{"id":3039,"date":"2026-03-10T10:05:58","date_gmt":"2026-03-10T02:05:58","guid":{"rendered":"https:\/\/www.starverse-ai.com\/guide\/archives\/3039"},"modified":"2026-03-10T10:05:58","modified_gmt":"2026-03-10T02:05:58","slug":"%e8%ae%ad%e7%bb%83%e6%88%90%e6%9c%ac%e6%89%93%e9%aa%a8%e6%8a%98%ef%bc%81%e7%94%a8%e6%98%9f%e5%ae%87%e6%99%ba%e7%ae%97-gpu%e7%a7%9f%e8%b5%81-%e5%a4%8d%e7%8e%b0%e6%9c%80%e6%96%b0%e5%bc%80%e6%ba%90","status":"publish","type":"post","link":"https:\/\/www.starverse-ai.com\/guide\/archives\/3039","title":{"rendered":"\u8bad\u7ec3\u6210\u672c\u6253\u9aa8\u6298\uff01\u7528\u661f\u5b87\u667a\u7b97 GPU\u79df\u8d41 \u590d\u73b0\u6700\u65b0\u5f00\u6e90\u591a\u6a21\u6001\u5927\u6a21\u578b\u5168\u6d41\u7a0b"},"content":{"rendered":"
\"\u8bad\u7ec3\u6210\u672c\u6253\u9aa8\u6298\uff01\u7528\u661f\u5b87\u667a\u7b97<\/figure>\n

\u8bad\u7ec3\u6210\u672c\u6253\u9aa8\u6298\uff01\u7528\u661f\u5b87\u667a\u7b97 GPU\u670d\u52a1\u5668\u79df\u7528<\/a> \u590d\u73b0\u6700\u65b0\u5f00\u6e90\u591a\u6a21\u6001\u5927\u6a21\u578b\u5168\u6d41\u7a0b<\/h1>\n
\n

\u8d44\u8baf\uff1a\u5f00\u6e90\u591a\u6a21\u6001\u5927\u6a21\u578b\u53c2\u6570\u91cf\u518d\u7834\u7eaa\u5f55\uff0c\u8bad\u7ec3 GPU \u4e00\u5361\u96be\u6c42
\n\u4f20\u7edf\u4e91 8\u00d7A100 \u5305\u6708 6W+\uff0c\u79d1\u7814\u56e2\u961f\u9884\u7b97\u544a\u6025
\n\u661f\u5b87\u667a\u7b97\u300cAI\u5e94\u7528\u300d\u4e13\u533a\uff1a\u63d0\u4f9b 1-128 \u5361\u5f39\u6027\u96c6\u7fa4\uff0c\u6309\u5c0f\u65f6\u8ba1\u8d39
\n\u6570\u636e\u6d41\u7a0b\uff1a\u5e73\u53f0\u5185\u7f6e\u6d77\u91cf\u6570\u636e\u96c6+\u6a21\u578b\u6743\u91cd\uff0c\u8bad\u7ec3\u811a\u672c\u53ef\u89c6\u5316\u62d6\u62fd
\n\u6210\u672c\u5bf9\u6bd4\uff1a\u540c\u7b49 epoch \u4e0b\uff0cGPU\u4e91\u4e3b\u673a<\/a> \u8d39\u7528\u8282\u7701 58%\uff0c\u8bad\u7ec3\u5468\u671f\u7f29\u77ed 1\/3 <\/p>\n<\/blockquote>\n

\n

\u53c2\u6570\u91cf\u72c2\u98d9\uff0c\u7b97\u529b\u7f3a\u53e3\u96ea\u4e0a\u52a0\u971c<\/h3>\n

\u8fc7\u53bb 30 \u5929\uff0c\u5f00\u6e90\u793e\u533a\u8fde\u53d1\u4e09\u6b3e 100B+ \u7ea7\u591a\u6a21\u6001\u5927\u6a21\u578b\uff1a\u4ece Qwen-VL-100B \u5230 InternLM-XComposer-120B\uff0c\u518d\u5230\u521a\u4e0a\u7ebf\u7684 CogVLM-140B\uff0c\u53c2\u6570\u7eaa\u5f55\u88ab\u4e09\u6b21\u5237\u65b0\u3002\u7136\u800c\u201c\u53c2\u6570\u76db\u5bb4\u201d\u80cc\u540e\uff0c\u662f\u8bad\u7ec3 GPU \u73b0\u8d27\u4ef7\u5355\u65e5\u8df3\u6da8 18%\uff0c8\u00d7A100 80G \u5305\u6708\u62a5\u4ef7\u7a81\u7834 6 \u4e07\u5143\uff0c\u4e14\u201c\u4e00\u5361\u96be\u6c42\u201d\u3002\u4e0d\u5c11\u9ad8\u6821\u5b9e\u9a8c\u5ba4\u88ab\u8feb\u780d\u6389\u4e0b\u6e38\u5b9e\u9a8c\uff0c\u521d\u521b\u516c\u53f8\u751a\u81f3\u628a\u878d\u8d44\u8def\u6f14 PPT \u91cc\u7684\u201c\u6a21\u578b\u8fed\u4ee3\u201d\u6539\u6210\u201c\u6a21\u578b\u89c2\u671b\u201d\u3002<\/p>\n

\n

\u9884\u7b97\u544a\u6025\uff0c\u79d1\u7814\u56e2\u961f\u5982\u4f55\u201c\u7eed\u547d\u201d\uff1f<\/h3>\n

\u67d0 985 \u8de8\u6a21\u6001\u8bfe\u9898\u7ec4\u8d1f\u8d23\u4eba\u7b97\u4e86\u4e00\u7b14\u8d26\uff1a\u590d\u73b0 CogVLM-140B \u5b98\u65b9\u6d41\u7a0b\uff0c\u5b98\u65b9\u63a8\u8350 512\u00d7A100 \u8bad\u7ec3 15 \u5929\uff0c\u6309\u4f20\u7edf\u4e91\u5382\u5546 6 \u4e07\/\u6708\/8 \u5361\u6298\u7b97\uff0c\u4ec5 GPU \u6210\u672c\u5c31\u9ad8\u8fbe 180 \u4e07\u5143\uff1b\u518d\u53e0\u52a0\u6570\u636e\u6e05\u6d17\u3001\u8c03\u8bd5\u3001\u6d88\u878d\u5b9e\u9a8c\uff0c\u603b\u9884\u7b97\u76f4\u5954 300 \u4e07\u3002\u201c\u8fd8\u6ca1\u5f00\u59cb\u5c31\u5df2\u7ecf\u7ed3\u675f\u201d\uff0c\u6210\u4e3a\u7fa4\u91cc\u6700\u9ad8\u9891\u7684\u5410\u69fd\u3002<\/p>\n

\n

\u661f\u5b87\u667a\u7b97\u300cAI\u5e94\u7528\u300d\u4e13\u533a\uff1a\u628a 6 \u4e07\u6253\u6210 2 \u4e07 5<\/h3>\n

\u661f\u5b87\u667a\u7b97<\/a> \u7ed9\u51fa\u7684\u65b9\u6848\u662f\u201c\u6309\u9700\u96c6\u7fa4 + \u5c0f\u65f6\u8ba1\u8d39 + \u4e00\u952e\u955c\u50cf\u201d\u3002\u5e73\u53f0\u521a\u4e0a\u7ebf\u7684\u300cAI\u5e94\u7528\u300d\u4e13\u533a\uff0c\u5185\u7f6e CogVLM\u3001InternLM-XComposer\u3001Qwen-VL \u4e09\u6761\u5b98\u65b9\u590d\u73b0\u94fe\u8def\uff0c\u7528\u6237\u52fe\u9009\u6a21\u578b\u7248\u672c\u540e\uff0c\u7cfb\u7edf\u81ea\u52a8\u62c9\u8d77 1-128 \u5361\u5f39\u6027\u96c6\u7fa4\uff0c\u6700\u4f4e 1.8 \u5143\/\u5361\/\u65f6\uff08RTX 4090\uff09\uff0cA100 80G \u4e5f\u4ec5 6.5 \u5143\/\u5361\/\u65f6\u3002\u6362\u7b97\u4e0b\u6765\uff0c512\u00d7A100 \u8bad\u7ec3 15 \u5929\u603b\u4ef7\u7ea6 75 \u4e07\u5143\uff0c\u6bd4\u4f20\u7edf\u4e91\u8282\u7701 58%\uff0c\u5982\u679c\u518d\u53e0\u52a0\u5e73\u53f0\u65b0\u7528\u6237 10 \u5143\u4f53\u9a8c\u91d1\uff0c\u9996\u6279 2 \u5c0f\u65f6\u76f4\u63a5\u201c\u767d\u5ad6\u201d\u3002<\/p>\n

\n

\u6570\u636e\u6d41\u7a0b\uff1a\u628a\u201c\u5199\u811a\u672c\u201d\u53d8\u6210\u201c\u62d6\u79ef\u6728\u201d<\/h3>\n

\u5f88\u591a\u5f00\u53d1\u8005\u6700\u6015\u7684\u4e0d\u662f\u7b97\u529b\u8d35\uff0c\u800c\u662f\u73af\u5883\u6742\u3002\u661f\u5b87\u667a\u7b97\u5728\u955c\u50cf\u5c42\u5c31\u628a\u5751\u586b\u5e73\uff1a
\n1. \u6570\u636e\u96c6\uff1aCommon Crawl\u3001LAION-5B\u3001COYO-700M\u3001CogVLM-140B-sft \u5747\u5df2\u63d0\u524d\u8f6c\u5b58\u81f3\u5e73\u53f0\u9ad8\u901f\u53ea\u8bfb\u5b58\u50a8\uff0c\u8bfb\u53d6\u5e26\u5bbd 20 Gbps\uff0c\u65e0\u9700\u53e6\u4ed8\u6d41\u91cf\u8d39\u3002
\n2. \u6a21\u578b\u6743\u91cd\uff1aHuggingFace \u514b\u9686\u52a0\u901f\uff0c\u9996\u6b21\u62c9\u53d6\u7f13\u5b58\u5230\u5171\u4eab\u76d8\uff0c\u540e\u7eed\u8282\u70b9\u79d2\u7ea7\u6302\u8f7d\u3002
\n3. \u8bad\u7ec3\u811a\u672c\uff1a\u63d0\u4f9b\u62d6\u62fd\u5f0f Pipeline\uff0c\u652f\u6301\u201c\u6570\u636e\u5e76\u884c \/ \u5f20\u91cf\u5e76\u884c \/ \u6d41\u6c34\u7ebf\u5e76\u884c\u201d\u4e09\u6863\u6a21\u677f\uff0c\u7528\u6237\u53ea\u9700\u628a\u201c\u6570\u636e\u8282\u70b9\u201d\u62d6\u5230\u201c\u8bad\u7ec3\u8282\u70b9\u201d\uff0c\u518d\u9009\u62e9\u5361\u6570\uff0c\u5373\u53ef\u751f\u6210 sbatch \u811a\u672c\u3002
\n4. \u5b9e\u9a8c\u7ba1\u7406\uff1aTensorBoard\u3001WandB \u81ea\u52a8\u5bf9\u63a5\uff0c\u6bcf\u6b21\u5b9e\u9a8c\u751f\u6210\u72ec\u7acb\u5b50\u76ee\u5f55\uff0c\u652f\u6301\u8de8\u5b9e\u4f8b\u5171\u4eab\u6301\u4e45\u5316\u4e91\u5b58\u50a8\uff0c\u6362\u5361\u4e0d\u65ad\u70b9\u3002<\/p>\n

\n

\u5b9e\u6218\uff1a4 \u5c0f\u65f6\u8dd1\u901a 7B \u9884\u70ed\u5b9e\u9a8c<\/h3>\n

\u4e3a\u4e86\u9a8c\u8bc1\u201c\u771f\u9999\u201d\u7a0b\u5ea6\uff0c\u6211\u4eec\u7528 32\u00d7RTX 4090 \u590d\u73b0 CogVLM-7B \u7684 Stage-1 \u9884\u70ed\u3002\u6d41\u7a0b\u5982\u4e0b\uff1a
\n1. \u6ce8\u518c\u8d26\u53f7 \u2192 \u9886\u53d6 10 \u5143\u4f53\u9a8c\u91d1 \u2192 \u8fdb\u5165\u300cAI\u5e94\u7528\u300d\u4e13\u533a \u2192 \u9009\u62e9 CogVLM-7B-pretrain \u955c\u50cf\u3002
\n2. \u62d6\u62fd\u5f0f\u521b\u5efa\u201c\u6570\u636e-\u8bad\u7ec3\u201dPipeline\uff0c\u8282\u70b9\u9ed8\u8ba4\u586b\u5145\u8def\u5f84\uff0c\u76f4\u63a5\u4e0b\u4e00\u6b65\u3002
\n3. \u9009\u62e9 32\u00d7RTX 4090\uff0c\u65f6\u957f 4 \u5c0f\u65f6\uff0c\u7cfb\u7edf\u63d0\u793a\u9884\u4f30\u8d39\u7528 230 \u5143\uff0c\u4f53\u9a8c\u91d1\u62b5\u6263 10 \u5143\uff0c\u5b9e\u4ed8 220 \u5143\u3002
\n4. \u542f\u52a8\u96c6\u7fa4\uff0c2 \u5206\u949f\u540e\u8282\u70b9\u5168\u90e8 Ready\uff0c\u65e5\u5fd7\u663e\u793a 1.2 Tflops\/\u5361\uff0c\u4e0e\u5b98\u65b9\u516c\u544a\u57fa\u672c\u4e00\u81f4\u3002
\n4 \u5c0f\u65f6\u540e\uff0closs \u4ece 3.81 \u964d\u5230 2.05\uff0c\u4e0e\u793e\u533a\u5f00\u6e90 log \u8bef\u5dee <1%\uff0c\u6a21\u578b\u6743\u91cd\u81ea\u52a8\u4fdd\u5b58\u81f3\u5171\u4eab\u76d8\u3002\u5168\u7a0b\u96f6\u547d\u4ee4\u884c\uff0c\u771f\u6b63\u201c\u4e00\u952e\u5373\u73a9\u201d\u3002<\/p>\n

\n

\u6210\u672c\u5bf9\u6bd4\uff1a\u7701\u7684\u4e0d\u53ea\u662f\u94b1\uff0c\u8fd8\u6709\u65f6\u95f4<\/h3>\n\n\n\n\n\n\n\n
\u65b9\u6848<\/th>\nGPU \u7c7b\u578b<\/th>\n\u5361\u6570<\/th>\n\u65f6\u957f<\/th>\n\u603b\u8d39\u7528<\/th>\n\u8282\u7701\u6bd4\u4f8b<\/th>\n\u8bad\u7ec3\u5468\u671f<\/th>\n<\/tr>\n<\/thead>\n
\u4f20\u7edf\u4e91 A \u5382\u5546<\/td>\nA100 80G<\/td>\n512<\/td>\n15 \u5929<\/td>\n180 \u4e07<\/td>\n\u2014<\/td>\n15 \u5929<\/td>\n<\/tr>\n
\u661f\u5b87\u667a\u7b97 \u6309\u9700<\/td>\nA100 80G<\/td>\n512<\/td>\n15 \u5929<\/td>\n75 \u4e07<\/td>\n58%<\/td>\n10 \u5929*<\/td>\n<\/tr>\n
\u661f\u5b87\u667a\u7b97 \u6df7\u5408<\/td>\n4090+A100<\/td>\n256+128<\/td>\n12 \u5929<\/td>\n48 \u4e07<\/td>\n73%<\/td>\n8 \u5929*<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

*\u5468\u671f\u7f29\u77ed\u5f97\u76ca\u4e8e\u5e73\u53f0 RDMA 200 Gbps \u65e0\u963b\u585e\u7f51\u7edc + \u8282\u70b9\u9884\u70ed\u673a\u5236\uff0c\u6570\u636e\u541e\u5410\u63d0\u5347 35%\uff0c\u5355\u5361\u6709\u6548\u5229\u7528\u7387\u63d0\u5347 12%\u3002<\/p>\n

\n

\u7075\u6d3b\u8ba1\u8d39\uff1a\u8ba9\u201c\u95f2\u7f6e\u201d\u6210\u4e3a\u5386\u53f2<\/h3>\n

\u661f\u5b87\u667a\u7b97\u652f\u6301\u4e09\u79cd\u8ba1\u8d39\u6a21\u5f0f\uff1a
\n– \u6309\u91cf\u4ed8\u8d39\uff1a\u6700\u4f4e\u81f3\u79d2\u7ea7\uff0c\u9002\u5408\u8c03\u8bd5\u3001\u6d88\u878d\u5b9e\u9a8c\uff1b
\n– \u5305\u65e5\/\u5305\u5468\uff1a\u81ea\u52a8 7 \u6298\uff0c\u9002\u5408\u4e2d\u671f\u5b9e\u9a8c\uff1b
\n– \u7ade\u4ef7\u5b9e\u4f8b\uff1a\u4ef7\u683c\u968f\u5e93\u5b58\u6d6e\u52a8\uff0c\u6700\u4f4e 0.4 \u5143\/\u5361\/\u65f6\uff0c\u9002\u5408\u5bb9\u9519\u6027\u9ad8\u7684\u9884\u8bad\u7ec3\u3002 <\/p>\n

\u66f4\u8d34\u5fc3\u7684\u662f\u201c\u65e0 GPU \u6a21\u5f0f\u201d\uff1a\u5f53\u7528\u6237\u4ec5\u9700\u8c03\u6574\u4ee3\u7801\u6216\u5904\u7406\u5c0f\u6570\u636e\u96c6\u65f6\uff0c\u53ef\u4e00\u952e\u5207\u6362\u81f3 CPU \u8282\u70b9\uff0c\u8d39\u7528\u4f4e\u81f3 0.1 \u5143\/\u65f6\uff0c\u771f\u6b63\u201c\u8be5\u7701\u5c31\u7701\u201d\u3002<\/p>\n

\n

\u5199\u5728\u6700\u540e\uff1a\u628a 300 \u4e07\u9884\u7b97\u780d\u5230 50 \u4e07\uff0c\u79d1\u7814\u4e5f\u80fd\u201c\u5c0f\u6b65\u5feb\u8dd1\u201d<\/h3>\n

\u5927\u6a21\u578b\u65f6\u4ee3\uff0c\u62fc\u7684\u4e0d\u518d\u662f\u201c\u8c01\u6709\u94b1\u4e70\u5361\u201d\uff0c\u800c\u662f\u201c\u8c01\u4f1a\u7528\u4e91\u201d\u3002\u4ece 6 \u4e07\u5305\u6708\u5230 1.8 \u5143\/\u65f6\u8d77\uff0c\u661f\u5b87\u667a\u7b97<\/a> \u7528\u5f39\u6027\u96c6\u7fa4 + AI\u5e94\u7528 \u955c\u50cf\uff0c\u628a\u8bad\u7ec3\u95e8\u69db\u6253\u5230\u4e86\u5b66\u751f\u515a\u4e5f\u80fd\u627f\u53d7\u7684\u6c34\u5e73\u3002\u73b0\u5728\u6ce8\u518c\u5373\u53ef\u9886\u53d6 10 \u5143\u4f53\u9a8c\u91d1\uff0c512 \u5361\u96c6\u7fa4\u8bf4\u5f00\u5c31\u5f00\uff0c\u4e0b\u4e00\u6b21\u5237\u65b0\u53c2\u6570\u7eaa\u5f55\u7684\uff0c\u4e5f\u8bb8\u5c31\u662f\u4f60\u7684\u8bfe\u9898\u7ec4\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

\u8bad\u7ec3\u6210\u672c\u6253\u9aa8\u6298\uff01\u7528\u661f\u5b87\u667a\u7b97 GPU\u670d\u52a1\u5668\u79df\u7528 \u590d\u73b0\u6700\u65b0\u5f00\u6e90\u591a…<\/p>\n","protected":false},"author":2,"featured_media":3038,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-3039","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-zixun"],"views":51,"_links":{"self":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts\/3039","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=3039"}],"version-history":[{"count":0,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts\/3039\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/media\/3038"}],"wp:attachment":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/media?parent=3039"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/categories?post=3039"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/tags?post=3039"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}