{"id":3105,"date":"2026-03-10T16:12:48","date_gmt":"2026-03-10T08:12:48","guid":{"rendered":"https:\/\/www.starverse-ai.com\/guide\/archives\/3105"},"modified":"2026-03-10T16:12:48","modified_gmt":"2026-03-10T08:12:48","slug":"%e5%9b%bd%e4%ba%a7%e5%a4%a7%e6%a8%a1%e5%9e%8b%e8%bf%81%e7%a7%bb%e5%bf%85%e8%af%bb%ef%bc%9a%e6%98%9f%e5%ae%87%e6%99%ba%e7%ae%97%e3%80%8c%e4%bf%a1%e5%88%9b%e4%b8%93%e5%8c%ba%e3%80%8d%e4%b8%80%e7%ab%99","status":"publish","type":"post","link":"https:\/\/www.starverse-ai.com\/guide\/archives\/3105","title":{"rendered":"\u56fd\u4ea7\u5927\u6a21\u578b\u8fc1\u79fb\u5fc5\u8bfb\uff1a\u661f\u5b87\u667a\u7b97\u300c\u4fe1\u521b\u4e13\u533a\u300d\u4e00\u7ad9\u5f0f GPU \u4e91\u4e3b\u673a\u5b9e\u8df5"},"content":{"rendered":"<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/www.starverse-ai.com\/guide\/wp-content\/uploads\/2026\/03\/1773130367_1f0ad7.png\" alt=\"\u56fd\u4ea7\u5927\u6a21\u578b\u8fc1\u79fb\u5fc5\u8bfb\uff1a\u661f\u5b87\u667a\u7b97\u300c\u4fe1\u521b\u4e13\u533a\u300d\u4e00\u7ad9\u5f0f GPU \u4e91\u4e3b\u673a\u5b9e\u8df5\" style=\"display:block; margin:10px auto; max-width:100%; height:auto;\" \/><\/figure>\n<blockquote>\n<p>\u201c\u51fa\u53e3\u7ba1\u5236\u518d\u5347\u7ea7\uff0c\u82f1\u4f1f\u8fbe\u9ad8\u7aef\u5361\u65ad\u4f9b\uff0c\u56fd\u4ea7 GPU \u670d\u52a1\u5668\u51fa\u8d27\u5360\u6bd4\u9996\u6b21\u7a81\u7834 30%\u3002\u201d<br \/>\n\u2014\u2014\u300a2024 Q1 \u4e2d\u56fd\u7b97\u529b\u5e02\u573a\u7b80\u62a5\u300b<\/p>\n<\/blockquote>\n<p>\u4e00\u7eb8\u7981\u4ee4\uff0c\u8ba9\u65e0\u6570\u6b63\u5728\u8bad\u7ec3\u7684 LLM \u88ab\u8feb\u6309\u4e0b\u6682\u505c\u952e\u3002CUDA \u751f\u6001\u518d\u6210\u719f\uff0c\u4e5f\u62b5\u4e0d\u8fc7\u201c\u65e0\u5361\u53ef\u7528\u201d\u7684\u5c34\u5c2c\u3002\u4e8e\u662f\uff0c\u653f\u52a1\u4e91\u3001\u91d1\u878d\u6838\u5fc3\u7cfb\u7edf\u3001\u56fd\u4ea7\u5927\u6a21\u578b\u6218\u961f\u96c6\u4f53\u628a\u76ee\u5149\u8f6c\u5411\u201c\u4fe1\u521b\u201d\u2014\u2014\u8fd9\u6761\u66fe\u88ab\u8d34\u4e0a\u201c\u6027\u80fd\u59a5\u534f\u201d\u6807\u7b7e\u7684\u8def\u7ebf\uff0c\u5982\u4eca\u6210\u4e86\u552f\u4e00\u786e\u5b9a\u6027\u7684\u7b97\u529b\u51fa\u53e3\u3002<\/p>\n<h2>\u4fe1\u521b\u4e0d\u662f\u5907\u80ce\uff0c\u662f\u65b0\u4e3b\u6218\u573a<\/h2>\n<p>\u8fc7\u53bb\u5341\u4e8c\u4e2a\u6708\uff0c\u6607\u817e 910B\u3001\u5bd2\u6b66\u7eaa MLU370\u3001\u6d77\u5149 DCU \u7684\u8fed\u4ee3\u901f\u5ea6\u51e0\u4e4e\u8ffd\u5e73\u539f\u5382\u8282\u594f\uff1a<br \/>\n&#8211; \u6607\u817e 910B \u5185\u5b58\u5e26\u5bbd\u63d0\u5347 62%\uff0cFP16 \u7b97\u529b 320 TFLOPS\uff1b<br \/>\n&#8211; \u5bd2\u6b66\u7eaa 370 \u652f\u6301 48 GB \u663e\u5b58\uff0c\u5355\u5361\u53ef\u541e\u4e0b 70 B \u53c2\u6570\u63a8\u7406\uff1b<br \/>\n&#8211; \u6d77\u5149 DCU \u517c\u5bb9 ROCm\uff0cKernel \u7ea7\u8fc1\u79fb\u53ea\u9700\u91cd\u7f16\u8bd1\u3002  <\/p>\n<p>\u66f4\u5173\u952e\u7684\u662f\uff0c\u56fd\u4ea7\u82af\u7247+\u56fd\u4ea7\u6846\u67b6\u7684\u7ec4\u5408\uff0c\u653f\u7b56\u8865\u8d34\u76f4\u63a5\u6298\u7b97\u6210\u79df\u91d1\u2014\u2014<strong>\u540c\u7b49\u7cbe\u5ea6\u4e0b\uff0c\u8bad\u7ec3\u6210\u672c\u964d\u5230 A100 \u7684 40%<\/strong>\uff0c\u518d\u52a0 10% \u7684\u4fe1\u521b\u4e0a\u4e91\u8865\u8d34\uff0c\u4ef7\u683c\u4f18\u52bf\u77ac\u95f4\u62c9\u6ee1\u3002<\/p>\n<h2>\u661f\u5b87\u667a\u7b97\u300c\u4fe1\u521b\u4e13\u533a\u300d\uff1a\u628a\u8fc1\u79fb\u505a\u6210\u201c\u4e00\u952e\u590d\u5236\u201d<\/h2>\n<p>\u5f53\u5e02\u573a\u8fd8\u5728\u4e89\u8bba\u201c\u56fd\u4ea7\u5361\u80fd\u4e0d\u80fd\u6253\u201d\u65f6\uff0c<a href=\"https:\/\/www.starverse-ai.com\">\u661f\u5b87\u667a\u7b97<\/a> \u5df2\u7ecf\u628a\u6607\u817e\u3001\u5bd2\u6b66\u7eaa\u3001\u6d77\u5149\u4e09\u5927\u673a\u578b\u642c\u4e0a\u4e86\u540c\u4e00\u6735\u4e91\uff0c\u5e76\u7ed9\u51fa\u4e00\u6761<strong>\u201cCUDA\u2192MindSpore\u201d3 \u6b65\u8def\u5f84<\/strong>\uff0c\u8ba9\u5f00\u53d1\u8005\u50cf\u6362\u663e\u5361\u4e00\u6837\u6362\u751f\u6001\uff1a<\/p>\n<ol>\n<li><strong>\u955c\u50cf\u5c42<\/strong>\uff1a\u5185\u7f6e CANN 5.1\u3001MindSpore 2.3\u3001Paddle-XTCL \u53cc\u6808\uff0c\u62c9\u53d6\u5373\u7528\uff1b  <\/li>\n<li><strong>\u7b97\u5b50\u5c42<\/strong>\uff1a30+ \u9ad8\u9891 Transformer \u7b97\u5b50\u5df2\u5b8c\u6210\u878d\u5408\u8fc1\u79fb\uff0cGLM\u3001Baichuan\u3001Qwen \u7cfb\u5217\u6a21\u578b<strong>\u8bad\u7ec3\u7cbe\u5ea6\u8bef\u5dee &lt; 0.8%<\/strong>\uff1b  <\/li>\n<li><strong>\u6570\u636e\u5c42<\/strong>\uff1a\u4e91\u786c\u76d8\u3001\u4e91\u5b58\u50a8\u3001<a href=\"https:\/\/www.starverse-ai.com\/node\/019b88ac-286a-70a3-bafa-cfa47c851b4d\">\u516c\u5171\u6a21\u578b\u5e93<\/a> \u4e09\u76d8\u8054\u52a8\uff0cTB \u7ea7\u6743\u91cd\u6587\u4ef6\u79d2\u7ea7\u6302\u8f7d\uff0c0 \u6539\u52a8\u7ee7\u7eed\u8bad\u7ec3\u3002<\/li>\n<\/ol>\n<p>\u4ece CUDA \u5230\u56fd\u4ea7\u6846\u67b6\uff0c<strong>\u5e73\u5747\u8fc1\u79fb\u5de5\u65f6\u4ece 4 \u5468\u538b\u7f29\u5230 3 \u5929<\/strong>\uff0c\u771f\u6b63\u505a\u5230\u201c\u4ee3\u7801\u4e0d\u52a8\uff0c\u6027\u80fd\u4e0d\u964d\u201d\u3002<\/p>\n<h2>30+ \u653f\u52a1\u3001\u91d1\u878d\u5ba2\u6237\u5b9e\u6218\u590d\u76d8<\/h2>\n<ul>\n<li>\u67d0\u7701\u653f\u52a1\u4e91\u77e5\u8bc6\u56fe\u8c31\u9879\u76ee\uff1a\u539f 64 \u5361 A100 \u96c6\u7fa4\uff0c\u6362\u7528 80 \u5361\u6607\u817e 910B\uff0c<strong>\u8bad\u7ec3\u65f6\u957f\u6301\u5e73\uff0c\u79df\u91d1\u4e0b\u964d 58%<\/strong>\uff1b  <\/li>\n<li>\u80a1\u4efd\u5236\u94f6\u884c\u53cd\u6b3a\u8bc8\u6a21\u578b\uff1a\u5bd2\u6b66\u7eaa 370 \u4e09\u8282\u70b9 FP16 \u63a8\u7406\uff0c<strong>\u5ef6\u8fdf 38 ms<\/strong>\uff0c\u6bd4 T4 \u964d\u4f4e 21%\uff0c\u5408\u89c4\u52a0\u5206\u9879\u76f4\u63a5\u5199\u5165\u62db\u6807\u8bc4\u5206\uff1b  <\/li>\n<li>\u56fd\u4ea7\u5927\u6a21\u578b\u521b\u4e1a\u56e2\u961f\uff1a\u5728\u6d77\u5149 DCU \u4e0a\u5b8c\u6210 176 B \u53c2\u6570\u9884\u8bad\u7ec3\uff0c<strong>MFU \u8dd1\u5230 62%<\/strong>\uff0c\u4e0e A100 \u5dee\u8ddd\u7f29\u5c0f\u81f3 8%\u3002  <\/li>\n<\/ul>\n<p>\u8fd9\u4e9b\u6848\u4f8b\u5168\u90e8\u6c89\u6dc0\u5728<a href=\"https:\/\/www.starverse-ai.com\">\u661f\u5b87\u667a\u7b97 GPU\u4e91\u4e3b\u673a<\/a>\u300c\u4fe1\u521b\u4e13\u533a\u300d\u955c\u50cf\u5e02\u573a\uff0c\u65b0\u7528\u6237\u6ce8\u518c\u5373\u53ef\u9886\u53d6 <strong>10 \u5143\u4f53\u9a8c\u91d1<\/strong>\uff0c\u4e00\u952e\u514b\u9686\uff0c<strong>10 \u5206\u949f\u8dd1\u51fa\u7b2c\u4e00\u4e2a\u6536\u655b\u66f2\u7ebf<\/strong>\u3002<\/p>\n<h2>\u6027\u80fd\u57fa\u51c6\uff1a\u6211\u4eec\u6562\u7528\u6570\u636e\u8bf4\u8bdd<\/h2>\n<p>\u5728 70 B \u53c2\u6570\u30011.2 T token \u7684\u6807\u51c6 GLM \u9884\u8bad\u7ec3\u4efb\u52a1\u4e2d\uff0c\u540c\u6837\u4fdd\u6301 DP=128\u3001TP=8 \u7684\u5e76\u884c\u7b56\u7565\uff1a<br \/>\n&#8211; A100 80 GB SXM\uff1a\u541e\u5410 2030 token\/s\uff1b<br \/>\n&#8211; \u6607\u817e 910B 32 GB\uff1a\u541e\u5410 1680 token\/s\uff0c<strong>\u8fbe\u5230 A100 \u7684 82.7%<\/strong>\uff1b<br \/>\n&#8211; \u5bd2\u6b66\u7eaa 370 48 GB\uff1a\u541e\u5410 1590 token\/s\uff0c<strong>\u8fbe\u5230 A100 \u7684 78.3%<\/strong>\uff1b<br \/>\n&#8211; \u6d77\u5149 DCU 32 GB\uff1a\u541e\u5410 1610 token\/s\uff0c<strong>\u8fbe\u5230 A100 \u7684 79.4%<\/strong>\u3002  <\/p>\n<p>\u82e5\u628a\u79df\u91d1\u56e0\u7d20\u6298\u7b97\u8fdb\u6765\uff0c<strong>\u6bcf\u4ebf token \u6210\u672c\u4ec5\u4e3a A100 \u7684 35%\uff5e42%<\/strong>\uff0c\u771f\u6b63\u505a\u5230\u201c\u82b1\u5c0f\u94b1\uff0c\u529e\u5927\u4e8b\u201d\u3002<\/p>\n<h2>\u4ef7\u683c\u5bf9\u6bd4\uff1a\u9ad8\u7aef\u5361\u79df\u91d1\u88ab\u6253\u5230 4 \u6298<\/h2>\n<p>\u4ee5 8 \u5361\u5e76\u884c\u3001\u79df\u7528 30 \u5929\u4e3a\u4f8b\uff1a<br \/>\n&#8211; A100 80 GB \u5e02\u573a\u5747\u4ef7\uff1a\u00a54.8 \/ \u5361\/\u65f6\uff0c\u603b\u4ef7 \u00a527.6 \u4e07\uff1b<br \/>\n&#8211; \u6607\u817e 910B\uff1a<a href=\"https:\/\/www.starverse-ai.com\">\u661f\u5b87\u667a\u7b97 GPU\u670d\u52a1\u5668\u79df\u7528<\/a> \u4ef7 \u00a51.9 \/ \u5361\/\u65f6\uff0c\u4fe1\u521b\u8865\u8d34\u540e\u518d\u51cf 10%\uff0c<strong>\u5b9e\u4ed8 \u00a59.8 \u4e07<\/strong>\uff1b<br \/>\n&#8211; \u5bd2\u6b66\u7eaa 370\uff1a\u00a51.7 \/ \u5361\/\u65f6\uff0c\u8865\u8d34\u540e <strong>\u00a58.9 \u4e07<\/strong>\uff1b<br \/>\n&#8211; \u6d77\u5149 DCU\uff1a\u00a51.6 \/ \u5361\/\u65f6\uff0c\u8865\u8d34\u540e <strong>\u00a58.3 \u4e07<\/strong>\u3002  <\/p>\n<p><strong>\u4e0d\u5230 10 \u4e07\u5143\uff0c\u5373\u53ef\u8dd1\u5b8c 176 B \u53c2\u6570\u5168\u6d41\u7a0b<\/strong>\uff0c\u9884\u7b97\u76f4\u63a5\u780d\u51fa\u4e24\u53f0\u9ad8\u7aef\u5de5\u4f5c\u7ad9\u3002<\/p>\n<h2>\u5f00\u53d1\u8005\u4f53\u9a8c\uff1a\u628a\u201c\u8fd0\u7ef4\u201d\u4ea4\u7ed9\u4e91\uff0c\u628a\u201c\u521b\u65b0\u201d\u7559\u7ed9\u81ea\u5df1<\/h2>\n<p>\u661f\u5b87\u667a\u7b97\u56f4\u7ed5 AI \u5de5\u4f5c\u6d41\u505a\u4e86\u4e09\u4ef6\u4e8b\uff1a<br \/>\n1. <strong>\u6570\u636e\u4e0d\u52a8<\/strong>\uff1a\u4e91\u786c\u76d8\u652f\u6301\u8de8\u5b9e\u4f8b\u70ed\u63d2\u62d4\uff0c<a href=\"https:\/\/www.starverse-ai.com\/node\/019b88ac-0730-7451-a8ab-9c3c873fef42\">\u4e91\u5b58\u50a8<\/a> \u63d0\u4f9b 10 Gb\/s \u5185\u7f51\u4e0a\u4f20\uff0c\u6570\u636e\u96c6\u4e00\u6b21\u4e0a\u4f20\uff0c\u5168\u751f\u547d\u5468\u671f\u590d\u7528\uff1b<br \/>\n2. <strong>\u6a21\u578b\u79d2\u8d77<\/strong>\uff1a\u5185\u7f6e 5000+ \u516c\u5171\u6a21\u578b\u4e0e\u6570\u636e\u96c6\uff0cTensorRT-llm\u3001vLLM\u3001MindFormers \u4e00\u952e\u5207\u6362\uff0c<strong>\u63a8\u7406\u6a21\u677f 30 \u79d2\u53d1\u5e03\u6210 API<\/strong>\uff1b<br \/>\n3. <strong>\u8ba1\u8d39\u7075\u6d3b<\/strong>\uff1a\u6309\u5c0f\u65f6\u3001\u6309\u5929\u3001\u6309\u6708\u4e09\u79cd\u6a21\u5f0f\uff0cGPU\u4e91\u4e3b\u673a \u968f\u5f00\u968f\u505c\uff0c\u8bad\u7ec3\u4efb\u52a1\u4e2d\u65ad\u81ea\u52a8\u4fdd\u5b58 Checkpoint\uff0c<strong>\u7a7a\u95f2\u4e0d\u8ba1\u8d39<\/strong>\uff0c\u771f\u6b63\u505a\u5230\u201c\u7b97\u529b\u50cf\u6c34\u7535\u4e00\u6837\u7528\u591a\u5c11\u4ed8\u591a\u5c11\u201d\u3002<\/p>\n<h2>\u5199\u5728\u6700\u540e\uff1a\u4fe1\u521b\u4e0d\u662f\u9009\u62e9\u9898\uff0c\u662f\u5fc5\u7b54\u9898<\/h2>\n<p>\u51fa\u53e3\u7ba1\u5236\u7684\u9774\u5b50\u5df2\u7ecf\u843d\u5730\uff0c\u56fd\u4ea7\u82af\u7247\u7684\u8fed\u4ee3\u901f\u5ea6\u5374\u8d8a\u6765\u8d8a\u5feb\u3002\u5bf9\u5f00\u53d1\u8005\u800c\u8a00\uff0c<strong>\u4e0e\u5176\u7b49\u5f85\u9ad8\u7aef\u5361\u89e3\u7981\uff0c\u4e0d\u5982\u5148\u628a\u6a21\u578b\u8dd1\u5728\u786e\u5b9a\u7684\u7b97\u529b\u4e0a<\/strong>\u3002\u661f\u5b87\u667a\u7b97\u300c\u4fe1\u521b\u4e13\u533a\u300d\u7528 30+ \u771f\u5b9e\u6848\u4f8b\u8bc1\u660e\uff1a\u8fc1\u79fb\u95e8\u69db\u53ef\u4ee5\u5f88\u4f4e\uff0c\u6027\u80fd\u635f\u5931\u53ef\u4ee5\u5f88\u5c0f\uff0c\u9884\u7b97\u538b\u529b\u53ef\u4ee5\u8f7b\u5230\u5ffd\u7565\u4e0d\u8ba1\u3002<\/p>\n<p>\u73b0\u5728\u6ce8\u518c <a href=\"https:\/\/www.starverse-ai.com\">\u661f\u5b87\u667a\u7b97<\/a>\uff0c\u5373\u53ef\u9886\u53d6 <strong>10 \u5143\u4f53\u9a8c\u91d1<\/strong>\uff0c0 \u6210\u672c\u8bd5\u8dd1\u6607\u817e\u3001\u5bd2\u6b66\u7eaa\u3001\u6d77\u5149\u5168\u7cfb\u5217 GPU\u4e91\u4e3b\u673a\u3002<br \/>\n\u628a\u98ce\u9669\u7559\u7ed9\u6628\u5929\uff0c\u628a\u786e\u5b9a\u6027\u63e1\u5728\u624b\u4e2d\u2014\u2014<strong>\u56fd\u4ea7\u5927\u6a21\u578b\u7684\u4e0b\u4e00\u7ad9\uff0c\u4ece\u661f\u5b87\u667a\u7b97\u5f00\u59cb<\/strong>\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u201c\u51fa\u53e3\u7ba1\u5236\u518d\u5347\u7ea7\uff0c\u82f1\u4f1f\u8fbe\u9ad8\u7aef\u5361\u65ad\u4f9b\uff0c\u56fd\u4ea7 GPU \u670d\u52a1\u5668\u51fa\u8d27&hellip;<\/p>\n","protected":false},"author":2,"featured_media":3104,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-3105","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-zixun"],"views":76,"_links":{"self":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts\/3105","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=3105"}],"version-history":[{"count":0,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts\/3105\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/media\/3104"}],"wp:attachment":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/media?parent=3105"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/categories?post=3105"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/tags?post=3105"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}