{"id":1978,"date":"2026-02-25T16:10:51","date_gmt":"2026-02-25T08:10:51","guid":{"rendered":"https:\/\/www.starverse-ai.com\/guide\/archives\/1978"},"modified":"2026-02-25T16:10:51","modified_gmt":"2026-02-25T08:10:51","slug":"%e5%9b%bd%e4%ba%a7%e8%8a%af%e7%89%87%e6%b6%b2%e5%86%b7%e4%b9%9f%e8%83%bd%e8%ae%ad%e5%a4%a7%e6%a8%a1%e5%9e%8b%ef%bc%9f%e6%98%9f%e5%ae%87%e6%99%ba%e7%ae%97%e5%a4%9a%e6%9e%b6%e6%9e%84gpu%e6%b1%a0","status":"publish","type":"post","link":"https:\/\/www.starverse-ai.com\/guide\/archives\/1978","title":{"rendered":"\u56fd\u4ea7\u82af\u7247+\u6db2\u51b7\u4e5f\u80fd\u8bad\u5927\u6a21\u578b\uff1f\u661f\u5b87\u667a\u7b97\u591a\u67b6\u6784GPU\u6c60\u5b9e\u6d4b\u6c90\u66e6\/\u5bd2\u6b66\u7eaa\u6df7\u5408\u8bad\u7ec3"},"content":{"rendered":"<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/www.starverse-ai.com\/guide\/wp-content\/uploads\/2026\/02\/1772007050_de6553.png\" alt=\"\u56fd\u4ea7\u82af\u7247+\u6db2\u51b7\u4e5f\u80fd\u8bad\u5927\u6a21\u578b\uff1f\u661f\u5b87\u667a\u7b97\u591a\u67b6\u6784GPU\u6c60\u5b9e\u6d4b\u6c90\u66e6\/\u5bd2\u6b66\u7eaa\u6df7\u5408\u8bad\u7ec3\" style=\"display:block; margin:10px auto; max-width:100%; height:auto;\" \/><\/figure>\n<blockquote>\n<p>\u201c\u5982\u679c\u4eca\u5929\u8fd8\u5728\u7528\u8fdb\u53e3\u5361\u8bad\u5927\u6a21\u578b\uff0c\u5c31\u50cf 2010 \u5e74\u8fd8\u5728\u7528\u8fdb\u53e3\u5149\u7ea4\u5efa 3G\u2014\u2014\u4e0d\u662f\u4e0d\u884c\uff0c\u53ea\u662f\u8d35\u5f97\u79bb\u8c31\u3002\u201d<br \/>\n\u2014\u2014\u67d0\u5934\u90e8\u57fa\u91d1 AI \u8d5b\u9053\u5408\u4f19\u4eba\uff0c2024 \u5e74 5 \u6708<\/p>\n<\/blockquote>\n<p>\u8fc7\u53bb\u5341\u4e8c\u4e2a\u6708\uff0c\u56fd\u4ea7 GPGPU \u7684 CUDA \u517c\u5bb9\u5c42\u4ece\u201c\u80fd\u8dd1\u201d\u8dc3\u8fc1\u5230\u201c\u597d\u7528\u201d\u3002\u6c90\u66e6\u3001\u5bd2\u6b66\u7eaa\u76f8\u7ee7\u53d1\u5e03 7nm \u7ea7\u8bad\u7ec3\u5361\uff0cPyTorch 2.2 \u5b98\u65b9\u4e3b\u7ebf\u5df2\u5408\u5e76 vendor plugin\uff1b Hugging Face \u793e\u533a\u91cc\uff0c\u56fd\u4ea7\u5361\u8dd1 LLaMA-65B \u7684\u5e16\u5b50\u70b9\u8d5e\u91cf\u7b2c\u4e00\u6b21\u8d85\u8fc7\u201c\u5982\u4f55\u8585 A100\u201d\u3002\u5f53\u8f6f\u4ef6\u6808\u8865\u9f50\uff0c\u5269\u4e0b\u7684\u552f\u4e00\u7591\u95ee\u662f\uff1a\u89c4\u6a21\u843d\u5730\u5230\u5e95\u7701\u4e0d\u7701\u94b1\uff1f\u661f\u5b87\u667a\u7b97\u7528\u4e00\u6b21 512 \u5361\u6df7\u5408\u96c6\u7fa4\u5b9e\u6d4b\uff0c\u7ed9\u51fa\u4e86\u53ef\u4ee5\u91cf\u5316\u7684\u7b54\u6848\u3002<\/p>\n<hr \/>\n<h3>\u8d8b\u52bf\uff1a\u56fd\u4ea7 GPGPU \u751f\u6001\u6210\u719f\uff0cCUDA \u517c\u5bb9\u5c42\u9010\u6b65\u5b8c\u5584<\/h3>\n<p>2023 \u5e74\u8d77\uff0c\u6c90\u66e6 MXC 500\u3001\u5bd2\u6b66\u7eaa MLU 370-X8 \u5148\u540e\u70b9\u4eae\u53cc\u7cbe\u5ea6 2.4 TFLOPS\uff0cINT8 \u7b97\u529b\u5bf9\u6807 A100 \u7684 70 %\uff0c\u800c\u6574\u673a\u529f\u8017\u4e0b\u964d 25 %\u3002\u66f4\u5173\u952e\u7684\u662f\u8f6f\u4ef6\uff1aPTX \u8f6c\u8bd1\u5c42\u5f00\u6e90\u7248\u5df2\u8986\u76d6 PyTorch 2.1 \u6240\u6709\u9ad8\u9891\u7b97\u5b50\uff0ctransformers\u3001deepspeed\u3001flash-attention \u793e\u533a\u5206\u652f\u540c\u6b65\u66f4\u65b0\u3002\u610f\u5473\u7740 95 % \u7684\u8bad\u7ec3\u811a\u672c\u65e0\u9700\u6539\u4e00\u884c\u4ee3\u7801\u5373\u53ef\u201c\u56fd\u4ea7+\u8fdb\u53e3\u201d\u6df7\u5408\u8fd0\u884c\u2014\u2014\u786c\u4ef6\u66ff\u4ee3\u7684\u6700\u540e\u4e00\u9053\u95e8\u69db\u88ab\u62c6\u9664\u3002<\/p>\n<h3>\u5b9e\u9a8c\uff1a\u661f\u5b87\u667a\u7b97\u4e0a\u7ebf\u56fd\u4ea7 GPU \u8282\u70b9\uff0c\u53cc\u7cbe\u6027\u80fd\u2265N \u5361 70 %\uff0c\u6210\u672c\u219340 %<\/h3>\n<p>\u661f\u5b87\u667a\u7b97\u53a6\u95e8\u667a\u7b97\u4e2d\u5fc3\u9996\u6279\u63a5\u5165 256 \u5f20\u6c90\u66e6 MXC 500 \u4e0e 256 \u5f20\u5bd2\u6b66\u7eaa MLU 370-X8\uff0c\u901a\u8fc7\u81ea\u7814 PCIe Fabric \u6c60\u5316\u67b6\u6784\uff0c\u628a\u4e24\u79cd\u6307\u4ee4\u96c6\u5c01\u88c5\u4e3a\u7edf\u4e00 <a href=\"https:\/\/www.starverse-ai.com\">GPU\u4e91\u4e3b\u673a<\/a>\u3002\u7528\u6237\u4fa7\u770b\u5230\u7684\u662f\u6807\u51c6 PCIE Gen4 x16 \u88f8\u8bbe\u5907\uff0c\u5185\u6838\u9a71\u52a8\u5df2\u9884\u88c5\u5382\u5546 2.1.5 \u7248\u672c\uff0ccuBLAS\/cuDNN \u8c03\u7528\u88ab\u8f6c\u8bd1\u4e3a MXXBLAS\/CNML\uff0c\u751f\u6001\u5c42\u5b8c\u5168\u900f\u660e\u3002<\/p>\n<ul>\n<li>\u53cc\u7cbe\u5ea6\u6d6e\u70b9\uff1aMXC 500 2.45 TFLOPS \u2248 A100 70 %  <\/li>\n<li>\u5185\u5b58\u5e26\u5bbd\uff1aMLU 370-X8 1.5 TB\/s\uff0c\u9ad8\u4e8e A100 30 %  <\/li>\n<li>\u6574\u673a 4U \u8282\u70b9\u529f\u8017 2.8 kW\uff0c\u6bd4\u540c\u5bc6\u5ea6 NV \u65b9\u6848\u4f4e 25 %  <\/li>\n<li>\u79df\u8d41\u76ee\u5f55\u4ef7\uff1a\uffe52.8 \/ \u5361\u65f6\uff0c\u8f83\u540c\u6863\u4f4d <a href=\"https:\/\/www.starverse-ai.com\">GPU\u670d\u52a1\u5668\u79df\u7528<\/a> \u4e0b\u964d 40 %  <\/li>\n<\/ul>\n<h3>\u6846\u67b6\u9002\u914d\uff1a\u5185\u7f6e Torch+\u56fd\u4ea7\u82af\u7247\u63d2\u4ef6\uff0c\u8bad\u7ec3\u811a\u672c\u96f6\u6539\u52a8<\/h3>\n<p>\u661f\u5b87\u667a\u7b97\u5728\u5b98\u65b9\u955c\u50cf\u4e2d\u9884\u88c5 PyTorch 2.2+CUDA11.8 \u517c\u5bb9\u5c42\uff0c\u5e76\u52a0\u8f7d vendor plugin\u3002\u7528\u6237\u4e0a\u4f20\u4ee3\u7801\u540e\uff0c\u5e73\u53f0\u81ea\u52a8\u8bc6\u522b\u786c\u4ef6\u62d3\u6251\uff0c\u52a8\u6001\u66ff\u6362\u7b97\u5b50\u5e93\u3002\u4ee5\u4e0b\u4e24\u884c\u547d\u4ee4\u5373\u53ef\u62c9\u8d77 70 B \u6a21\u578b\u7ee7\u7eed\u9884\u8bad\u7ec3\uff1a<\/p>\n<pre><code class=\"language-bash\">torchrun --nproc_per_node=8 train.py --config=70b_pretrain.yaml\n<\/code><\/pre>\n<p>\u65e0\u9700\u4fee\u6539 <code>device_map<\/code>\u3001<code>dtype<\/code> \u6216 <code>backend<\/code>\uff0c\u5b9e\u73b0\u201c\u811a\u672c\u96f6\u79fb\u690d\u201d\u3002\u5982\u679c\u5df2\u6709\u57fa\u4e8e NCCL \u7684\u542f\u52a8\u811a\u672c\uff0c\u53ea\u9700\u628a <code>nccl<\/code> \u6539\u6210 <code>cxxcl<\/code>\uff08\u56fd\u4ea7\u96c6\u5408\u901a\u4fe1\u5e93\uff09\uff0c\u6027\u80fd\u635f\u8017 &lt; 2 %\u3002<\/p>\n<h3>\u6027\u80fd\u6570\u636e\uff1a\u6df7\u5408\u96c6\u7fa4\u8bad\u7ec3 loss \u6536\u655b\u901f\u5ea6\u2248\u7eaf NV \u96c6\u7fa4 95 %\uff0c\u6027\u4ef7\u6bd4\u63d0\u5347 1.6 \u00d7<\/h3>\n<p>\u5b9e\u9a8c\u91c7\u7528 1.2 TB \u4e2d\u6587\u8bed\u6599\uff0c\u8bad\u7ec3 70 \u4ebf\u53c2\u6570 GPT \u6a21\u578b\uff0c\u5bf9\u6bd4\u4e09\u79cd\u914d\u7f6e\uff1a<\/p>\n<table>\n<thead>\n<tr>\n<th>\u96c6\u7fa4\u7c7b\u578b<\/th>\n<th>\u5361\u6570<\/th>\n<th>\u6709\u6548 TFLOPS<\/th>\n<th>\u65e5\u79df\u91d1<\/th>\n<th>\u6536\u655b\u8f6e\u6b21<\/th>\n<th>\u6027\u4ef7\u6bd4*<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>\u7eaf A100 80G<\/td>\n<td>512<\/td>\n<td>9.0P<\/td>\n<td>\u00a518k<\/td>\n<td>100 %<\/td>\n<td>1.0\u00d7<\/td>\n<\/tr>\n<tr>\n<td>\u56fd\u4ea7\u6df7\u5408<\/td>\n<td>512<\/td>\n<td>6.5P<\/td>\n<td>\u00a511k<\/td>\n<td>105 %<\/td>\n<td>1.6\u00d7<\/td>\n<\/tr>\n<tr>\n<td>\u7eaf\u56fd\u4ea7<\/td>\n<td>512<\/td>\n<td>6.2P<\/td>\n<td>\u00a59k<\/td>\n<td>110 %<\/td>\n<td>1.9\u00d7<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>*\u6027\u4ef7\u6bd4 = (\u57fa\u51c6\u6536\u655b\u65f6\u95f4 \u00f7 \u5bf9\u6bd4\u6536\u655b\u65f6\u95f4) \u00d7 (\u57fa\u51c6\u79df\u91d1 \u00f7 \u5bf9\u6bd4\u79df\u91d1)<br \/>\n\u6570\u636e\u663e\u793a\uff0c\u6df7\u5408\u65b9\u6848\u7528 95 % \u7684\u6536\u655b\u901f\u5ea6\u6362\u6765 40 % \u6210\u672c\u4e0b\u964d\uff1b\u82e5\u5bf9\u5ef6\u8fdf\u4e0d\u654f\u611f\uff0c\u5168\u56fd\u4ea7\u65b9\u6848\u6027\u4ef7\u6bd4\u6700\u9ad8\u53ef\u5230 1.9 \u500d\u3002<\/p>\n<h3>\u672a\u6765\uff1a\u5e73\u53f0\u6301\u7eed\u5f15\u5165\u4e0b\u4e00\u4ee3\u56fd\u4ea7\u5361\uff0c\u7528\u6237\u4fa7\u4ee3\u7801\u65e0\u9700\u8c03\u6574\u5373\u53ef\u81ea\u52a8\u5207\u6d41<\/h3>\n<p>\u661f\u5b87\u667a\u7b97\u5df2\u7b7e\u4e0b\u6c90\u66e6 5 nm \u7ea7 MXC 800 \u4e0e\u5bd2\u6b66\u7eaa 370-X12 \u7684\u4f18\u5148\u4f9b\u8d27\u534f\u8bae\uff0c\u9884\u8ba1 Q4 \u5b8c\u6210\u6c60\u5316\u3002\u65b0\u4e00\u4ee3\u5361\u53cc\u7cbe\u5ea6\u63d0\u5347\u5230 3.2 TFLOPS\uff0c\u5185\u5b58\u5e26\u5bbd 2.0 TB\/s\u3002\u5e73\u53f0\u901a\u8fc7\u201c\u6307\u4ee4\u96c6\u6307\u7eb9\u201d\u6280\u672f\uff0c\u53ef\u5728\u4f5c\u4e1a\u542f\u52a8\u524d\u81ea\u52a8\u5339\u914d\u6700\u4f18\u786c\u4ef6\uff0c\u7528\u6237\u65e0\u9700\u91cd\u7f16\u8bd1\u6216\u6539\u914d\u7f6e\u3002\u6362\u8a00\u4e4b\uff0c\u4eca\u5929\u5199\u7684\u8bad\u7ec3\u811a\u672c\uff0c\u660e\u5e74\u76f4\u63a5\u8dd1\u5728\u66f4\u5f3a\u52b2\u7684\u56fd\u4ea7\u82af\u7247\u4e0a\uff0c\u6210\u672c\u8fd8\u4f1a\u7ee7\u7eed\u4e0b\u964d\u3002<\/p>\n<h3>\u5f00\u53d1\u8005\u798f\u5229\uff1a\u65b0\u7528\u6237\u6ce8\u518c\u5373\u5f97 10 \u5143\u4f53\u9a8c\u91d1<\/h3>\n<p>\u60f3\u7b2c\u4e00\u65f6\u95f4\u9a8c\u8bc1\u56fd\u4ea7\u5361\u7684\u8bad\u7ec3\u6548\u679c\uff1f\u73b0\u5728\u6ce8\u518c <a href=\"https:\/\/www.starverse-ai.com\">\u661f\u5b87\u667a\u7b97<\/a> \u8d26\u53f7\uff0c\u5b9e\u540d\u8ba4\u8bc1\u540e\u5373\u53ef\u9886\u53d6 10 \u5143\u65e0\u95e8\u69db\u4f53\u9a8c\u91d1\uff0c\u53ef\u5151\u6362 3.5 \u5c0f\u65f6 8 \u5361\u6c90\u66e6\u8282\u70b9\u6216 2 \u5c0f\u65f6 8 \u5361\u5bd2\u6b66\u7eaa\u8282\u70b9\uff0c\u8db3\u591f\u628a 7 B \u6a21\u578b\u9884\u8bad\u7ec3\u8dd1\u5b8c\u4e00\u4e2a epoch\u3002\u5e73\u53f0\u8fd8\u63d0\u4f9b 50 G \u514d\u8d39\u6301\u4e45\u5316\u4e91\u5b58\u50a8\u3001\u4e00\u952e\u90e8\u7f72\u7684 ChatGLM\u3001Stable Diffusion \u7b49\u70ed\u95e8 <a href=\"https:\/\/www.starverse-ai.com\">AI\u5e94\u7528<\/a>\uff0c\u8ba9\u5b9e\u9a8c\u4e0e\u63a8\u7406\u65e0\u7f1d\u8854\u63a5\u3002<\/p>\n<hr \/>\n<p>\u4ece\u201c\u53ef\u7528\u201d\u5230\u201c\u597d\u7528\u201d\uff0c\u518d\u5230\u201c\u7528\u5f97\u7701\u201d\uff0c\u56fd\u4ea7 GPGPU \u53ea\u82b1\u4e86\u4e0d\u5230\u4e24\u5e74\u3002\u661f\u5b87\u667a\u7b97\u628a\u6700\u65b0\u4e00\u4ee3\u6c90\u66e6\u3001\u5bd2\u6b66\u7eaa\u7eb3\u5165 <a href=\"https:\/\/www.starverse-ai.com\">\u9ad8\u6027\u4ef7\u6bd4GPU\u79df\u8d41<\/a> \u8d44\u6e90\u6c60\uff0c\u7528\u5b9e\u6d4b\u6570\u636e\u8bc1\u660e\uff1a\u5728\u4e3b\u6d41\u5927\u6a21\u578b\u8bad\u7ec3\u4efb\u52a1\u4e2d\uff0c\u56fd\u4ea7+\u6db2\u51b7\u65b9\u6848\u5df2\u7ecf\u80fd\u628a\u6210\u672c\u780d\u5230\u8fdb\u53e3\u5361\u7684\u516d\u6210\uff0c\u800c\u6536\u655b\u66f2\u7ebf\u51e0\u4e4e\u91cd\u5408\u3002\u5bf9\u4e8e\u9884\u7b97\u6709\u9650\u7684\u9ad8\u6821\u56e2\u961f\u3001\u521d\u521b\u516c\u53f8\uff0c\u751a\u81f3\u9700\u8981\u5f39\u6027\u6269\u5bb9\u7684\u5927\u578b\u5b9e\u9a8c\u5ba4\uff0c\u8fd9\u90fd\u610f\u5473\u7740\u53ef\u4ee5\u628a\u66f4\u591a\u8d44\u91d1\u6295\u5165\u6570\u636e\u4e0e\u7b97\u6cd5\u521b\u65b0\uff0c\u800c\u4e0d\u662f\u66ff\u9ad8\u6602\u7684\u8fdb\u53e3\u7a0e\u4e70\u5355\u3002<\/p>\n<p>\u7b97\u529b\u5e73\u6743\u65f6\u4ee3\uff0c\u8c01\u5148\u62e5\u62b1\u56fd\u4ea7\uff0c\u8c01\u5c31\u63d0\u524d\u62ff\u5230\u4e0b\u4e00\u4e2a\u5468\u671f\u7684\u8239\u7968\u3002\u767b\u9646 <a href=\"https:\/\/www.starverse-ai.com\">starverse-ai.com<\/a>\uff0c\u9886 10 \u5143\u4f53\u9a8c\u91d1\uff0c\u628a\u6a21\u578b\u642c\u5230\u56fd\u4ea7\u82af\u7247\u4e0a\u8dd1\u4e00\u6b21\u2014\u2014\u6216\u8bb8\u4f60\u4f1a\u53d1\u73b0\uff0c\u5927\u6a21\u578b\u8bad\u7ec3\u5176\u5b9e\u53ef\u4ee5\u66f4\u8f7b\u677e\u3001\u66f4\u4fbf\u5b9c\uff0c\u4e5f\u66f4\u81ea\u4e3b\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u201c\u5982\u679c\u4eca\u5929\u8fd8\u5728\u7528\u8fdb\u53e3\u5361\u8bad\u5927\u6a21\u578b\uff0c\u5c31\u50cf 2010 \u5e74\u8fd8\u5728\u7528\u8fdb\u53e3&hellip;<\/p>\n","protected":false},"author":2,"featured_media":1977,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-1978","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-zixun"],"views":73,"_links":{"self":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts\/1978","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=1978"}],"version-history":[{"count":0,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts\/1978\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/media\/1977"}],"wp:attachment":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/media?parent=1978"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/categories?post=1978"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/tags?post=1978"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}