{"id":2626,"date":"2026-03-05T14:15:20","date_gmt":"2026-03-05T06:15:20","guid":{"rendered":"https:\/\/www.starverse-ai.com\/guide\/archives\/2626"},"modified":"2026-03-05T14:15:20","modified_gmt":"2026-03-05T06:15:20","slug":"%e5%9b%bd%e4%ba%a7%e8%8a%af%e7%89%87%e6%9b%bf%e4%bb%a3%e5%8a%a0%e9%80%9f%ef%bc%8c%e6%98%9f%e5%ae%87%e6%99%ba%e7%ae%97%e5%85%bc%e5%ae%b9%e6%98%87%e8%85%be%e3%80%81%e5%af%92%e6%ad%a6%e7%ba%aa%e3%80%8c","status":"publish","type":"post","link":"https:\/\/www.starverse-ai.com\/guide\/archives\/2626","title":{"rendered":"\u56fd\u4ea7\u82af\u7247\u66ff\u4ee3\u52a0\u901f\uff0c\u661f\u5b87\u667a\u7b97\u517c\u5bb9\u6607\u817e\u3001\u5bd2\u6b66\u7eaa\u300c\u6df7\u5408 GPU\u79df\u8d41\u300d\u5b9e\u8df5"},"content":{"rendered":"
\"\u56fd\u4ea7\u82af\u7247\u66ff\u4ee3\u52a0\u901f\uff0c\u661f\u5b87\u667a\u7b97\u517c\u5bb9\u6607\u817e\u3001\u5bd2\u6b66\u7eaa\u300c\u6df7\u5408<\/figure>\n
\n

2024 \u5e74 10 \u6708\uff0c\u7f8e\u56fd\u518d\u6b21\u6536\u7d27\u5bf9\u534e\u9ad8\u7aef GPU \u51fa\u53e3\u7ba1\u5236\uff0cA100\u3001H100 \u53ca\u5176\u201c\u9609\u5272\u7248\u201dA800\u3001H800 \u5168\u90e8\u5217\u5165\u7981\u8fd0\u6e05\u5355\u3002\u82f1\u4f1f\u8fbe\u5e02\u503c\u4e00\u591c\u84b8\u53d1 500 \u4ebf\u7f8e\u5143\uff0c\u800c\u56fd\u5185\u5927\u6a21\u578b\u516c\u53f8\u5374\u4e0d\u5f97\u4e0d\u628a\u201c\u8fdb\u53e3\u5907\u80ce\u201d\u65b9\u6848\u63d0\u524d\u63d0\u4e0a\u65e5\u7a0b\u3002\u4e0e\u6b64\u540c\u65f6\uff0c\u6607\u817e 910B\u3001\u5bd2\u6b66\u7eaa MLU 370 \u7684\u91cf\u4ea7\u826f\u7387\u9996\u6b21\u7a81\u7834 70%\uff0c\u56fd\u4ea7\u9ad8\u7aef\u7b97\u529b\u6b63\u5f0f\u4ece\u201c\u80fd\u7528\u201d\u8d70\u5411\u201c\u597d\u7528\u201d\u3002\u5f53\u201c\u5361\u8116\u5b50\u201d\u53d8\u6210\u201c\u5361\u8ba2\u5355\u201d\uff0c\u4e00\u573a\u56f4\u7ed5 GPU \u670d\u52a1\u5668\u79df\u7528\u7684\u56fd\u4ea7\u66ff\u4ee3\u7ade\u8d5b\uff0c\u5df2\u7ecf\u6084\u7136\u6253\u54cd\u3002<\/p>\n<\/blockquote>\n

\u56fd\u4ea7\u82af\u7247\u66ff\u4ee3\u52a0\u901f\uff0c\u661f\u5b87\u667a\u7b97\u517c\u5bb9\u6607\u817e\u3001\u5bd2\u6b66\u7eaa\u300c\u6df7\u5408 GPU \u79df\u8d41\u300d\u5b9e\u8df5<\/h2>\n

01 \u884c\u4e1a\u80cc\u666f\uff1a\u7f3a\u5361\u3001\u8d35\u5361\u3001\u7b49\u5361\uff0c\u56fd\u4ea7 GPU \u5d1b\u8d77\u6b63\u5f53\u65f6<\/h3>\n

\u201c\u539f\u6765\u6392\u961f 8 \u5468\u80fd\u62ff 100 \u5f20 A100\uff0c\u73b0\u5728 16 \u5468\u90fd\u672a\u5fc5\u6709\u8d27\u3002\u201d\u67d0\u5934\u90e8\u5927\u6a21\u578b\u5382\u5546\u91c7\u8d2d\u8d1f\u8d23\u4eba\u76f4\u8a00\uff0c\u8fdb\u53e3\u5361\u201c\u4e00\u5361\u96be\u6c42\u201d\u76f4\u63a5\u62d6\u6162\u4e86\u8fed\u4ee3\u8282\u594f\u3002
\n\u53e6\u4e00\u9762\uff0c\u534e\u4e3a\u6607\u817e 910B \u5355\u5361 FP16 \u7b97\u529b 320 TFLOPS\uff0c\u5bd2\u6b66\u7eaa MLU 370-X8 \u5728 350W \u529f\u8017\u4e0b\u4e5f\u80fd\u8dd1\u51fa 256 TFLOPS\uff0c\u6307\u6807\u5df2\u903c\u8fd1 A100\u3002\u66f4\u91cd\u8981\u7684\u662f\uff0c\u56fd\u4ea7\u5361\u4e0d\u53d7\u51fa\u53e3\u7ba1\u5236\uff0c\u53ef\u4ee5\u7a33\u5b9a\u4f9b\u8d27\u3002
\n\u653f\u7b56\u3001\u4f9b\u7ed9\u3001\u9700\u6c42\u4e09\u529b\u4ea4\u6c47\uff0c\u8ba9\u201c\u56fd\u4ea7\u66ff\u4ee3\u201d\u4e0d\u518d\u662f\u4e00\u53e5\u53e3\u53f7\uff0c\u800c\u662f\u53ef\u843d\u5730\u7684\u5de5\u7a0b\u9009\u9879\u3002<\/p>\n

02 \u661f\u5b87\u667a\u7b97\u6df7\u5408\u8282\u70b9\uff1aA100+\u6607\u817e 910B \u534f\u540c\u8bad\u7ec3<\/h3>\n

\u4f5c\u4e3a\u4e13\u6ce8 AI \u7b97\u529b\u57fa\u7840\u8bbe\u65bd\u7684\u53a6\u95e8\u661f\u5b87\u667a\u7b97\u667a\u80fd\u79d1\u6280\u6709\u9650\u516c\u53f8\uff0c\u8fc7\u53bb 18 \u4e2a\u6708\u5728 3 \u5927\u6838\u5fc3\u673a\u623f\u90e8\u7f72\u4e86\u8d85\u8fc7 6000 \u5f20 GPU\uff0c\u5176\u4e2d 35% \u4e3a\u56fd\u4ea7\u5361\u3002
\n\u6700\u65b0\u4e0a\u7ebf\u7684\u201c\u6df7\u5408 GPU \u79df\u8d41\u201d\u8282\u70b9\uff0c\u5355\u673a\u7bb1\u5185 4 \u5f20 A100 \u4e0e 4 \u5f20\u6607\u817e 910B \u901a\u8fc7 400 Gbps RoCEv2 \u7f51\u7edc\u76f4\u8fde\uff0c\u914d\u5408\u81ea\u7814\u7684 StarLink \u62d3\u6251\u611f\u77e5\u8c03\u5ea6\uff0c\u53ef\u5728\u6846\u67b6\u5c42\u81ea\u52a8\u8bc6\u522b\u7b97\u529b\u7279\u5f81\uff0c\u5b9e\u73b0\u201c\u6a21\u578b\u5e76\u884c+\u6570\u636e\u5e76\u884c\u201d\u6df7\u5408\u7b56\u7565\u3002
\n\u7528\u6237\u53ea\u9700\u5728\u5e73\u53f0\u52fe\u9009\u201cAuto-Mix\u201d\u6a21\u5f0f\uff0c\u7cfb\u7edf\u5373\u53ef\u628a\u8ba1\u7b97\u56fe\u62c6\u6210 17 \u4e2a\u7b97\u5b50\u7c07\uff1a\u5bf9 CUDA \u53cb\u597d\u90e8\u5206\u7559\u5728 A100\uff0c\u5bf9\u77e9\u9635\u4e58\u5bc6\u5ea6\u9ad8\u7684\u7c07\u52a8\u6001\u8fc1\u79fb\u5230\u6607\u817e\uff0c\u5b9e\u6d4b 1024 \u5361\u89c4\u6a21\u4e0b\u7aef\u5230\u7aef\u541e\u5410\u4e0b\u964d <4%\uff0c\u800c\u79df\u8d41\u6210\u672c\u76f4\u964d 40%\u3002<\/p>\n

03 \u6846\u67b6\u9002\u914d\uff1aPyTorch 2.2 + CANN 6.3 \u4e00\u952e\u5207\u6362<\/h3>\n

\u4f20\u7edf\u5370\u8c61\u4e2d\uff0c\u56fd\u4ea7\u5361\u201c\u9a71\u52a8\u96be\u3001\u8fc1\u79fb\u82e6\u201d\u3002\u661f\u5b87\u667a\u7b97\u628a\u6846\u67b6\u9002\u914d\u505a\u6210\u201c\u767d\u76d2\u201d\u670d\u52a1\uff1a
\n– \u955c\u50cf\u5e02\u573a\u5185\u7f6e PyTorch 2.2\u3001TensorFlow 2.15\u3001MindSpore 2.3\uff0c\u5df2\u9884\u88c5 CUDA 12.1 \u4e0e CANN 6.3 \u53cc\u6808\uff1b
\n– \u63d0\u4f9b torch.cuda<\/code> \u21c4 torch_npu<\/code> \u81ea\u52a8\u66ff\u6362\u811a\u672c\uff0c30 \u4e07\u884c\u4ee3\u7801\u4ed3\u5e93\u5e73\u5747 18 \u5206\u949f\u5b8c\u6210\u8fc1\u79fb\uff1b
\n– \u5e73\u53f0\u8d60\u9001 10 \u5143\u4f53\u9a8c\u91d1\uff0c\u65b0\u7528\u6237\u53ef 0 \u6210\u672c\u542f\u52a8 GPU\u4e91\u4e3b\u673a<\/a> \u8fdb\u884c\u9002\u914d\u9a8c\u8bc1\u3002
\n\u5982\u679c\u4f01\u4e1a\u9700\u8981\u8054\u5408\u8c03\u8bd5\uff0c\u661f\u5b87\u667a\u7b97\u7b97\u6cd5\u56e2\u961f\u63d0\u4f9b 5\u00d78 \u5c0f\u65f6\u514d\u8d39\u9002\u914d\u54a8\u8be2\uff0c\u4ece\u7b97\u5b50\u62a5\u9519\u5230\u7cbe\u5ea6\u5bf9\u9f50\u4e00\u7ad9\u5f0f\u515c\u5e95\u3002<\/p>\n

04 \u6027\u80fd\u6d4b\u8bd5\uff1a\u76f8\u540c\u6a21\u578b 1024 \u5361\u89c4\u6a21 FP16 \u541e\u5410\u5bf9\u6bd4<\/h3>\n

\u6d4b\u8bd5\u6a21\u578b\uff1aLLaMA-65B \u9884\u8bad\u7ec3\uff0c\u5e8f\u5217\u957f\u5ea6 4096\uff0c\u5168\u5c40 Batch 4M\u3002
\n\u786c\u4ef6\u914d\u6bd4\uff1aA100 \u7eaf\u5361\u96c6\u7fa4 vs \u661f\u5b87\u667a\u7b97\u6df7\u5408\u96c6\u7fa4\uff0850% A100 + 50% \u6607\u817e 910B\uff09\u3002
\n\u7ed3\u679c\uff1a
\n– A100 \u7eaf\u5361\uff1a\u6bcf\u5361\u5e73\u5747 163 token\/s\uff0c\u96c6\u7fa4\u603b\u541e\u5410 167 k token\/s\uff1b
\n– \u6df7\u5408\u96c6\u7fa4\uff1a\u6bcf\u5361\u6298\u7b97 156 token\/s\uff0c\u96c6\u7fa4\u603b\u541e\u5410 160 k token\/s\uff0c\u6027\u80fd\u635f\u5931 4.2%\uff1b
\n– \u529f\u8017\u4e0b\u964d 11%\uff0c\u5355\u5361\u79df\u8d41\u4ef7\u4e0b\u964d 40%\uff0cTCO \u4f18\u52bf\u663e\u8457\u3002
\n\u5728\u53e6\u4e00\u7ec4 CV \u591a\u6a21\u6001\u6a21\u578b ViT-g\/14 \u6d4b\u8bd5\u4e2d\uff0c\u7531\u4e8e\u5bd2\u6b66\u7eaa MLU 370 \u5bf9 INT8 \u91cf\u5316\u539f\u751f\u53cb\u597d\uff0c\u6df7\u5408\u96c6\u7fa4\u53cd\u800c\u6bd4\u7eaf A100 \u96c6\u7fa4\u9ad8\u51fa 7% \u7684\u63a8\u7406\u541e\u5410\u3002<\/p>\n

05 \u6210\u672c\u4f18\u52bf\uff1a\u56fd\u4ea7\u5361\u79df\u8d41\u4ef7\u4f4e 40%\uff0c\u957f\u671f\u5408\u540c\u518d\u964d 10%<\/h3>\n

\u82f1\u4f1f\u8fbe\u9ad8\u7aef\u5361\u88ab\u9650\u540e\uff0c\u9ed1\u5e02 A100 \u79df\u91d1\u4e00\u5ea6\u88ab\u7092\u5230 2.8 \u5143\/\u5361\u65f6\u3002\u661f\u5b87\u667a\u7b97\u4f9d\u6258\u53a6\u95e8\u3001\u829c\u6e56\u3001\u4e2d\u536b\u4e09\u5927\u81ea\u6709\u673a\u623f\uff0c\u628a\u56fd\u4ea7\u5361\u65e5\u5e38\u4ef7\u5b9a\u5728 1.2 \u5143\/\u5361\u65f6\uff0c\u6bd4\u8fdb\u53e3\u5361\u5b98\u65b9\u4ef7\u4f4e 40%\u3002
\n\u82e5\u7528\u6237\u7b7e\u7f72 6 \u4e2a\u6708\u4ee5\u4e0a GPU\u670d\u52a1\u5668\u79df\u7528<\/a> \u5408\u540c\uff0c\u53ef\u518d\u4eab 10% \u4f18\u60e0\uff0c\u5e76\u9501\u5b9a\u4f9b\u8d27\u6570\u91cf\uff0c\u907f\u514d\u201c\u4eca\u5929\u79df 50 \u5f20\uff0c\u660e\u5929\u53ea\u80fd\u62ff\u5230 30 \u5f20\u201d\u7684\u65ad\u4f9b\u98ce\u9669\u3002
\n\u5bf9\u4e8e\u8d44\u91d1\u654f\u611f\u7684\u9ad8\u6821\u4e0e\u521d\u521b\u56e2\u961f\uff0c\u5e73\u53f0\u8fd8\u63d0\u4f9b\u201c\u5148\u4f7f\u7528\u540e\u4ed8\u8d39\u201d\u6a21\u5f0f\u2014\u2014\u9879\u76ee\u9a8c\u6536\u524d\u53ea\u8ba1\u91cf\u4e0d\u7b97\u8d39\uff0c\u9a8c\u6536\u540e\u4e00\u6b21\u6027\u62b5\u6263\u4f53\u9a8c\u91d1\uff0c\u771f\u6b63\u505a\u5230\u201c\u7b97\u529b\u5148\u7528\u3001\u98ce\u9669\u6211\u62c5\u201d\u3002<\/p>\n

06 \u751f\u6001\u52a0\u6210\uff1a\u6570\u636e\u96c6\u3001\u6a21\u578b\u3001\u5e94\u7528\u4e00\u952e\u5373\u73a9<\/h3>\n

\u661f\u5b87\u667a\u7b97\u4e0d\u4ec5\u51fa\u79df\u786c\u4ef6\uff0c\u66f4\u6253\u9020\u201cAI \u5e94\u7528\u5546\u5e97\u201d\uff1a
\n– \u5185\u7f6e 1200+ \u516c\u5171\u6570\u636e\u96c6\u3001280+ \u4e3b\u6d41\u6a21\u578b\uff0c\u53ef\u76f4\u63a5\u6302\u8f7d\u5230\u5b9e\u4f8b\uff1b
\n– \u63d0\u4f9b Stable Diffusion\u3001ChatGLM3\u3001CodeLlama \u7b49 40 \u4f59\u4e2a\u70ed\u95e8 AI\u5e94\u7528<\/a> \u955c\u50cf\uff0c\u70b9\u51fb\u5373\u5f00\uff1b
\n– \u652f\u6301\u8de8\u5b9e\u4f8b\u5171\u4eab\u7684 10 TB \u6301\u4e45\u5316\u4e91\u76d8\uff0c\u8bad\u7ec3\u4e2d\u65ad\u540e\u6362\u5361\u7ee7\u7eed\u8dd1\uff0c\u6570\u636e 0 \u4e22\u5931\u3002
\n\u5f00\u53d1\u8005\u65e0\u9700\u518d\u4e3a\u201c\u88c5\u9a71\u52a8\u3001\u914d\u73af\u5883\u3001\u62f7\u6570\u636e\u201d\u6d6a\u8d39\u65f6\u95f4\uff0c\u628a\u7cbe\u529b\u6295\u5165\u5230\u6838\u5fc3\u7b97\u6cd5\u521b\u65b0\u5373\u53ef\u3002<\/p>\n

07 \u7acb\u5373\u884c\u52a8\uff1a\u514d\u8d39\u9002\u914d\u54a8\u8be2 + 10 \u5143\u4f53\u9a8c\u91d1<\/h3>\n

\u56fd\u4ea7\u66ff\u4ee3\u4e0d\u662f\u201c\u4f4e\u4ef7\u4f4e\u914d\u201d\u7684\u4ee3\u540d\u8bcd\uff0c\u800c\u662f\u201c\u9ad8\u6027\u4ef7\u6bd4+\u53ef\u6301\u7eed\u4f9b\u8d27\u201d\u7684\u65b0\u9009\u62e9\u3002
\n\u73b0\u5728\u8bbf\u95ee \u661f\u5b87\u667a\u7b97\u5b98\u7f51<\/a>\uff0c\u6ce8\u518c\u5373\u53ef\u83b7\u5f97 10 \u5143\u4f53\u9a8c\u91d1\uff0c\u53ef\u62b5\u6263 8 \u5c0f\u65f6\u6607\u817e 910B \u6216 5 \u5c0f\u65f6 A100 \u8bd5\u7528\u3002
\n\u5982\u679c\u4f60\u7684\u6a21\u578b\u8fd8\u5728\u6392\u961f\u7b49\u5361\uff0c\u4e0d\u59a8\u628a\u201c\u6df7\u5408 GPU \u79df\u8d41\u201d\u65b9\u6848\u642c\u8fdb\u5b9e\u9a8c\u5ba4\uff0c\u8ba9\u56fd\u4ea7\u7b97\u529b\u5148\u8dd1\u8d77\u6765\u3002
\n\u70b9\u51fb\u201c\u514d\u8d39\u9002\u914d\u54a8\u8be2\u201d\uff0c\u661f\u5b87\u667a\u7b97\u7b97\u6cd5\u56e2\u961f 24 \u5c0f\u65f6\u5185\u54cd\u5e94\uff0c\u966a\u4f60\u5b8c\u6210\u4ece CUDA \u5230 CANN \u7684\u6700\u540e\u4e00\u516c\u91cc\u3002 <\/p>\n

AI \u7ade\u8d5b\u7684\u4e0b\u534a\u573a\uff0c\u62fc\u7684\u4e0d\u518d\u662f\u201c\u8c01\u6709\u94b1\u4e70\u5361\u201d\uff0c\u800c\u662f\u201c\u8c01\u5148\u7528\u597d\u56fd\u4ea7\u5361\u201d\u3002\u661f\u5b87\u667a\u7b97\u613f\u505a\u90a3\u6761\u628a\u7b97\u529b\u53d8\u6210\u6c34\u7535\u7684\u7ba1\u9053\uff0c\u8ba9\u6bcf\u4e00\u4f4d\u521b\u4f5c\u8005\u90fd\u80fd\u4f4e\u6210\u672c\u3001\u9ad8\u6548\u7387\u5730\u62b5\u8fbe\u667a\u80fd\u672a\u6765\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

2024 \u5e74 10 \u6708\uff0c\u7f8e\u56fd\u518d\u6b21\u6536\u7d27\u5bf9\u534e\u9ad8\u7aef GPU \u51fa\u53e3\u7ba1…<\/p>\n","protected":false},"author":2,"featured_media":2625,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-2626","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-zixun"],"views":67,"_links":{"self":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts\/2626","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=2626"}],"version-history":[{"count":0,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts\/2626\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/media\/2625"}],"wp:attachment":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/media?parent=2626"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/categories?post=2626"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/tags?post=2626"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}