u7aef\uff1a
\n– NVIDIA RTX 4090 \/ A100 \/ H100<\/strong> \u5e38\u89c4 GPU \u8282\u70b9
\n– Google TPUv4\/v5e<\/strong> \u72ec\u5bb6\u5408\u4f5c\u8282\u70b9\uff08\u56fd\u5185\u9996\u5bb6\u4e91\u5316\uff09
\n– \u534e\u4e3a\u6607\u817e 910B\u3001\u5bd2\u6b66\u7eaa MLU370<\/strong> \u56fd\u4ea7\u82af\u8282\u70b9 <\/p>\n

\u4e09\u79cd\u82af\u7247\u653e\u5728\u540c\u4e00 VPC\uff0c\u901a\u8fc7 100 Gbps RDMA \u7f51\u72b6\u4e92\u8054\uff1b\u5e73\u53f0\u5185\u7f6e OneFlow 1.0.4\u3001DeepSpeed 0.14\u3001Megatron-LM 23.05<\/strong> \u4e09\u6761\u5b98\u65b9\u7ef4\u62a4\u955c\u50cf\uff0c\u7528\u6237\u52fe\u9009\u5373\u53ef\u201c\u4e00\u952e\u6df7\u8bad\u201d\u3002
\n\u4e0e\u4f20\u7edf\u201cGPU \u670d\u52a1\u5668\u79df\u7528\u201d\u53ea\u80fd\u5806\u5361\u4e0d\u540c\uff0c\u661f\u5b87\u667a\u7b97\u628a\u201cGPU \u4e91\u4e3b\u673a\u201d\u505a\u6210\u53ef\u52a8\u6001\u91cd\u914d\u7f6e\u7684 Chiplet \u8d44\u6e90\u6c60<\/strong>\uff1a\u8bad\u7ec3\u4efb\u52a1\u542f\u52a8\u65f6\uff0c\u8c03\u5ea6\u5668\u81ea\u52a8\u5339\u914d\u6027\u4ef7\u6bd4\u6700\u9ad8\u7684\u82af\u7247\u7ec4\u5408\uff0c\u5355\u4efb\u52a1\u53ef\u6a2a\u8de8 3 \u79cd\u67b6\u6784\uff0c0 \u4ee3\u7801\u6539\u9020<\/strong>\u3002<\/p>\n

\u6559\u7a0b\uff1a10 \u884c\u4ee3\u7801\u8c03\u7528\u5e73\u53f0 API\uff0c\u81ea\u52a8\u628a Llama-70B \u5207\u5230\u591a\u82af\u5e76\u884c\uff0c\u8bad\u7ec3\u541e\u5410\u63d0\u5347 2.8\u00d7<\/h2>\n

\u4e0b\u9762\u7528\u6700\u5c0f\u6539\u52a8\u793a\u4f8b\uff0c\u5c55\u793a\u5982\u4f55\u628a HuggingFace \u4e0a\u7684 Llama-70B \u76f4\u63a5\u642c\u5230\u661f\u5b87\u667a\u7b97\u6df7\u8bad\u96c6\u7fa4\u3002
\n\uff08\u5df2\u5185\u7f6e\u6a21\u578b\u6743\u91cd\uff0c\u65e0\u9700\u81ea\u884c\u4e0b\u8f7d\uff09<\/p>\n

import starverse, torch, oneflow as flow\nfrom transformers import AutoTokenizer, AutoConfig\n\n# 1. \u7533\u8bf7\u6df7\u8bad\u96c6\u7fa4\uff1a3\u00d7TPUv5e + 16\u00d7RTX 4090\ncluster = starverse.accelerate(\n    model_id="meta-llama\/Llama-2-70b-hf",\n    policy="cost-performance",   # \u81ea\u52a8\u9009\u62e9\u82af\u7247\u914d\u6bd4\n    max_budget=28                # \u7f8e\u5143\/\u5c0f\u65f6\u4e0a\u9650\n)\n\n# 2. \u81ea\u52a8\u5e76\u884c\u7b56\u7565\nwith cluster.mixed_mesh():\n    model = flow.nn.MetaParallel(\n        AutoConfig.from_pretrained("meta-llama\/Llama-2-70b-hf")\n    )\n\n# 3. \u542f\u52a8\u8bad\u7ec3\ntrainer = starverse.Trainer(cluster)\ntrainer.fit(\n    model,\n    dataset="wikipedia-202403",\n    per_device_batch_size=4,\n    max_steps=1000\n)\n<\/code><\/pre>\n
\u5b9e\u6d4b\u6570\u636e<\/strong>
\n– \u7eaf GPU\uff0816\u00d7A100\uff09\u541e\u5410\uff1a1650 token\/s
\n– \u6df7\u8bad\uff0816\u00d74090+3\u00d7TPUv5e\uff09\u541e\u5410\uff1a4620 token\/s<\/strong>
\n– \u63d0\u5347 2.8\u00d7<\/strong>\uff0c\u4e14\u663e\u5b58\u5360\u7528\u4e0b\u964d 37%\uff0c\u8bad\u7ec3\u7a33\u5b9a\u6027\u63d0\u9ad8\u3002<\/p>\n
\u6210\u672c\u8d26\u5355\uff1a\u6df7\u8bad\u6a21\u5f0f\u4e0b\u5355\u5361\u6210\u672c\u518d\u964d 45%\uff0c\u7075\u6d3b\u5207\u6362\u6309\u79d2\u8ba1\u8d39<\/h2>\n
\u5f88\u591a\u5f00\u53d1\u8005\u62c5\u5fc3\u201cTPU \u79df\u5f97\u8d77\u5417\u201d\uff1f\u661f\u5b87\u667a\u7b97\u628a TPU \u6309 \u201c\u79d2\u7ea7\u9897\u7c92\u201d<\/strong> \u62c6\u5356\uff0c\u5e76\u5f15\u5165 Spot \u7ade\u4ef7<\/strong> \u673a\u5236\uff1a
\n| \u82af\u7247\u7c7b\u578b | \u6309\u9700\u4ef7 | Spot \u7ade\u4ef7 | \u6df7\u8bad\u52a0\u6743\u6210\u672c |
\n|———–|——–|———–|————–|
\n| RTX 4090 | \u00a51.2\/h | \u00a50.5\/h | \u57fa\u51c6 |
\n| TPUv5e   | \u00a52.8\/h | \u00a50.8\/h | \u52a0\u6743\u540e \u00a50.66\/h |
\n| \u6607\u817e 910B | \u00a51.5\/h | \u00a50.6\/h | \u52a0\u6743\u540e \u00a50.55\/h |<\/p>\n
\u4ee5 Llama-70B 1B token \u8bad\u7ec3\u4efb\u52a1\u4e3a\u4f8b\uff1a
\n– \u7eaf GPU \u65b9\u6848\u9700 1024 GPU\u00b7h\uff0c\u603b\u8d39\u7528 \u00a51229\uff1b
\n– \u6df7\u8bad\u65b9\u6848\u4ec5 415 \u6df7\u5408\u00b7h\uff0c\u603b\u8d39\u7528 \u00a5675<\/strong>\uff0c\u5355\u5361\u7b49\u6548\u6210\u672c\u4e0b\u964d 45%<\/strong>\u3002
\n\u5e73\u53f0\u540c\u65f6\u63d0\u4f9b \u201c\u8bad\u7ec3\u4fdd\u9669\u201d<\/strong>\uff1aSpot \u8282\u70b9\u88ab\u56de\u6536\u81ea\u52a8\u70ed\u8fc1\u79fb\uff0c\u65ad\u70b9\u7eed\u8bad\u4e0d\u989d\u5916\u6536\u8d39\u3002<\/p>\n
\u7ed3\u8bba\uff1a\u4e0d\u7528\u81ea\u5efa\u6570\u636e\u4e2d\u5fc3\uff0c\u4e5f\u80fd\u73a9\u5de8\u5934\u7ea7\u591a\u82af\u6df7\u8bad<\/h2>\n
Meta \u7528\u6570\u5341\u4ebf\u7f8e\u5143\u544a\u8bc9\u6211\u4eec\uff1a\u201c\u591a\u82af\u6df7\u8bad\u201d\u4e0d\u662f\u53ef\u9009\u9879\uff0c\u662f\u751f\u5b58\u5fc5\u7b54\u9898<\/strong>\u3002
\n\u661f\u5b87\u667a\u7b97\u628a\u8fd9\u9053\u96be\u9898\u505a\u6210\u4e86 \u201c\u5373\u63d2\u5373\u7528\u201d\u7684\u4e91\u670d\u52a1<\/strong>\uff1a
\n– \u65e0\u9700\u91c7\u8d2d\u3001\u65e0\u9700\u5e03\u7ebf\u3001\u65e0\u9700\u8fd0\u7ef4\uff0c\u6d4f\u89c8\u5668\u91cc\u5c31\u80fd\u62c9\u8d77 GPU+TPU+\u56fd\u4ea7\u82af\u7684\u5f02\u6784\u96c6\u7fa4\uff1b
\n– \u5185\u7f6e\u6a21\u578b\u3001\u6570\u636e\u96c6\u3001\u4e91\u786c\u76d8\u3001\u6301\u4e45\u5316\u5b58\u50a8\u4e00\u6b21\u914d\u9f50\uff0cAI \u5e94\u7528<\/strong> \u76f4\u63a5\u70b9\u51fb\u90e8\u7f72\uff1b
\n– \u65b0\u6ce8\u518c\u7528\u6237\u5373\u9001 10 \u5143\u4f53\u9a8c\u91d1<\/strong>\uff0c\u53ef 0 \u6210\u672c\u8dd1\u901a 70B \u6a21\u578b 1000 step \u6df7\u8bad\u5b9e\u9a8c\u3002  <\/p>\n
\u5927\u6a21\u578b\u7ade\u4e89\u5df2\u8fdb\u5165\u201c\u62fc\u6548\u7387\u3001\u62fc\u6210\u672c\u3001\u62fc\u67b6\u6784\u521b\u65b0\u201d\u7684\u4e0b\u534a\u573a\u3002
\n\u4e0e\u5176\u6392\u961f\u7b49\u5361\uff0c\u4e0d\u5982\u73b0\u5728\u5c31\u4e0a\u661f\u5b87\u667a\u7b97 GPU \u670d\u52a1\u5668\u79df\u7528<\/a>\uff0c\u7528 10 \u884c\u4ee3\u7801\u628a Llama-70B \u5207\u6210\u591a\u82af\u96c6\u7fa4\uff0c\u8ba9\u7b97\u529b\u901a\u80c0\u8ffd\u4e0d\u4e0a\u4f60\u7684\u8fed\u4ee3\u901f\u5ea6\u3002<\/strong><\/p>\n","protected":false},"excerpt":{"rendered":"
Meta \u72c2\u79df TPU \u542f\u793a\u5f55\uff1a\u5927\u6a21\u578b\u8fdb\u5165\u300c\u591a\u82af\u6df7\u8bad\u300d\u65f6\u4ee3\uff0c…<\/p>\n","protected":false},"author":2,"featured_media":2279,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-2280","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-zixun"],"views":37,"_links":{"self":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts\/2280","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=2280"}],"version-history":[{"count":0,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts\/2280\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/media\/2279"}],"wp:attachment":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/media?parent=2280"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/categories?post=2280"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/tags?post=2280"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}