{"id":2221,"date":"2026-02-28T14:16:34","date_gmt":"2026-02-28T06:16:34","guid":{"rendered":"https:\/\/www.starverse-ai.com\/guide\/archives\/2221"},"modified":"2026-02-28T14:16:34","modified_gmt":"2026-02-28T06:16:34","slug":"%e6%8e%a8%e7%90%86%e6%97%b6%e4%bb%a3%e7%9a%84%e7%ae%97%e5%8a%9b%e6%96%b0%e7%a7%a9%e5%ba%8f%ef%bc%9a%e4%b8%ba%e4%bb%80%e4%b9%8810x%e4%bd%8e%e5%bb%b6%e8%bf%9f%e6%af%941000tflops%e6%9b%b4%e9%87%8d","status":"publish","type":"post","link":"https:\/\/www.starverse-ai.com\/guide\/archives\/2221","title":{"rendered":"\u63a8\u7406\u65f6\u4ee3\u7684\u7b97\u529b\u65b0\u79e9\u5e8f\uff1a\u4e3a\u4ec0\u4e4810\u00d7\u4f4e\u5ef6\u8fdf\u6bd41000TFLOPS\u66f4\u91cd\u8981\uff1f"},"content":{"rendered":"<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/www.starverse-ai.com\/guide\/wp-content\/uploads\/2026\/02\/1772259393_01829d.png\" alt=\"\u63a8\u7406\u65f6\u4ee3\u7684\u7b97\u529b\u65b0\u79e9\u5e8f\uff1a\u4e3a\u4ec0\u4e4810\u00d7\u4f4e\u5ef6\u8fdf\u6bd41000TFLOPS\u66f4\u91cd\u8981\uff1f\" style=\"display:block; margin:10px auto; max-width:100%; height:auto;\" \/><\/figure>\n<blockquote>\n<p>\u201c\u8bad\u7ec3\u7b97\u529b\u8fc7\u5269\uff0c\u63a8\u7406\u7b97\u529b\u77ed\u7f3a\u3002\u201d<br \/>\n\u8fd9\u662f\u82f1\u4f1f\u8fbe 2024 \u5f00\u53d1\u8005\u5927\u4f1a\uff08GTC\uff09\u4e0a\u88ab\u53cd\u590d\u63d0\u53ca\u7684\u4e00\u53e5\u8bdd\u3002\u5f53\u5927\u6a21\u578b\u53c2\u6570\u7a81\u7834\u4e07\u4ebf\u95e8\u69db\uff0c\u884c\u4e1a\u9aa4\u7136\u53d1\u73b0\uff1a\u7528\u6237\u771f\u6b63\u4e70\u5355\u7684\uff0c\u4e0d\u662f\u5b9e\u9a8c\u5ba4\u91cc\u8dd1\u51fa\u7684 loss \u66f2\u7ebf\uff0c\u800c\u662f\u7ebf\u4e0a\u6bcf\u4e00\u6b21\u201c\u79d2\u56de\u201d\u7684\u4e1d\u6ed1\u4f53\u9a8c\u3002IDC \u6700\u65b0\u62a5\u544a\u4f50\u8bc1\u2014\u20142025 \u5e74\u5168\u7403 AI \u7b97\u529b\u652f\u51fa\u4e2d\uff0c<strong>\u63a8\u7406\u5360\u6bd4\u5c06\u9996\u6b21\u8d85\u8fc7 70%<\/strong>\u3002\u5ef6\u8fdf\u6bcf\u589e\u52a0 100ms\uff0c\u5bf9\u8bdd\u7c7b\u5e94\u7528\u7559\u5b58\u7387\u5c31\u6389 3%\uff1b\u7535\u5546\u5ba2\u670d\u673a\u5668\u4eba\u6296\u52a8\u4e00\u6b21\uff0c\u5ba2\u5355\u4ef7\u5c31\u84b8\u53d1 5%\u3002\u5728\u201c\u4f53\u9a8c\u5373\u6210\u672c\u201d\u7684\u6b8b\u9177\u516c\u5f0f\u91cc\uff0c<strong>10\u00d7\u4f4e\u5ef6\u8fdf\u7684\u4ef7\u503c\uff0c\u65e9\u5df2\u78be\u538b 1000 TFLOPS \u7684\u7eb8\u9762\u7b97\u529b<\/strong>\u3002<\/p>\n<\/blockquote>\n<hr \/>\n<h2>\u4ece\u201c\u7b97\u529b\u519b\u5907\u201d\u5230\u201c\u5ef6\u8fdf\u6218\u4e89\u201d<\/h2>\n<p>\u8fc7\u53bb\u4e09\u5e74\uff0c\u4e91\u5382\u5546\u7684 PR \u7a3f\u91cc\u5199\u6ee1\u4e86\u201c\u4e07\u5361\u96c6\u7fa4\u201d\u201cEFLOPS \u7ea7\u8bad\u7ec3\u6c60\u201d\u3002\u4f46\u5f53\u6a21\u578b\u843d\u5730\u5230\u771f\u5b9e\u573a\u666f\uff0c\u5f00\u53d1\u8005\u5374\u9677\u5165\u53e6\u4e00\u756a\u82e6\u6218\uff1a<\/p>\n<ul>\n<li>\u591a\u79df\u6237\u62a2\u5360\uff0cGPU \u4e0a\u4e0b\u6587\u9891\u7e41\u5207\u6362\uff0cP99 \u5ef6\u8fdf\u52a8\u8f84\u7ffb\u500d\uff1b  <\/li>\n<li>\u865a\u62df\u5316\u5c42\u539a\u91cd\uff0cGPU \u76f4\u901a\u6bd4\u7387\u4e0d\u8db3 40%\uff0ctoken \u6d41\u50cf\u201c\u6324\u7259\u818f\u201d\uff1b  <\/li>\n<li>\u8de8 NUMA \u8282\u70b9\u8bbf\u5b58\uff0c\u4e00\u6b21\u63a8\u7406\u8981\u8d70 16 \u8df3\uff0c\u5ef6\u8fdf\u88ab\u653e\u5927 5 \u500d\u3002<\/li>\n<\/ul>\n<p>\u7ed3\u679c\u5f88\u9b54\u5e7b\uff1a\u8bad\u7ec3\u9636\u6bb5\u53ef\u4ee5\u5806\u5361\u201c\u5927\u529b\u51fa\u5947\u8ff9\u201d\uff0c\u63a8\u7406\u9636\u6bb5\u5374\u201c\u5341\u5361\u4e5d\u6296\u201d\u3002\u4e00\u4f4d AIGC \u521b\u4e1a\u8005\u5728\u670b\u53cb\u5708\u5410\u69fd\uff1a\u201c\u540c\u6837 Llama-3-8B\uff0c\u522b\u4eba 2600 token\/s\uff0c\u6211\u7684 800 token\/s \u8fd8\u6389\u7ebf\uff0c\u7528\u6237\u5168\u8dd1\u7ade\u54c1\u90a3\u513f\u53bb\u4e86\u3002\u201d<\/p>\n<hr \/>\n<h2>\u661f\u5b87\u667a\u7b97\uff1a\u628a\u201c\u5ef6\u8fdf\u201d\u505a\u6210\u53ef\u627f\u8bfa\u7684 SLA<\/h2>\n<p>\u5f53\u884c\u4e1a\u8fd8\u5728\u5377\u5cf0\u503c\u7b97\u529b\uff0c<a href=\"https:\/\/www.starverse-ai.com\">\u661f\u5b87\u667a\u7b97<\/a> \u628a\u67aa\u53e3\u5bf9\u51c6\u4e86\u201c\u6700\u540e\u4e00\u6beb\u79d2\u201d\u3002\u8fd9\u5bb6\u6765\u81ea\u53a6\u95e8\u7684 AI \u667a\u7b97\u53ca\u5e94\u7528\u751f\u6001\u5e73\u53f0\uff0c\u7528\u4e00\u5957\u201c\u4e91\u539f\u751f + \u88f8\u91d1\u5c5e\u201d\u6df7\u5408\u67b6\u6784\uff0c\u628a\u63a8\u7406\u5ef6\u8fdf\u538b\u8fdb\u4e86 <strong>P99&lt;40ms<\/strong> \u7684\u533a\u95f4\uff0c\u8ba9 GPU\u670d\u52a1\u5668\u79df\u7528 \u7b2c\u4e00\u6b21\u62e5\u6709\u4e86\u53ef\u5199\u5165\u5408\u540c\u7684 SLA\u3002<\/p>\n<h3>\u81ea\u7814\u8c03\u5ea6\u5668\uff1a\u544a\u522b\u201c\u90bb\u5c45\u6270\u52a8\u201d<\/h3>\n<p>\u4f20\u7edf\u4e91\u5382\u5546\u7684 Kubernetes \u8c03\u5ea6\u5668\u9ed8\u8ba4\u201c\u88c5\u7bb1\u7387\u4f18\u5148\u201d\uff0c\u661f\u5b87\u667a\u7b97\u91cd\u5199\u8c03\u5ea6\u903b\u8f91\uff0c\u4ee5\u201c\u5ef6\u8fdf\u654f\u611f\u201d\u4e3a\u7b2c\u4e00\u4f18\u5148\u7ea7\uff1a<br \/>\n&#8211; GPU \u7ed1\u5b9a\u7269\u7406\u6838\uff0cvCPU \u62a2\u5360\u7387\u964d\u5230 0\uff1b<br \/>\n&#8211; \u5185\u5b58\u5927\u9875 + NUMA \u4eb2\u548c\uff0c\u8de8\u8282\u70b9\u8bbf\u95ee\u5ef6\u8fdf\u964d\u4f4e 62%\uff1b<br \/>\n&#8211; \u7f51\u7edc\u5c42\u57fa\u4e8e DPDK \u505a\u7528\u6237\u6001\u534f\u8bae\u6808\uff0c\u5185\u6838\u65c1\u8def\u8ba9\u6296\u52a8&lt;1ms\u3002<\/p>\n<h3>GPU \u76f4\u901a\uff1a\u865a\u62df\u5316\u201c0\u201d\u635f\u8017<\/h3>\n<p>\u5e73\u53f0\u6240\u6709 <a href=\"https:\/\/www.starverse-ai.com\">GPU\u4e91\u4e3b\u673a<\/a> \u5747\u91c7\u7528 PCIe Gen4 \u76f4\u901a\uff0cVFIO \u65b9\u6848\u7ed5\u8fc7 QEMU\uff0c\u7b97\u529b\u635f\u8017\u63a7\u5236\u5728 1% \u4ee5\u5185\u3002\u5bf9\u6bd4\u884c\u4e1a 10%-15% \u7684\u865a\u62df\u5316\u7a0e\uff0c\u76f8\u5f53\u4e8e<strong>\u6bcf 10 \u5f20\u5361\u767d\u6361 1 \u5f20<\/strong>\u3002<\/p>\n<h3>\u7cbe\u54c1\u8282\u70b9\uff1a\u4e13\u4e3a\u63a8\u7406\u800c\u751f<\/h3>\n<p>\u661f\u5b87\u667a\u7b97\u5728\u534e\u5317\u3001\u534e\u4e1c\u3001\u534e\u5357\u4e09\u5927\u673a\u623f\u5212\u51fa\u201c\u63a8\u7406\u4e13\u5c5e\u6c60\u201d\uff0c\u5355\u8282\u70b9\u4ec5\u90e8\u7f72 8 \u5f20 H100\uff0cCPU \u4e0e GPU \u6bd4\u4f8b 1:1\uff0c\u914d\u5957 400 Gbps RDMA \u7f51\u7edc\uff0c<strong>\u5f7b\u5e95\u9694\u79bb\u8bad\u7ec3\u6d41\u91cf<\/strong>\u3002\u4e0a\u7ebf\u4e09\u4e2a\u6708\uff0c\u5e73\u5747 P99 \u5ef6\u8fdf\u7a33\u5b9a\u5728 38 ms\uff0c\u6700\u5927\u6296\u52a8\u4e0d\u8d85\u8fc7 5 ms\u3002<\/p>\n<hr \/>\n<h2>\u5b9e\u6d4b\uff1aLlama-3-8B-Q4KS 2600 token\/s \u662f\u600e\u6837\u70bc\u6210\u7684\uff1f<\/h2>\n<p>\u6211\u4eec\u5728\u661f\u5b87\u667a\u7b97\u534e\u5357-A \u533a\u79df\u7528\u4e86\u4e00\u53f0 <code>h100-1x<\/code> \u89c4\u683c <a href=\"https:\/\/www.starverse-ai.com\">GPU\u4e91\u4e3b\u673a<\/a>\uff08\u5355\u5361 80 GB SXM\uff09\uff0c\u64cd\u4f5c\u7cfb\u7edf\u91c7\u7528 Ubuntu 22.04 + CUDA 12.3\uff0c\u63a8\u7406\u6846\u67b6\u4e3a vLLM-0.4.2\uff0c\u6d4b\u8bd5\u6570\u636e\u96c6\u53d6 Wikipedia \u4e2d\u6587 10 k \u6761\u968f\u673a Prompt\uff0c\u8f93\u5165\u957f\u5ea6 512 token\uff0c\u8f93\u51fa\u957f\u5ea6 2048 token\u3002<\/p>\n<table>\n<thead>\n<tr>\n<th>\u6307\u6807<\/th>\n<th>\u661f\u5b87\u667a\u7b97<\/th>\n<th>\u5e38\u89c4\u516c\u6709\u4e91 A<\/th>\n<th>\u5e38\u89c4\u516c\u6709\u4e91 B<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>\u5e73\u5747\u5ef6\u8fdf<\/td>\n<td>33 ms<\/td>\n<td>71 ms<\/td>\n<td>68 ms<\/td>\n<\/tr>\n<tr>\n<td>P99 \u5ef6\u8fdf<\/td>\n<td>38 ms<\/td>\n<td>152 ms<\/td>\n<td>145 ms<\/td>\n<\/tr>\n<tr>\n<td>\u541e\u5410<\/td>\n<td>2600 token\/s<\/td>\n<td>1920 token\/s<\/td>\n<td>1830 token\/s<\/td>\n<\/tr>\n<tr>\n<td>\u6296\u52a8<\/td>\n<td>\u00b15 ms<\/td>\n<td>\u00b147 ms<\/td>\n<td>\u00b141 ms<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u5728 1 \u5c0f\u65f6\u6301\u7eed\u538b\u6d4b\u4e2d\uff0c\u661f\u5b87\u667a\u7b97\u8282\u70b9<strong>\u65e0\u4e00\u6b21\u6389\u7ebf\u3001\u65e0\u4e00\u6b21\u8d85\u65f6<\/strong>\uff0c\u800c\u4e24\u5bb6\u5bf9\u7167\u7ec4\u5206\u522b\u51fa\u73b0 12 \u6b21\u4e0e 9 \u6b21 500 \u9519\u8bef\u3002\u5bf9\u4e8e\u76f4\u64ad\u5f39\u5e55\u3001\u5b9e\u65f6\u5ba2\u670d\u3001AI \u966a\u4f34\u7b49\u573a\u666f\uff0c\u8fd9\u610f\u5473\u7740\u7528\u6237\u4fa7\u201c\u96f6\u611f\u77e5\u201d\u5361\u987f\uff0c\u7559\u5b58\u7387\u76f4\u63a5\u62c9\u6ee1\u3002<\/p>\n<hr \/>\n<h2>\u5f00\u53d1\u8005\u751f\u6001\uff1a\u8ba9\u201c\u4f4e\u5ef6\u8fdf\u201d\u4e0d\u6b62\u4e8e\u53e3\u53f7<\/h2>\n<p>\u661f\u5b87\u667a\u7b97\u6df1\u77e5\uff0c<strong>\u7b97\u529b\u53ea\u662f\u5165\u573a\u5238\uff0c\u751f\u6001\u624d\u662f\u62a4\u57ce\u6cb3<\/strong>\u3002\u5e73\u53f0\u5185\u7f6e 300+ \u516c\u5171\u6a21\u578b\u3001500 TB \u5408\u89c4\u6570\u636e\u96c6\uff0c\u914d\u5408\u201c\u4e00\u952e\u5373\u73a9\u201d <a href=\"https:\/\/www.starverse-ai.com\">AI\u5e94\u7528<\/a> \u5e02\u573a\uff0c5 \u5206\u949f\u5c31\u80fd\u62c9\u8d77\u4e00\u5957\u201c\u79c1\u6709 ChatGPT\u201d\u3002<\/p>\n<ul>\n<li>\u6301\u4e45\u5316\u4e91\u5b58\u50a8\u8de8\u5b9e\u4f8b\u5171\u4eab\uff0c\u6a21\u578b\u6743\u91cd\u4e00\u6b21\u4e0b\u8f7d\uff0c\u5168\u6c60\u590d\u7528\uff1b  <\/li>\n<li>\u955c\u50cf\u5e02\u573a\u9884\u7f6e vLLM\u3001TensorRT-LLM\u3001TGI \u7b49\u63a8\u7406\u52a0\u901f\u5305\uff0c\u5f00\u673a\u5373\u4e0a\u7ebf\uff1b  <\/li>\n<li>\u7075\u6d3b\u8ba1\u8d39\u6309\u79d2\u6263\u8d39\uff0c\u652f\u6301 Spot \u5b9e\u4f8b\u518d\u7701 70%\uff0c\u521d\u521b\u516c\u53f8\u4e5f\u80fd\u201c\u65e5\u79df\u201dH100\u3002<\/li>\n<\/ul>\n<hr \/>\n<h2>0 \u5143\u4f53\u9a8c\uff1a\u628a 72 \u5c0f\u65f6\u505a\u6210\u201c\u63a8\u7406\u52a0\u901f\u5238\u201d<\/h2>\n<p>\u5373\u65e5\u8d77\u81f3 6 \u6708 30 \u65e5\uff0c\u65b0\u7528\u6237\u6ce8\u518c <a href=\"https:\/\/www.starverse-ai.com\">\u661f\u5b87\u667a\u7b97<\/a> \u5373\u53ef\u9886\u53d6 <strong>10 \u5143\u4f53\u9a8c\u91d1 + \u63a8\u7406\u52a0\u901f\u5238<\/strong>\uff0c0 \u5143\u7545\u4eab H100 GPU\u4e91\u4e3b\u673a 72 \u5c0f\u65f6\uff0c\u8db3\u591f\u628a Llama-3\u3001Qwen1.5\u3001ChatGLM3 \u5168\u90e8\u8dd1\u4e00\u904d\u3002\u4f53\u9a8c\u5230\u671f\u540e\uff0c\u6309\u9700\u91ca\u653e\uff0c<strong>\u4e0d\u82b1\u4e00\u5206\u51a4\u6789\u94b1<\/strong>\u3002<\/p>\n<hr \/>\n<h2>\u7ed3\u8bed\uff1a\u5ef6\u8fdf\u624d\u662f\u4e0b\u4e00\u5ea7\u91d1\u77ff<\/h2>\n<p>\u5f53\u5927\u6a21\u578b\u8d70\u5411\u5343\u884c\u767e\u4e1a\uff0c<strong>\u201c\u5feb\u201d\u4e0d\u518d\u662f\u4e00\u79cd\u4f53\u9a8c\uff0c\u800c\u662f\u4e00\u6761\u751f\u6b7b\u7ebf<\/strong>\u3002\u5728\u661f\u5b87\u667a\u7b97\u7684\u5de5\u7a0b\u54f2\u5b66\u91cc\uff0c10\u00d7\u4f4e\u5ef6\u8fdf\u4e0d\u662f\u8425\u9500\u8bdd\u672f\uff0c\u800c\u662f\u628a\u6bcf\u4e00\u6b21 kernel \u542f\u52a8\u3001\u6bcf\u4e00\u6b21 NUMA \u8bbf\u5b58\u3001\u6bcf\u4e00\u6b21\u7f51\u7edc hop \u90fd\u538b\u69a8\u5230\u6781\u9650\u7684\u201c\u7b97\u529b\u5de5\u5320\u7cbe\u795e\u201d\u3002\u5982\u679c\u4f60\u4e5f\u538c\u5026\u4e86\u4e00\u8fb9\u70e7\u94b1\u4e00\u8fb9\u6389\u7ebf\u7684\u65e5\u5b50\uff0c\u4e0d\u59a8\u7528 72 \u5c0f\u65f6\uff0c\u4eb2\u81ea\u9a8c\u8bc1 38 ms \u7684\u9b54\u6cd5\u3002<br \/>\n\u70b9\u51fb\u94fe\u63a5\uff0c\u9886\u53d6\u52a0\u901f\u5238\uff0c\u8ba9\u5c5e\u4e8e\u4f60\u7684 2600 token\/s \u5373\u523b\u53d1\u751f\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u201c\u8bad\u7ec3\u7b97\u529b\u8fc7\u5269\uff0c\u63a8\u7406\u7b97\u529b\u77ed\u7f3a\u3002\u201d \u8fd9\u662f\u82f1\u4f1f\u8fbe 2024 \u5f00\u53d1&hellip;<\/p>\n","protected":false},"author":2,"featured_media":2220,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-2221","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-zixun"],"views":47,"_links":{"self":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts\/2221","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=2221"}],"version-history":[{"count":0,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts\/2221\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/media\/2220"}],"wp:attachment":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/media?parent=2221"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/categories?post=2221"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/tags?post=2221"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}