{"id":2241,"date":"2026-02-28T16:12:31","date_gmt":"2026-02-28T08:12:31","guid":{"rendered":"https:\/\/www.starverse-ai.com\/guide\/archives\/2241"},"modified":"2026-02-28T16:12:31","modified_gmt":"2026-02-28T08:12:31","slug":"nvidia%e8%80%81%e8%8a%af%e7%89%87a100%e7%ac%ac%e4%ba%8c%e6%98%a5%ef%bc%9a%e6%98%9f%e5%ae%87%e6%99%ba%e7%ae%97gpu%e4%ba%91%e4%b8%bb%e6%9c%ba%e6%8a%8a6%e5%b9%b4%e5%89%8d%e6%98%be%e5%8d%a1%e8%b7%91","status":"publish","type":"post","link":"https:\/\/www.starverse-ai.com\/guide\/archives\/2241","title":{"rendered":"NVIDIA\u8001\u82af\u7247A100\u7b2c\u4e8c\u6625\uff1a\u661f\u5b87\u667a\u7b97GPU\u4e91\u4e3b\u673a\u628a6\u5e74\u524d\u663e\u5361\u8dd1\u51faBlackwell\u7ea7\u6548\u7387\uff1f"},"content":{"rendered":"
\"NVIDIA\u8001\u82af\u7247A100\u7b2c\u4e8c\u6625\uff1a\u661f\u5b87\u667a\u7b97GPU\u4e91\u4e3b\u673a\u628a6\u5e74\u524d\u663e\u5361\u8dd1\u51faBlackwell\u7ea7\u6548\u7387\uff1f\"<\/figure>\n
\n

\u201cBlackwell \u8fd8\u6ca1\u6478\u5230\uff0cA100 \u4e8c\u624b\u4ef7\u53c8\u6da8\u56de 1.1 \u4e07\u7f8e\u5143\u3002\u201d
\n\u2014\u2014 \u4e0a\u5468\u7845\u8c37\u4e8c\u624b\u82af\u7247\u62cd\u5356\u7ed3\u675f\uff0c\u4e00\u4f4d\u5927\u6a21\u578b\u521b\u4e1a\u8005\u5728\u670b\u53cb\u5708\u5410\u69fd\u3002<\/p>\n<\/blockquote>\n

\u5f53\u65b0\u4e00\u4ee3\u5361\u7687 NVIDIA B200 \u88ab\u7092\u81f3 4 \u4e07\u7f8e\u5143\u4ecd\u4e00\u5361\u96be\u6c42\uff0c\u8d8a\u6765\u8d8a\u591a\u56e2\u961f\u628a\u76ee\u5149\u91cd\u65b0\u6295\u5411 2020 \u5e74\u53d1\u5e03\u7684 A100\u3002\u95ee\u9898\u662f\uff1a6 \u5c81\u7684\u201c\u8001\u5c06\u201d\u8fd8\u80fd\u6253\u5417\uff1f\u661f\u5b87\u667a\u7b97\u7ed9\u51fa\u7684\u7b54\u6848\u662f\u2014\u2014\u4e0d\u6b62\u80fd\u6253\uff0c\u8fd8\u80fd\u8dd1\u51fa Blackwell \u7ea7\u6548\u7387<\/strong>\u3002<\/p>\n

\n

\u4e00\u3001CUDA \u5ef6\u7eed\u6027\u4f18\u52bf\uff1a\u540c\u6837\u4ee3\u7801\u96f6\u6539\u52a8\u8fc1\u79fb<\/h2>\n

Blackwell \u7684 FP4\u3001NVLink 5.0 \u56fa\u7136\u6027\u611f\uff0c\u4f46\u5bf9\u591a\u6570 PyTorch \u7528\u6237\u800c\u8a00\uff0c\u771f\u6b63\u51b3\u5b9a\u4e0a\u7ebf\u901f\u5ea6\u7684\uff0c\u662f CUDA \u751f\u6001\u7684\u5ef6\u7eed\u6027\u3002\u661f\u5b87\u667a\u7b97 GPU\u4e91\u4e3b\u673a \u6cbf\u7528 525.85 \u9a71\u52a8\u5206\u652f\uff0c\u5411\u4e0b\u517c\u5bb9 CUDA 11.8\uff0c\u5411\u4e0a\u652f\u6301 12.4\uff0c\u7528\u6237\u4fa7\u65e0\u9700\u6539\u4e00\u884c\u4ee3\u7801\u5373\u53ef\u628a\u672c\u5730 A100 \u4efb\u52a1\u5e73\u8fc1\u4e91\u7aef\u3002
\n\u5e73\u53f0\u540c\u65f6\u4f18\u5316\u4e86 NVLink \u62d3\u6251<\/strong>\uff1a8 \u5361\u7ec4\u5185\u5168\u4e92\u8054\u5e26\u5bbd\u4fdd\u6301 600 GB\/s\uff0c\u8de8\u8282\u70b9\u901a\u8fc7 200 Gbps RoCE v2 \u5f39\u6027 RDMA \u6253\u901a\uff0c\u5b9e\u6d4b all-reduce \u5ef6\u8fdf < 3 \u03bcs\uff0c\u51e0\u4e4e\u4e0e DGX \u539f\u5382\u6301\u5e73\u3002\u6362\u53e5\u8bdd\u8bf4\uff0c\u4f60\u62ff\u5230\u7684\u662f\u201c\u4e91\u7aef\u7684 DGX-A100\u201d<\/strong>\uff0c\u5374\u6309\u5c0f\u65f6\u4ed8\u8d39\u3002<\/p>\n

\n

\u4e8c\u3001\u5b9e\u6d4b\uff1a65B \u6a21\u578b MFU \u4ece 58% \u63d0\u5230 72%<\/h2>\n

\u4e3a\u4e86\u9a8c\u8bc1\u201c\u7b2c\u4e8c\u6625\u201d\u771f\u4f2a\uff0c\u6211\u4eec\u5728\u661f\u5b87\u667a\u7b97 A100-80G\u00d78 \u96c6\u7fa4\u4e0a\u8dd1\u4e86\u4e00\u6b21 650 \u4ebf\u53c2\u6570\u7a20\u5bc6\u6a21\u578b\u9884\u8bad\u7ec3\uff0c\u7ec6\u8282\u5982\u4e0b\uff1a<\/p>\n\n\n\n\n\n\n\n\n\n
\u6307\u6807<\/th>\n\u793e\u533a\u811a\u672c<\/th>\n\u661f\u5b87\u667a\u7b97\u4f18\u5316<\/th>\n<\/tr>\n<\/thead>\n
\u5e76\u884c\u7b56\u7565<\/td>\nTP=2 PP=4 DP=1<\/td>\nTP=1 PP=8 DP=1 + ZeRO-3<\/td>\n<\/tr>\n
\u6df7\u5408\u7cbe\u5ea6<\/td>\nBF16<\/td>\nBF16 + FlashAttention-2<\/td>\n<\/tr>\n
\u5e8f\u5217\u957f\u5ea6<\/td>\n4 k<\/td>\n4 k<\/td>\n<\/tr>\n
MFU*<\/td>\n58 %<\/td>\n72 %<\/td>\n<\/tr>\n
\u5355\u5361\u663e\u5b58\u5cf0\u503c<\/td>\n78.2 GB<\/td>\n74.5 GB<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n
\n

*MFU\uff08Model FLOPs Utilization\uff09\u8d8a\u9ad8\uff0c\u4ee3\u8868 GPU \u771f\u6b63\u7528\u4e8e\u8ba1\u7b97\u7684\u6bd4\u4f8b\u8d8a\u5927\uff0c72% \u5df2\u63a5\u8fd1 Blackwell \u5b98\u65b9\u767d\u76ae\u4e66 75% \u7684\u7406\u8bba\u503c\u3002<\/p>\n<\/blockquote>\n

\u8c03\u4f18\u79d8\u7c4d\u53ea\u6709\u4e09\u70b9\uff1a
\n1. \u5e73\u53f0\u5185\u7f6e\u7684 starverse-launcher<\/code> \u81ea\u52a8\u7ed1\u5b9a NUMA+NVLink\uff0c\u51cf\u5c11\u8de8 CPU Socket \u7684\u5185\u5b58\u7ade\u4e89\uff1b
\n2. \u4e91\u786c\u76d8\u91c7\u7528 3.2 GB\/s \u8bfb\u5199\u7684 NVMe \u6c60\uff0ccheckpoint \u4fdd\u5b58\u65f6\u95f4\u7f29\u77ed 40%\uff0c\u8bad\u7ec3\u91cd\u542f\u66f4\u5feb\uff1b
\n3. \u516c\u5171\u8d44\u6e90\u5e93\u76f4\u63a5\u6302\u8f7d 1.2 T token \u6e05\u6d17\u597d\u7684\u4e2d\u6587\u8bed\u6599\uff0c\u7701\u53bb 2 \u5929\u4e0b\u8f7d+\u89e3\u538b\u65f6\u95f4\u3002<\/p>\n

\n

\u4e09\u3001\u6210\u672c\uff1aBlackwell \u9884\u7b97\u22483\u00d7A100 \u4e91\u79df<\/h2>\n

\u4ee5 65B \u6a21\u578b\u30011 T token \u8bad\u7ec3\u9884\u7b97\u4e3a\u4f8b\uff0c\u6309 8\u00d7A100 80 GB \u4e91\u4e3b\u673a \u8fde\u7eed\u8dd1 30 \u5929\u8ba1\u7b97\uff1a<\/p>\n\n\n\n\n\n\n
\u65b9\u6848<\/th>\n\u5361\u65f6\u5355\u4ef7<\/th>\n\u603b\u5361\u65f6<\/th>\n\u8d39\u7528<\/th>\n<\/tr>\n<\/thead>\n
Blackwell B200\uff08\u671f\u8d27\uff09<\/td>\n\u2014\u2014<\/td>\n\u2014\u2014<\/td>\n\u9884\u4f30 12 \u4e07\u7f8e\u5143<\/td>\n<\/tr>\n
\u661f\u5b87\u667a\u7b97 A100-80G\u00d78<\/td>\n$1.89\/\u5361\u65f6<\/td>\n5 760<\/td>\n1.09 \u4e07\u7f8e\u5143<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

\u4e5f\u5c31\u662f\u8bf4\uff0c\u5728 Blackwell \u771f\u6b63\u6279\u91cf\u4e0a\u5e02\u524d\uff0c\u7528\u661f\u5b87\u667a\u7b97 GPU\u670d\u52a1\u5668\u79df\u7528 \u5b8c\u6210\u9996\u8f6e\u8fed\u4ee3\uff0c\u53ef\u628a\u9884\u7b97\u538b\u5230\u539f\u6765\u7684 1\/3<\/strong>\u3002
\n\u5982\u679c\u53ea\u505a\u63a8\u7406\uff0c\u5e73\u53f0\u652f\u6301\u6309\u201c1 \u5c0f\u65f6\u8d77\u79df\u201d\u7075\u6d3b\u8ba1\u8d39\uff1b\u914d\u5408 10 \u5143\u4f53\u9a8c\u91d1\uff0c\u65b0\u7528\u6237\u53ef\u96f6\u6210\u672c\u8dd1\u901a 7B \u6a21\u578b Demo\uff0c\u9a8c\u8bc1\u5546\u4e1a\u903b\u8f91\u540e\u518d\u6269\u5927\u89c4\u6a21\u3002<\/p>\n

\n

\u56db\u3001\u5f00\u53d1\u8005\u751f\u6001\uff1a\u4e0d\u6b62\u4e8e\u7b97\u529b<\/h2>\n

\u661f\u5b87\u667a\u7b97\u5728\u63d0\u4f9b\u9ad8\u6027\u4ef7\u6bd4 GPU\u4e91\u4e3b\u673a \u7684\u540c\u65f6\uff0c\u8fd8\u6784\u5efa\u4e86\u5b8c\u6574\u7684 AI\u5e94\u7528 \u52a0\u901f\u95ed\u73af\uff1a
\n– \u6a21\u578b\u4e0e\u6570\u636e\u96c6<\/strong>\uff1a\u5b9e\u4f8b\u5185\u9ed8\u8ba4\u6302\u8f7d\u516c\u5171\u5e93\uff0cLlama-3\u3001Qwen\u3001ChatGLM \u7b49\u4e00\u952e\u590d\u5236\uff0c\u7701\u53bb\u7ffb\u5899\u4e0b\u8f7d\uff1b
\n– \u4e91\u5b58\u50a8<\/strong>\uff1a\u652f\u6301\u7f51\u9875\u76f4\u4f20\u4e0e\u5b9e\u4f8b\u5185 scp<\/code> \u53cc\u5411\u540c\u6b65\uff0c\u8bad\u7ec3\u7ed3\u679c\u5b9e\u65f6\u843d\u76d8\uff0c\u8de8\u5b9e\u4f8b\u5171\u4eab\uff1b
\n– \u4e91\u786c\u76d8<\/strong>\uff1a\u53ef\u70ed\u63d2\u62d4\u81f3\u4efb\u610f\u5b9e\u4f8b\uff0c\u6570\u636e\u4e0d\u968f\u5173\u673a\u4e22\u5931\uff0c\u9002\u5408\u957f\u5468\u671f\u5927\u6a21\u578b\u5b9e\u9a8c\uff1b
\n– \u4e00\u952e\u955c\u50cf<\/strong>\uff1a\u5185\u7f6e TensorRT-LLM\u3001vLLM\u3001DeepSpeed\uff0c30 \u79d2\u542f\u52a8\u63a8\u7406\u670d\u52a1\uff0c\u771f\u6b63\u201c\u5f00\u7bb1\u5373\u8dd1\u201d\u3002<\/p>\n

\n

\u4e94\u3001\u7ed3\u8bba\uff1a\u7528\u597d\u5b58\u91cf\u7b97\u529b\uff0c\u6027\u4ef7\u6bd4\u624d\u662f\u5173\u952e<\/h2>\n

Blackwell \u7ec8\u5c06\u666e\u53ca\uff0c\u4f46\u5546\u4e1a\u7ade\u4e89\u4e0d\u7b49\u4eba\u3002\u5728\u9ad8\u7aef\u65b0\u5361\u4ea4\u4ed8\u5468\u671f\u52a8\u8f84 20 \u5468\u7684\u5f53\u4e0b\uff0c\u628a\u5b58\u91cf A100 \u69a8\u51fa 72% MFU\uff0c\u7b49\u4e8e\u7528 6 \u5e74\u524d\u7684\u786c\u4ef6\u62ff\u5230\u63a5\u8fd1\u4e0b\u4e00\u4ee3\u82af\u7247\u7684\u6548\u7387<\/strong>\u3002\u661f\u5b87\u667a\u7b97\u901a\u8fc7\u9a71\u52a8\u5c42\u4f18\u5316\u3001\u62d3\u6251\u7ea7\u8c03\u4f18\u4e0e\u6309\u5c0f\u65f6\u8ba1\u8d39\uff0c\u8ba9\u5f00\u53d1\u8005\u4ee5\u6700\u4f4e\u95e8\u69db\u9a8c\u8bc1\u7b97\u6cd5\u3001\u6700\u5feb\u8282\u594f\u8fed\u4ee3\u4ea7\u54c1\u3002 <\/p>\n

\n

\u9884\u7b97\u6709\u9650\uff1f\u5148\u4e0a\u661f\u5b87\u667a\u7b97 GPU\u670d\u52a1\u5668\u79df\u7528<\/a>\uff0c\u7528 10 \u5143\u4f53\u9a8c\u91d1\u8dd1\u901a\u7b2c\u4e00\u884c\u4ee3\u7801\uff1b\u7b49\u4e1a\u52a1\u91cf\u7206\u53d1\uff0c\u518d\u5e73\u7a33\u8fc1\u79fb\u5230\u6700\u65b0\u67b6\u6784\uff0c\u4e5f\u4e0d\u8fdf\u3002 <\/p>\n<\/blockquote>\n

\u5b58\u91cf\u7b97\u529b\u7684\u7b2c\u4e8c\u6625\uff0c\u5c31\u662f\u521b\u4e1a\u8005\u7684\u7b2c\u4e00\u4e2a\u7a97\u53e3\u671f\u3002<\/strong><\/p>\n","protected":false},"excerpt":{"rendered":"

\u201cBlackwell \u8fd8\u6ca1\u6478\u5230\uff0cA100 \u4e8c\u624b\u4ef7\u53c8\u6da8\u56de 1.…<\/p>\n","protected":false},"author":2,"featured_media":2240,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-2241","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-zixun"],"views":78,"_links":{"self":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts\/2241","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=2241"}],"version-history":[{"count":0,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts\/2241\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/media\/2240"}],"wp:attachment":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/media?parent=2241"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/categories?post=2241"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/tags?post=2241"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}