{"id":2312,"date":"2026-03-01T16:18:08","date_gmt":"2026-03-01T08:18:08","guid":{"rendered":"https:\/\/www.starverse-ai.com\/guide\/archives\/2312"},"modified":"2026-03-01T16:18:08","modified_gmt":"2026-03-01T08:18:08","slug":"%e8%ae%ad%e7%bb%83%e9%80%9f%e5%ba%a6%e7%bf%bb%e5%80%8d%e5%8d%b4%e6%9b%b4%e4%be%bf%e5%ae%9c%ef%bc%9a%e6%98%9f%e5%ae%87%e6%99%ba%e7%ae%97%e5%b9%b3%e5%8f%b0fp8%e6%b7%b7%e5%90%88%e7%b2%be%e5%ba%a6","status":"publish","type":"post","link":"https:\/\/www.starverse-ai.com\/guide\/archives\/2312","title":{"rendered":"\u8bad\u7ec3\u901f\u5ea6\u7ffb\u500d\u5374\u66f4\u4fbf\u5b9c\uff1a\u661f\u5b87\u667a\u7b97\u5e73\u53f0FP8\u6df7\u5408\u7cbe\u5ea6+\u6db2\u51b7GPU\u96c6\u7fa4\u5b9e\u6d4b"},"content":{"rendered":"
\"\u8bad\u7ec3\u901f\u5ea6\u7ffb\u500d\u5374\u66f4\u4fbf\u5b9c\uff1a\u661f\u5b87\u667a\u7b97\u5e73\u53f0FP8\u6df7\u5408\u7cbe\u5ea6+\u6db2\u51b7GPU\u96c6\u7fa4\u5b9e\u6d4b\"<\/figure>\n
\n

\u201c\u8bad\u7ec3 70B \u5927\u6a21\u578b\uff0cFP16 \u65f6\u4ee3 8 \u5361 A100 \u8981\u8dd1 3 \u5468\uff1f\u73b0\u5728\u53ea\u9700 4 \u5f20 H200 \u6db2\u51b7\u8282\u70b9\uff0c14 \u5929\u5c31\u80fd\u6536\u655b\u3002\u201d\u2014\u2014\u8fd9\u4e0d\u662f\u5b9e\u9a8c\u5ba4 PPT\uff0c\u800c\u662f\u661f\u5b87\u667a\u7b97\u5e73\u53f0\u5728\u672c\u5468\u521a\u7ed3\u675f\u7684\u516c\u5f00\u5b9e\u6d4b\u6570\u636e\u3002\u968f\u7740 Blackwell \u67b6\u6784\u539f\u751f\u652f\u6301 FP8\uff0c\u8bad\u7ec3\u7cbe\u5ea6\u4e0e\u901f\u5ea6\u7684\u8df7\u8df7\u677f\u7b2c\u4e00\u6b21\u5411\u201c\u65e2\u8981\u53c8\u8981\u201d\u503e\u659c\uff0c\u800c\u7387\u5148\u628a FP8 \u6df7\u5408\u7cbe\u5ea6 + \u6db2\u51b7 GPU \u96c6\u7fa4\u642c\u8fdb\u5546\u7528\u73af\u5883\u7684\uff0c\u6b63\u662f\u56fd\u5185 GPU\u670d\u52a1\u5668\u79df\u7528 \u8d5b\u9053\u7684\u65b0\u53d8\u91cf\uff1a\u661f\u5b87\u667a\u7b97\u3002<\/p>\n<\/blockquote>\n

\u884c\u4e1a\u80cc\u666f\uff1a\u7b97\u529b\u7126\u8651\u4e0e\u7535\u8d39\u8d26\u5355\u53cc\u91cd\u5939\u51fb<\/h2>\n

\u8fc7\u53bb 12 \u4e2a\u6708\uff0c\u5927\u6a21\u578b\u53c2\u6570\u91cf\u4ece\u767e\u4ebf\u7ea7\u8df3\u6da8\u5230\u5343\u4ebf\u7ea7\uff0cGPU \u4e91\u4e3b\u673a \u7684\u79df\u91d1\u548c IDC \u7535\u8d39\u540c\u6b65\u98d9\u5347\u3002\u67d0\u5934\u90e8 VC \u8c03\u7814\u663e\u793a\uff0cAI \u521d\u521b\u516c\u53f8 40% \u7684\u878d\u8d44\u6b3e\u6700\u7ec8\u6d41\u8fdb\u4e86\u4e91\u5382\u5546\u7684\u7535\u8868\u3002\u5f53\u201c\u5361\u8352\u201d\u7565\u6709\u7f13\u89e3\uff0c\u201c\u7535\u8352\u201d\u6210\u4e3a\u65b0\u74f6\u9888\u2014\u2014\u4f20\u7edf\u98ce\u51b7 8\u00d7A100 \u8282\u70b9\u6ee1\u8f7d PUE\u22481.6\uff0c\u76f8\u5f53\u4e8e\u6bcf\u8bad\u7ec3 1 \u5ea6\u7535\u7684\u6a21\u578b\uff0c\u5c31\u8981\u989d\u5916\u652f\u4ed8 0.6 \u5ea6\u7535\u7684\u201c\u7a7a\u8c03\u7a0e\u201d\u3002\u5982\u4f55\u628a\u8bad\u7ec3\u901f\u5ea6\u7ffb\u500d\u3001\u518d\u628a\u7535\u8d39\u6253\u4e0b\u6765\uff1f\u7b54\u6848\u5c31\u85cf\u5728\u4e24\u6761\u6280\u672f\u66f2\u7ebf\u91cc\uff1a
\n1. NVIDIA \u5728 Blackwell \u67b6\u6784\u9996\u6b21\u539f\u751f\u652f\u6301 FP8\uff0cTensor Core \u5cf0\u503c\u7b97\u529b\u6bd4 FP16 \u63d0\u5347 2.2\u00d7\uff0c\u663e\u5b58\u5360\u7528\u76f4\u964d 25%\uff1b
\n2. \u6db2\u51b7 GPU \u96c6\u7fa4\u53ef\u5c06 PUE \u538b\u5230 1.15 \u4ee5\u4e0b\uff0c\u673a\u623f\u566a\u97f3 <45dB\uff0c\u76f4\u63a5\u90e8\u7f72\u5199\u5b57\u697c\uff0c\u7701\u6389\u4f20\u7edf\u6570\u636e\u4e2d\u5fc3\u9009\u5740\u3001\u571f\u5efa\u3001\u7535\u529b\u589e\u5bb9\u7684 6 \u4e2a\u6708\u6f2b\u957f\u5468\u671f\u3002<\/p>\n

\u661f\u5b87\u667a\u7b97\uff1a\u628a\u5b9e\u9a8c\u5ba4\u6280\u672f\u505a\u6210\u201c\u53ef\u79df\u7528\u201d\u7684\u5546\u54c1<\/h2>\n

\u5f53\u591a\u6570\u4e91\u5382\u5546\u8fd8\u5728\u7528 FP16 \u62fc\u4ef7\u683c\u65f6\uff0c\u661f\u5b87\u667a\u7b97\u5df2\u9996\u6279\u4e0a\u67b6 H200 \u6db2\u51b7\u8282\u70b9\uff0c\u5e76\u9488\u5bf9 FP8 \u6df7\u5408\u7cbe\u5ea6\u505a\u4e86\u5168\u6808\u9002\u914d\uff1a
\n– \u786c\u4ef6\u5c42\uff1a\u5355\u8282\u70b9 8\u00d7H200 NVLink 4.0\uff0c\u6db2\u51b7\u677f+CDU \u6574\u673a\u67dc\u4ea4\u4ed8\uff0c\u6ee1\u8f7d PUE \u5b9e\u6d4b 1.12\uff0c\u7535\u8d39\u7acb\u7701 20%\u3002
\n– \u8f6f\u4ef6\u5c42\uff1a\u5185\u7f6e PyTorch 2.3 + TransformerEngine \u955c\u50cf\uff0c\u7528\u6237 ssh \u767b\u5f55\u540e\u53ea\u9700\u52a0\u4e00\u884c --dtype float8<\/code> \u5373\u53ef\u5f00\u542f FP8 \u8bad\u7ec3\uff0c\u65e0\u9700\u81ea\u5df1\u7f16\u8bd1 CUDA extension\u3002
\n– \u6570\u636e\u5c42\uff1a\u5e73\u53f0\u9ed8\u8ba4\u6302\u8f7d 3PB \u516c\u5171\u6a21\u578b\u4e0e\u6570\u636e\u96c6\uff0cLLaMA-2\u3001Falcon\u3001\u4e2d\u6587\u609f\u9053\u7b49\u6743\u91cd\u4e00\u952e\u590d\u5236\u5230\u5b9e\u4f8b\uff0c\u7701\u53bb\u6df1\u591c\u6302\u673a\u4e0b\u8f7d 500GB \u6587\u4ef6\u7684\u7126\u8651\u3002
\n– \u751f\u6001\u5c42\uff1aGPU\u670d\u52a1\u5668\u79df\u7528 \u652f\u6301\u6309\u5c0f\u65f6\u3001\u6309\u5929\u3001\u6309\u6708\u4e09\u79cd\u8ba1\u8d39\uff0c\u65b0\u6ce8\u518c\u8d26\u53f7\u5373\u9001 10 \u5143\u4f53\u9a8c\u91d1\uff0c\u8db3\u591f\u8dd1\u5b8c 70B \u6a21\u578b\u5b8c\u6574\u9a8c\u8bc1\u96c6\u3002 <\/p>\n

\u5b9e\u6d4b\uff1aLLaMA-2 70B \u6536\u655b\u65f6\u95f4\u7f29\u77ed 42%\uff0c\u7cbe\u5ea6\u5dee\u8ddd <0.05<\/h2>\n

\u6211\u4eec\u5728\u661f\u5b87\u667a\u7b97\u5e73\u53f0\u5206\u522b\u62c9\u8d77\u4e24\u7ec4 4\u00d7H200 \u6db2\u51b7\u8282\u70b9\uff0c\u4fdd\u6301\u6279\u5c3a\u5bf8\u3001\u5b66\u4e60\u7387\u3001\u6570\u636e\u5206\u7247\u5b8c\u5168\u4e00\u81f4\uff0c\u4ec5\u5207\u6362\u7cbe\u5ea6\u6a21\u5f0f\u3002\u7ed3\u679c\u5982\u4e0b\uff1a <\/p>\n\n\n\n\n\n\n\n\n\n
\u6307\u6807<\/th>\nFP16 \u57fa\u7ebf<\/th>\nFP8 \u5b9e\u6d4b<\/th>\n\u63d0\u5347<\/th>\n<\/tr>\n<\/thead>\n
\u6bcf\u8fed\u4ee3\u65f6\u95f4<\/td>\n18.3 s<\/td>\n10.6 s<\/td>\n-42%<\/td>\n<\/tr>\n
\u663e\u5b58\u5360\u7528<\/td>\n78 GB<\/td>\n58 GB<\/td>\n-26%<\/td>\n<\/tr>\n
\u6536\u655b epoch<\/td>\n12<\/td>\n12<\/td>\n\u6301\u5e73<\/td>\n<\/tr>\n
\u9a8c\u8bc1 perplexity<\/td>\n3.45<\/td>\n3.50<\/td>\n+0.05<\/td>\n<\/tr>\n
\u603b\u7535\u8d39<\/td>\n126 \u5143<\/td>\n95 \u5143<\/td>\n-25%<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

\u4ece\u66f2\u7ebf\u53ef\u4ee5\u770b\u51fa\uff0cFP8 \u5728\u524d 3 \u4e2a epoch \u5c31\u80fd\u903c\u8fd1 FP16 \u7684\u9a8c\u8bc1\u635f\u5931\uff0c\u6700\u7ec8 perplexity \u5dee\u8ddd\u4ec5 0.05\uff0c\u5b8c\u5168\u843d\u5728\u5b9e\u9a8c\u8bef\u5dee\u8303\u56f4\u5185\u3002\u66f4\u5173\u952e\u7684\u662f\uff0c\u5f53\u663e\u5b58\u5360\u7528\u4e0b\u964d 26%\uff0c\u7528\u6237\u53ef\u4ee5\u628a\u539f\u672c\u9700\u8981 8 \u5361\u7684\u8bad\u7ec3\u4efb\u52a1\u538b\u7f29\u5230 4 \u5361\uff0c\u76f4\u63a5\u780d\u6389\u4e00\u534a GPU\u670d\u52a1\u5668\u79df\u7528 \u6210\u672c\u3002<\/p>\n

\u4e0d\u6b62\u7701\u7535\uff0c\u66f4\u662f\u201c\u9759\u201d\u7535<\/h2>\n

\u4f20\u7edf\u98ce\u51b7\u673a\u623f 80dB \u7684\u8f70\u9e23\u8ba9\u5199\u5b57\u697c\u9000\u907f\u4e09\u820d\uff0c\u800c\u661f\u5b87\u667a\u7b97\u6db2\u51b7\u8282\u70b9\u628a\u566a\u97f3\u63a7\u5236\u5728 45dB \u4ee5\u4e0b\uff0c\u4e0e\u56fe\u4e66\u9986\u9605\u89c8\u533a\u76f8\u5f53\u3002\u5317\u4eac\u671b\u4eac\u67d0\u6e38\u620f\u516c\u53f8\u76f4\u63a5\u628a 16 \u8282\u70b9\u6db2\u51b7\u67dc\u90e8\u7f72\u5728 12 \u697c\u7814\u53d1\u533a\uff0c\u9694\u58c1\u5c31\u662f\u7b56\u5212\u5de5\u4f4d\uff0c\u8bad\u7ec3 70B \u6a21\u578b\u65f6\u5458\u5de5\u4ecd\u80fd\u5728 3 \u7c73\u5916\u6b63\u5e38\u5f00\u4f1a\u3002\u65e0\u9700\u5bfb\u627e\u504f\u8fdc\u6570\u636e\u4e2d\u5fc3\uff0c\u4e5f\u4e0d\u7528\u652f\u4ed8\u52a8\u8f84\u6bcf\u5343\u74e6 1.2 \u4e07\u5143\u7684\u7535\u529b\u589e\u5bb9\u8d39\uff0cAI \u56e2\u961f\u628a\u7701\u4e0b\u7684\u4eba\u529b\u3001\u65f6\u95f4\u6295\u5165\u5230\u6570\u636e\u6e05\u6d17\u548c\u7b97\u6cd5\u521b\u65b0\uff0c\u8fd9\u624d\u662f AI\u5e94\u7528 \u843d\u5730\u7684\u6700\u77ed\u8def\u5f84\u3002<\/p>\n

\u6210\u672c\u8d26\u5355\uff1a\u540c\u7b49\u6027\u80fd\u518d\u964d 9% \u6708\u79df\uff0c\u7535\u8d39\u989d\u5916\u7701 15%<\/h2>\n

\u4ee5 8\u00d7H200 \u6db2\u51b7\u8282\u70b9\u4e3a\u4f8b\uff0c\u5e02\u9762\u4e0a FP16 \u65b9\u6848\u6708\u79df 4.8 \u4e07\u5143\uff0cPUE 1.6\uff1b\u661f\u5b87\u667a\u7b97 FP8 \u65b9\u6848\u6708\u79df 4.35 \u4e07\u5143\uff0cPUE 1.12\u3002\u6309 0.8 \u5143\/\u5ea6\u7535\u3001\u6ee1\u8f7d 6.5 kW \u8ba1\u7b97\uff0c30 \u5929\u7535\u8d39\u5dee 1.1 \u4e07\u5143\uff0c\u76f8\u5f53\u4e8e\u518d\u7701 15%\u3002\u7efc\u5408\u79df\u91d1\u4e0e\u7535\u8d39\uff0c\u603b\u6210\u672c\u4e0b\u964d 24%\uff0c\u8bad\u7ec3\u65f6\u95f4\u5374\u7f29\u77ed 42%\uff0c\u5355\u4f4d\u6709\u6548\u7b97\u529b\u6210\u672c\u76f4\u964d 56%\u3002\u5bf9\u4e8e\u9884\u7b97\u6709\u9650\u7684\u79d1\u7814\u7ec4\u3001\u521d\u521b\u516c\u53f8\uff0c\u8fd9\u610f\u5473\u7740\u540c\u6837\u4e00\u7b14\u878d\u8d44\u53ef\u4ee5\u591a\u505a 2 \u8f6e\u8d85\u53c2\u641c\u7d22\uff0c\u6216\u8005\u628a\u6a21\u578b\u5bb9\u91cf\u76f4\u63a5\u7ffb\u500d\u3002<\/p>\n

\u7acb\u5373\u4f53\u9a8c\uff1a\u6ce8\u518c\u5373\u9001 10 \u5143\uff0c0.69 \u5143\/\u5361\u65f6\u8d77<\/h2>\n

\u73b0\u5728\u767b\u5f55 \u661f\u5b87\u667a\u7b97\u5b98\u7f51<\/a> \u6ce8\u518c\uff0c\u65b0\u7528\u6237\u81ea\u52a8\u5230\u8d26 10 \u5143\u4f53\u9a8c\u91d1\uff0c\u53ef\u4efb\u9009 RTX 4090\u3001H200\u3001H100 \u7b49 GPU\u4e91\u4e3b\u673a \u673a\u578b\u3002\u5e73\u53f0\u5df2\u9884\u88c5 50+ AI \u5e94\u7528\u955c\u50cf\uff0c\u5305\u542b Stable Diffusion\u3001ChatGLM-6B\u3001Code Llama \u7b49\u70ed\u95e8\u6a21\u578b\uff0c\u771f\u6b63\u505a\u5230\u201c\u4e00\u952e\u5373\u73a9\u201d\u3002\u5982\u679c\u4f60\u60f3\u628a FP8 \u6df7\u5408\u7cbe\u5ea6\u642c\u8fdb\u81ea\u5df1\u7684\u5927\u6a21\u578b\u8bad\u7ec3\uff0c\u53ea\u9700\u5728\u63a7\u5236\u53f0\u9009\u62e9\u201cPyTorch 2.3 FP8\u201d\u955c\u50cf\uff0cssh \u5f00\u673a\u540e\u8f93\u5165\uff1a<\/p>\n

torchrun --nproc_per_node=8 train.py --dtype float8 --model_size 70B\n<\/code><\/pre>\n
\u8bad\u7ec3\u901f\u5ea6\u7ffb\u500d\u3001\u7535\u8d39\u8d26\u5355\u7f29\u6c34\u3001\u5199\u5b57\u697c\u91cc\u4e5f\u80fd\u9759\u97f3\u8dd1\u5927\u6a21\u578b\u2014\u2014\u661f\u5b87\u667a\u7b97\u5df2\u7ecf\u628a\u672a\u6765\u88c5\u8fdb 4U \u6db2\u51b7\u67dc\uff0c\u7b49\u4f60\u9a8c\u8bc1\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"
\u201c\u8bad\u7ec3 70B \u5927\u6a21\u578b\uff0cFP16 \u65f6\u4ee3 8 \u5361 A100 \u8981…<\/p>\n","protected":false},"author":2,"featured_media":2311,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-2312","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-zixun"],"views":44,"_links":{"self":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts\/2312","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=2312"}],"version-history":[{"count":0,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts\/2312\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/media\/2311"}],"wp:attachment":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/media?parent=2312"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/categories?post=2312"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/tags?post=2312"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}