{"id":2782,"date":"2026-03-07T10:21:30","date_gmt":"2026-03-07T02:21:30","guid":{"rendered":"https:\/\/www.starverse-ai.com\/guide\/archives\/2782"},"modified":"2026-03-07T10:21:30","modified_gmt":"2026-03-07T02:21:30","slug":"%e5%a4%a7%e6%a8%a1%e5%9e%8b%e6%97%b6%e4%bb%a3%e3%80%8c%e7%ae%97%e5%8a%9b%e9%93%81%e4%b8%89%e8%a7%92%e3%80%8d%ef%bc%9a%e6%98%9f%e5%ae%87%e6%99%ba%e7%ae%97%e5%a6%82%e4%bd%95%e5%90%8c%e6%ad%a5%e8%a7%a3","status":"publish","type":"post","link":"https:\/\/www.starverse-ai.com\/guide\/archives\/2782","title":{"rendered":"\u5927\u6a21\u578b\u65f6\u4ee3\u300c\u7b97\u529b\u94c1\u4e09\u89d2\u300d\uff1a\u661f\u5b87\u667a\u7b97\u5982\u4f55\u540c\u6b65\u89e3\u51b3\u8ba1\u7b97\u3001\u5b58\u50a8\u3001\u7f51\u7edc"},"content":{"rendered":"
\"\u5927\u6a21\u578b\u65f6\u4ee3\u300c\u7b97\u529b\u94c1\u4e09\u89d2\u300d\uff1a\u661f\u5b87\u667a\u7b97\u5982\u4f55\u540c\u6b65\u89e3\u51b3\u8ba1\u7b97\u3001\u5b58\u50a8\u3001\u7f51\u7edc\"<\/figure>\n

\u5927\u6a21\u578b\u65f6\u4ee3\u300c\u7b97\u529b\u94c1\u4e09\u89d2\u300d\uff1a\u661f\u5b87\u667a\u7b97\u5982\u4f55\u540c\u6b65\u89e3\u51b3\u8ba1\u7b97\u3001\u5b58\u50a8\u3001\u7f51\u7edc<\/h1>\n
\n

\u201c\u7f3a\u5c11\u7b97\u529b\uff0c\u6a21\u578b\u8dd1\u4e0d\u52a8\uff1b\u7f3a\u5c11\u5b58\u529b\uff0c\u6570\u636e\u642c\u4e0d\u52a8\uff1b\u7f3a\u5c11\u8fd0\u529b\uff0c\u68af\u5ea6\u7b49\u4e0d\u8d77\u3002\u201d
\n\u2014\u2014 Mirantis \u6700\u65b0\u767d\u76ae\u4e66\u300aThe AI Triad\u300b<\/p>\n<\/blockquote>\n

\u5f53 1750 \u4ebf\u53c2\u6570\u7684 GPT-3 \u521a\u628a\u201c\u5927\u6a21\u578b\u201d\u4e09\u4e2a\u5b57\u5199\u8fdb\u79d1\u6280\u5934\u6761\u65f6\uff0c\u4f01\u4e1a\u53ea\u9700\u5806 GPU \u5c31\u80fd\u6362\u6765 N \u500d\u7684\u6027\u80fd\u63d0\u5347\u3002\u4eca\u5929\uff0c\u4e07\u4ebf\u53c2\u6570\u5df2\u6210\u6807\u914d\uff0c\u5355\u5361\u7b97\u529b\u5374\u903c\u8fd1\u7269\u7406\u6781\u9650\uff0c\u8bad\u7ec3\u96c6\u7fa4\u7684\u74f6\u9888\u6084\u7136\u8f6c\u79fb\uff1a
\ncheckpoint \u5199\u5165\u6162\u3001All-Reduce \u5ef6\u8fdf\u9ad8\u3001\u5e76\u884c\u6548\u7387\u8dcc\u5230 70% \u4ee5\u4e0b\u2014\u2014\u201c\u7b97\u529b\u3001\u5b58\u529b\u3001\u8fd0\u529b\u201d\u4efb\u4f55\u4e00\u5757\u77ed\u677f\uff0c\u90fd\u4f1a\u8ba9\u9ad8\u6602\u7684 GPU \u6295\u8d44\u77ac\u95f4\u8d2c\u503c<\/strong>\u3002 <\/p>\n

\u661f\u5b87\u667a\u7b97\u628a\u8fd9\u4e09\u5757\u677f\u540c\u65f6\u62c9\u957f\uff0c\u62fc\u6210\u4e00\u5957\u53ef\u7b7e\u5bf9\u8d4c\u534f\u8bae\u7684\u300c\u7b97\u5b58\u7f51\u300d\u4e00\u4f53\u5316 SLA\uff0c\u8ba9 GPU\u670d\u52a1\u5668\u79df\u7528 \u771f\u6b63\u50cf\u6c34\u7535\u4e00\u6837\u968f\u53d6\u968f\u7528\u3002<\/p>\n

\n

\u4e00\u3001\u7b97\u529b\uff1aRTX 4090 \u4e0d\u662f\u7ec8\u70b9\uff0c\u800c\u662f\u8d77\u70b9<\/h2>\n

\u4f20\u7edf\u4e91\u5382\u5546\u5356\u7684\u662f\u201c\u88f8\u91d1\u5c5e\u201d\uff0c\u661f\u5b87\u667a\u7b97\u5356\u7684\u662f\u201c\u6d41\u6c34\u7ebf\u201d\u3002
\n\u5e73\u53f0\u9884\u7f6e CUDA\u3001PyTorch\u3001DeepSpeed\u3001Megatron-LM \u7b49\u4e3b\u6d41\u6846\u67b6\uff0cGPU\u4e91\u4e3b\u673a<\/strong>\u5f00\u673a\u5373\u81ea\u5e26\u4f18\u5316\u540e\u7684 NCCL \u73af\u5883\uff0c\u7528\u6237\u65e0\u9700\u518d\u4e3a 4090 \u7684\u529f\u8017\u5899\u6216 PCIe \u62d3\u6251\u8c03\u4f18\u3002
\n\u66f4\u5173\u952e\u7684\u662f\u2014\u2014\u65b0\u7528\u6237\u6ce8\u518c\u5373\u9001 10 \u5143\u4f53\u9a8c\u91d1<\/strong>\uff0c\u53ef 0 \u6210\u672c\u62c9\u8d77 8 \u5361 4090 \u5b9e\u4f8b\uff0c\u8dd1\u901a 7B \u6a21\u578b\u5fae\u8c03\uff0c\u9a8c\u8bc1\u601d\u8def\u540e\u518d\u5f39\u6027\u6269\u5bb9\u5230 64 \u5361\u3001128 \u5361\uff0c\u6309\u5206\u949f\u8ba1\u8d39\uff0c\u65e0\u9884\u7559\u6210\u672c<\/strong>\u3002 <\/p>\n

\n

\u4e8c\u3001\u5b58\u529b\uff1aNVMe-oF \u5206\u5e03\u5f0f\u5b58\u50a8\uff0c\u8ba9 checkpoint \u4ece\u5206\u949f\u7ea7\u5230\u79d2\u7ea7<\/h2>\n

\u5927\u6a21\u578b\u8bad\u7ec3\u6700\u6015\u201c\u5199\u65ad\u70b9\u201d\u3002
\n\u4f20\u7edf NFS \u65b9\u6848\u5728 100 GB \u7ea7\u522b\u7684 checkpoint \u9762\u524d\u5199\u5165\u901f\u5ea6\u4ec5 600 MB\/s\uff0c\u4e00\u4e2a epoch \u8981\u7b49\u5f85 3 \u5206\u949f<\/strong>\uff0cGPU \u7a7a\u8f6c\u70e7\u7684\u662f\u771f\u91d1\u767d\u94f6\u3002 <\/p>\n

\u661f\u5b87\u667a\u7b97\u81ea\u7814\u7684 NVMe-oF \u5206\u5e03\u5f0f\u5b58\u50a8\uff0c\u5355\u5ba2\u6237\u7aef\u6301\u7eed\u541e\u5410 3 GB\/s<\/strong>\uff0c\u914d\u5408 RDMA \u7f51\u7edc\uff0ccheckpoint \u4fdd\u5b58\u63d0\u901f 5\u00d7<\/strong>\uff0c\u8ba9 175B \u6a21\u578b\u4e5f\u80fd 30 \u79d2\u5b8c\u6210\u4e00\u6b21\u65ad\u70b9\u843d\u5730\u3002
\n\u6570\u636e\u7ba1\u7406\u66f4\u8d34\u5fc3\uff1a
\n– \u4e91\u786c\u76d8<\/a> \u53ef\u5728\u591a\u5b9e\u4f8b\u95f4\u70ed\u63d2\u62d4\uff0c\u8bad\u7ec3\u4e0e\u63a8\u7406\u65e0\u7f1d\u5207\u6362\uff1b
\n– \u4e91\u5b58\u50a8<\/a> \u652f\u6301 Web \u7aef\u4e00\u952e\u4e0a\u4f20\uff0cPB \u7ea7\u6570\u636e\u96c6\u65e0\u9700\u91cd\u590d\u4e0b\u8f7d\uff1b
\n– \u6a21\u578b\u548c\u6570\u636e\u96c6<\/a> \u516c\u5171\u5e93\u5df2\u6302\u8f7d\u81f3\u5b9e\u4f8b\uff0ccp -r \/public\/llama-2-70b .\/<\/code> \u5373\u53ef\u5f00\u7ec3\u3002 <\/p>\n

\n

\u4e09\u3001\u8fd0\u529b\uff1a100 Gbps RDMA\uff0cAll-Reduce \u5ef6\u8fdf 2 \u03bcs<\/h2>\n

\u7b97\u529b\u4e0e\u5b58\u529b\u518d\u5f3a\uff0c\u7f51\u7edc\u62d6\u540e\u817f\u7167\u6837\u201c\u68af\u5ea6\u7b49\u4eba\u201d\u3002
\n\u661f\u5b87\u667a\u7b97\u5168\u7ebf\u63a5\u5165 100 Gbps RoCE v2<\/strong>\uff0c\u4ea4\u6362\u673a\u652f\u6301 SHARP v3 \u5f15\u64ce\uff0cAll-Reduce \u5ef6\u8fdf\u4f4e\u81f3 2 \u03bcs<\/strong>\uff0c\u76f8\u6bd4\u4f20\u7edf 25 Gbps TCP \u7f51\u7edc\uff0c\u5343\u4ebf\u6a21\u578b\u5e76\u884c\u6548\u7387\u63d0\u5347 18%<\/strong>\u3002 <\/p>\n

\u5b9e\u6d4b\u6570\u636e\uff1a
\n– 128 \u5361 4090 \u8bad\u7ec3 175B \u6a21\u578b\uff0cDP+TP+PP \u4e09\u9636\u6df7\u5408\u5e76\u884c\uff0c\u6548\u7387\u7a33\u5b9a\u5728 90% \u4ee5\u4e0a<\/strong>\uff1b
\n– \u6bcf\u4e07\u4ebf token \u8bad\u7ec3\u6210\u672c\u8f83\u4f20\u7edf\u4e91\u4e0b\u964d 23%<\/strong>\uff0c\u8bad\u7ec3\u5468\u671f\u4ece 30 \u5929\u538b\u7f29\u5230 24 \u5929<\/strong>\u3002 <\/p>\n

\n

\u56db\u3001B \u7aef\u515c\u5e95\uff1a\u7b7e\u5f97\u4e0b\u5bf9\u8d4c\uff0c\u624d\u62ff\u5f97\u51fa\u5e95\u6c14<\/h2>\n

\u4f01\u4e1a\u5ba2\u6237\u6700\u6015\u201c\u53e3\u5934\u9ad8\u6027\u80fd\uff0c\u843d\u5730\u6253\u5bf9\u6298\u201d\u3002
\n\u661f\u5b87\u667a\u7b97\u628a\u300c\u7b97\u5b58\u7f51\u300d\u6307\u6807\u5199\u8fdb SLA\uff1a
\n– GPU \u5229\u7528\u7387 \u2265 95% \u6301\u7eed 7\u00d724 \u5c0f\u65f6\uff1b
\n– \u5b58\u50a8\u5199\u5165\u541e\u5410\u4e0d\u4f4e\u4e8e 2.5 GB\/s\uff1b
\n– \u7f51\u7edc\u5ef6\u8fdf P99 \u2264 5 \u03bcs\uff1b
\n\u4efb\u4e00\u6307\u6807\u672a\u8fbe\u6807\uff0c\u6309\u5c0f\u65f6\u8d54\u4ed8 10% \u79df\u91d1<\/strong>\uff0c\u53ef\u53e0\u52a0\uff0c\u4e0d\u8bbe\u4e0a\u9650\u3002 <\/p>\n

\u76ee\u524d\u5df2\u6709\u4e09\u5bb6\u5934\u90e8\u5927\u6a21\u578b\u521b\u4f01\u7b7e\u4e0b\u5bf9\u8d4c\u534f\u8bae\uff0c\u628a 300 \u5361 4090 \u96c6\u7fa4\u7684\u6708\u8d26\u671f\u4ece\u9884\u4ed8\u6539\u4e3a\u540e\u4ed8<\/strong>\uff0c\u661f\u5b87\u667a\u7b97\u7528\u771f\u91d1\u767d\u94f6\u4e3a\u81ea\u5df1\u7684\u6280\u672f\u80cc\u4e66\u3002 <\/p>\n

\n

\u4e94\u3001\u751f\u6001\uff1a\u8ba9 AI \u5e94\u7528\u201c\u4e00\u952e\u5373\u73a9\u201d<\/h2>\n

\u7b97\u529b\u3001\u5b58\u50a8\u3001\u7f51\u7edc\u53ea\u662f\u5730\u57fa\uff0cAI\u5e94\u7528<\/strong> \u624d\u662f\u6700\u7ec8\u7684\u5546\u54c1\u623f\u3002
\n\u661f\u5b87\u667a\u7b97\u4e0a\u7ebf\u300c\u5e94\u7528\u5e02\u573a\u300d\uff1a
\n– \u5f00\u53d1\u8005\u4e0a\u4f20\u955c\u50cf\uff0c\u5e73\u53f0\u81ea\u52a8\u5b8c\u6210 CUDA \u9a71\u52a8\u3001Python \u4f9d\u8d56\u3001\u7aef\u53e3\u6620\u5c04\uff1b
\n– \u9700\u6c42\u65b9\u50cf\u8ba2\u9605 SaaS \u4e00\u6837\u4e0b\u5355\uff0c\u79d2\u7ea7\u62c9\u8d77 GPU\u4e91\u4e3b\u673a<\/strong>\uff0c\u652f\u6301 Gradio\u3001Streamlit\u3001FastAPI \u7b49\u4e3b\u6d41\u6846\u67b6\uff1b
\n– \u6536\u76ca\u5206\u6210 7:3\uff0c\u5f00\u53d1\u8005\u62ff\u5927\u5934\uff0c\u661f\u5b87\u667a\u7b97\u53ea\u6536\u5e73\u53f0\u670d\u52a1\u8d39<\/strong>\u3002 <\/p>\n

\u65e0\u8bba\u662f\u9ad8\u6821\u5e08\u751f\u505a\u79d1\u7814\uff0c\u8fd8\u662f\u521d\u521b\u516c\u53f8\u8dd1 AIGC \u5546\u4e1a\u5316\uff0c\u90fd\u80fd\u5728\u6700\u77ed\u65f6\u95f4\u5185\u628a\u521b\u610f\u53d8\u6210\u53ef\u8bbf\u95ee\u7684 URL<\/strong>\u3002 <\/p>\n

\n

\u516d\u3001\u7acb\u5373\u4f53\u9a8c\uff1a10 \u5143\u767d\u5ad6\uff0c\u5148\u8dd1\u518d\u8bf4<\/h2>\n

\u5927\u6a21\u578b\u8bad\u7ec3\u4ece\u6765\u4e0d\u662f\u201c\u6709\u94b1\u5c31\u80fd\u5806\u51fa\u6765\u201d\u7684\u6e38\u620f\uff0c\u9009\u5bf9\u7b97\u529b\u5e73\u53f0\u624d\u662f ROI \u7684\u7b2c\u4e00\u6027\u539f\u7406<\/strong>\u3002
\n\u73b0\u5728\u6ce8\u518c \u661f\u5b87\u667a\u7b97<\/a>\uff0c\u65b0\u7528\u6237\u76f4\u63a5\u5230\u8d26 10 \u5143\u4f53\u9a8c\u91d1<\/strong>\uff0c\u53ef\u96f6\u6210\u672c\u62c9\u8d77 8 \u5361 RTX 4090 \u5b9e\u4f8b\uff0c5 \u5206\u949f\u5b8c\u6210 Llama-2-7B \u5fae\u8c03<\/strong>\u3002
\n\u522b\u8ba9 GPU \u7a7a\u8f6c\uff0c\u522b\u8ba9\u6570\u636e\u7b49\u5f85\uff0c\u522b\u8ba9\u7f51\u7edc\u62d6\u540e\u817f\u2014\u2014
\n\u628a\u7b97\u529b\u3001\u5b58\u529b\u3001\u8fd0\u529b\u4e00\u6b21\u6027\u6253\u5e73\uff0c\u5269\u4e0b\u7684\u4ea4\u7ed9\u521b\u65b0\u672c\u8eab\u3002<\/strong><\/p>\n","protected":false},"excerpt":{"rendered":"

\u5927\u6a21\u578b\u65f6\u4ee3\u300c\u7b97\u529b\u94c1\u4e09\u89d2\u300d\uff1a\u661f\u5b87\u667a\u7b97\u5982\u4f55\u540c\u6b65\u89e3\u51b3\u8ba1\u7b97\u3001\u5b58\u50a8\u3001\u7f51…<\/p>\n","protected":false},"author":2,"featured_media":2781,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-2782","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\/2782","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=2782"}],"version-history":[{"count":0,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts\/2782\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/media\/2781"}],"wp:attachment":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/media?parent=2782"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/categories?post=2782"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/tags?post=2782"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}