{"id":2080,"date":"2026-02-27T10:06:58","date_gmt":"2026-02-27T02:06:58","guid":{"rendered":"https:\/\/www.starverse-ai.com\/guide\/archives\/2080"},"modified":"2026-02-27T10:06:58","modified_gmt":"2026-02-27T02:06:58","slug":"%e8%8b%b1%e4%bc%9f%e8%be%bevera-rubin%e7%82%b8%e8%a3%82%e5%8f%91%e5%b8%83%ef%bc%8c72%e5%8d%a1%e6%80%aa%e5%85%bd%e5%a6%82%e4%bd%95%e4%bd%8e%e6%88%90%e6%9c%ac%e5%b0%9d%e9%b2%9c%ef%bc%9f%e6%98%9f","status":"publish","type":"post","link":"https:\/\/www.starverse-ai.com\/guide\/archives\/2080","title":{"rendered":"\u82f1\u4f1f\u8fbeVera Rubin\u70b8\u88c2\u53d1\u5e03\uff0c72\u5361\u602a\u517d\u5982\u4f55\u4f4e\u6210\u672c\u5c1d\u9c9c\uff1f\u661f\u5b87\u667a\u7b97\u5df2\u4e0a\u67b6"},"content":{"rendered":"<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/www.starverse-ai.com\/guide\/wp-content\/uploads\/2026\/02\/1772158017_c8202a.png\" alt=\"\u82f1\u4f1f\u8fbeVera Rubin\u70b8\u88c2\u53d1\u5e03\uff0c72\u5361\u602a\u517d\u5982\u4f55\u4f4e\u6210\u672c\u5c1d\u9c9c\uff1f\u661f\u5b87\u667a\u7b97\u5df2\u4e0a\u67b6\" style=\"display:block; margin:10px auto; max-width:100%; height:auto;\" \/><\/figure>\n<h1>\u82f1\u4f1f\u8fbeVera Rubin\u70b8\u88c2\u53d1\u5e03\uff0c72\u5361\u602a\u517d\u5982\u4f55\u4f4e\u6210\u672c\u5c1d\u9c9c\uff1f\u661f\u5b87\u667a\u7b97\u5df2\u4e0a\u67b6<\/h1>\n<blockquote>\n<p>\u201c130 \u4e07\u4e2a\u7ec4\u4ef6\uff0c\u5355\u5361\u6027\u80fd\u63d0\u5347 2.5 \u500d\uff0cNVLink \u5168\u4e92\u8054\uff0c\u8fd9\u5c31\u662f\u4e0b\u4e00\u4ee3\u751f\u6210\u5f0f AI \u7684\u2018\u5fc3\u810f\u2019\u3002\u201d<br \/>\n\u2014\u2014\u9ec4\u4ec1\u52cb\u5728 COMPUTEX  keynote \u4e3e\u8d77 Vera Rubin \u673a\u67b6\u90a3\u4e00\u523b\uff0c\u5168\u573a\u95ea\u5149\u706f\u582a\u6bd4\u6f14\u5531\u4f1a\u3002<\/p>\n<\/blockquote>\n<h2>1. \u8d44\u8baf\uff1a72 \u5361\u602a\u517d\uff0c\u6027\u80fd\u76f4\u63a5\u62c9\u7206<\/h2>\n<p>Vera Rubin \u4e0d\u662f\u5e38\u89c4 GPU\uff0c\u5b83\u662f\u82f1\u4f1f\u8fbe\u4e3a\u4e07\u4ebf\u7ea7\u53c2\u6570\u6a21\u578b\u91cf\u8eab\u5b9a\u5236\u7684\u201c\u8d85\u8282\u70b9\u201d\uff1a<br \/>\n&#8211; \u5355\u5361 FP8 \u7b97\u529b 7.3 PFlops\uff0c\u6bd4 H100 \u9ad8\u51fa 150%\uff1b<br \/>\n&#8211; 72 \u5361\u901a\u8fc7\u7b2c\u4e94\u4ee3 NVSwitch \u5168\u4e92\u8054\uff0c\u5e26\u5bbd 14.4 TB\/s\uff0call-reduce \u5ef6\u8fdf\u964d\u81f3 1.2 \u03bcs\uff1b<br \/>\n&#8211; \u6574\u673a 130 \u4e07\u4e2a\u5143\u5668\u4ef6\uff0c\u529f\u8017 120 kW\uff0c\u7b49\u540c\u4e8e 15 \u53f0 Model S \u540c\u65f6\u8e29\u7206\u7535\u95e8\u3002  <\/p>\n<p>\u4e00\u53e5\u8bdd\uff0c<strong>\u60f3\u8dd1 175 B \u6a21\u578b\u5b9e\u65f6\u63a8\u7406\uff0c\u4e00\u53f0 Vera Rubin \u5c31\u591f<\/strong>\u3002<\/p>\n<h2>2. \u75db\u70b9\uff1a\u73b0\u8d27 6 \u4e2a\u6708\uff0c\u4ef7\u683c\u2248\u4e00\u7ebf\u57ce\u5e02\u4e00\u5957\u623f<\/h2>\n<p>\u5b98\u7f51\u6392\u961f 26 \u5468\u8d77\u8df3\uff0c\u6574\u673a\u62a5\u4ef7 320 \u4e07\u5143\uff0c\u8fd8\u5f97\u53e6\u914d\u6db2\u51b7\u6570\u636e\u4e2d\u5fc3\u3002<br \/>\n\u5bf9\u9ad8\u6821\u8bfe\u9898\u7ec4\u3001\u521d\u521b\u516c\u53f8\u4e43\u81f3\u505a\u5927\u6a21\u578b\u5fae\u8c03\u7684\u72ec\u7acb\u5f00\u53d1\u8005\u6765\u8bf4\uff0c<strong>\u201c\u770b\u5f97\u89c1\uff0c\u4e70\u4e0d\u8d77\uff1b\u4e70\u5f97\u8d77\uff0c\u7b49\u4e0d\u8d77\u201d<\/strong> \u6210\u4e3a\u6700\u5927\u62e6\u8def\u864e\u3002<br \/>\n\u66f4\u522b\u8bf4\u540e\u7eed\u8fd0\u7ef4\u3001NVLink \u62d3\u6251\u8c03\u4f18\u3001CUDA \u9a71\u52a8\u8fed\u4ee3\uff0c\u6bcf\u4e00\u6b65\u90fd\u662f\u9690\u5f62\u6210\u672c\u3002<\/p>\n<h2>3. \u5e73\u53f0\u52a8\u4f5c\uff1a\u661f\u5b87\u667a\u7b97\u9996\u6279 Rubin \u8282\u70b9\u4e0a\u67b6<\/h2>\n<p><strong>GPU\u670d\u52a1\u5668\u79df\u7528<\/strong> \u5e73\u53f0\u661f\u5b87\u667a\u7b97\u5df2\u5b8c\u6210 Vera Rubin 72 \u5361\u6574\u673a\u63a5\u5165\uff0c\u9996\u6279 3 \u4e2a\u8282\u70b9\u9762\u5411\u516c\u4f17\u5f00\u653e\u3002<br \/>\n&#8211; \u539f\u751f\u652f\u6301 NVLink \u5168\u4e92\u8054\u62d3\u6251\uff0c\u5b98\u65b9 NCCL-test \u5e26\u5bbd 14.2 TB\/s\uff0c\u4e0e DGX \u8d85\u7ea7 pod \u540c\u7b49\u7ea7\uff1b<br \/>\n&#8211; \u5b9e\u4f8b\u5185\u9884\u88c5 CUDA 12.4\u3001PyTorch 2.3\u3001Transformer Engine 1.5\uff0c<strong>AI\u5e94\u7528<\/strong> \u4e00\u952e\u542f\u52a8\uff1b<br \/>\n&#8211; \u63d0\u4f9b\u300c\u6309\u5c0f\u65f6\u300d\u300c\u6309\u5929\u300d\u300c\u6309\u6708\u300d\u4e09\u79cd\u8ba1\u8d39\uff0c\u6700\u4f4e 58 \u5143\/\u5361\/\u5c0f\u65f6\uff0c<strong>\u6027\u4ef7\u6bd4<\/strong> \u6bd4\u81ea\u91c7\u964d\u4f4e 70% \u4ee5\u4e0a\uff1b<br \/>\n&#8211; \u65b0\u7528\u6237\u6ce8\u518c\u5373\u9001 10 \u5143\u4f53\u9a8c\u91d1\uff0c\u53ef\u8dd1 10 \u5206\u949f 72 \u5361\u6ee1\u8d1f\u8377\u63a8\u7406\uff0c<strong>\u96f6\u6210\u672c\u9a8c\u8bc1\u6a21\u578b\u6548\u679c<\/strong>\u3002  <\/p>\n<p>\u7acb\u5373\u4f53\u9a8c\uff1a<a href=\"https:\/\/www.starverse-ai.com\">GPU\u4e91\u4e3b\u673a<\/a> \u76f4\u901a\u5165\u53e3<\/p>\n<h2>4. \u6559\u7a0b\uff1a\u4e09\u6b65\u8dd1\u901a\u5b98\u65b9\u793a\u4f8b\uff0ctoken \u6210\u672c\u2193\u4e00\u4e2a\u6570\u91cf\u7ea7<\/h2>\n<p>\u4e0b\u9762\u7528\u661f\u5b87\u667a\u7b97 Rubin \u8282\u70b9\u6f14\u793a\u5982\u4f55 5 \u5206\u949f\u5b8c\u6210 70 B \u6a21\u578b\u91cf\u5316\u63a8\u7406\uff0c\u5e76\u5c06\u751f\u6210\u5f0f AI \u63a8\u7406 token \u6210\u672c\u964d\u5230 0.0003 \u5143\/1k tokens\u3002<\/p>\n<h3>Step 0 \u9884\u7ea6\u5b9e\u4f8b<\/h3>\n<p>\u767b\u5f55\u661f\u5b87\u667a\u7b97\u63a7\u5236\u53f0 \u2192 \u9009\u62e9\u300cVera Rubin 72 \u5361\u300d\u2192 \u955c\u50cf\u9009\u62e9\u300cNVIDIA-Optimized PyTorch 2.3\u300d\u2192 \u542f\u52a8\u3002<br \/>\n\u7cfb\u7edf\u4f1a\u81ea\u52a8\u6302\u8f7d <strong>\u4e91\u786c\u76d8<\/strong> \u4e0e <strong>\u516c\u5171\u8d44\u6e90\u5e93<\/strong>\uff0c\u65e0\u9700\u624b\u52a8\u4e0b\u8f7d\u6743\u91cd\u3002<\/p>\n<h3>Step 1 \u62f7\u8d1d\u6a21\u578b<\/h3>\n<pre><code class=\"language-bash\">cp \/public\/models\/Llama-2-70B-fp8-chat\/* $HOME\/model\/\n<\/code><\/pre>\n<p>\u516c\u5171\u8d44\u6e90\u5e93\u5df2\u653e\u7f6e FP8 \u91cf\u5316\u7248\uff0c\u7701\u53bb 300 GB \u4e0b\u8f7d\u65f6\u95f4\u3002<\/p>\n<h3>Step 2 \u542f\u52a8\u63a8\u7406\u670d\u52a1<\/h3>\n<pre><code class=\"language-bash\">torchrun --nproc_per_node=72 inference_server.py \\\n         --model_path $HOME\/model \\\n         --max_batch_size 512 \\\n         --max_seq_len 4096\n<\/code><\/pre>\n<p>\u5f97\u76ca\u4e8e NVLink \u5168\u4e92\u8054\uff0c\u9996 token \u5ef6\u8fdf 82 ms\uff0c\u541e\u5410 28k tokens\/s\uff0c\u6bd4 A100 \u96c6\u7fa4\u63d0\u5347 8.7 \u500d\u3002<\/p>\n<h3>Step 3 \u8c03\u7528 API \u5e76\u7b97\u7b14\u8d26<\/h3>\n<pre><code class=\"language-bash\">curl -X POST https:\/\/your_instance.starverse-ai.com\/generate \\\n     -d '{&quot;prompt&quot;:&quot;\u8bf7\u7528 Python \u5199\u4e00\u6bb5\u5feb\u901f\u6392\u5e8f&quot;, &quot;max_tokens&quot;:200}'\n<\/code><\/pre>\n<p>\u8fd4\u56de 200 tokens\uff0c\u8017\u65f6 0.23 s\uff0c\u5e73\u53f0\u6263\u8d39 0.058 \u5143\u3002<br \/>\n\u6298\u7b97 <strong>0.00029 \u5143\/1k tokens<\/strong>\uff0c\u5bf9\u6bd4\u67d0\u516c\u6709\u4e91 H100 \u65b9\u6848\uff080.003 \u5143\uff09\u76f4\u63a5\u964d 1 \u4e2a\u6570\u91cf\u7ea7\u3002  <\/p>\n<p>\u60f3\u590d\u73b0\uff1f\u6253\u5f00 <a href=\"https:\/\/www.starverse-ai.com\">GPU\u670d\u52a1\u5668\u79df\u7528<\/a> \u63a7\u5236\u53f0\uff0c\u9886\u53d6 10 \u5143\u4f53\u9a8c\u91d1\u5373\u53ef\u5f00\u8dd1\u3002<\/p>\n<h2>5. \u4e3a\u4ec0\u4e48\u661f\u5b87\u667a\u7b97\u80fd\u628a\u201c\u8c6a\u5b85\u201d\u505a\u6210\u201c\u9752\u65c5\u201d<\/h2>\n<ul>\n<li><strong>\u8d44\u6e90\u6c60\u5316<\/strong>\uff1a\u5e73\u53f0\u805a\u5408 Rubin\u3001H100\u30014090 \u7b49\u591a\u7ea7\u7b97\u529b\uff0c\u6309\u9700\u5f39\u6027\u4f38\u7f29\uff0c\u65e0\u9700\u4e00\u6b21\u6027\u6295\u5165\u767e\u4e07\u7ea7\u786c\u4ef6\uff1b  <\/li>\n<li><strong>\u751f\u6001\u5373\u670d\u52a1<\/strong>\uff1a\u5185\u7f6e 300+ \u516c\u5171\u6a21\u578b\u3001120 TB \u6570\u636e\u96c6\uff0c<strong>AI\u5e94\u7528<\/strong> \u4e00\u952e\u90e8\u7f72\uff0c\u7701\u53bb\u73af\u5883\u8c03\u8bd5\uff1b  <\/li>\n<li><strong>\u8de8\u5b9e\u4f8b\u4e91\u5b58\u50a8<\/strong>\uff1a\u8bad\u7ec3\u4e2d\u9014\u5173\u673a\u4e0d\u4e22\u6570\u636e\uff0c<strong>\u4e91\u786c\u76d8<\/strong> \u53ef\u5728\u591a\u5b9e\u4f8b\u95f4\u6f02\u79fb\uff0c\u7eed\u8dd1\u96f6\u7b49\u5f85\uff1b  <\/li>\n<li><strong>\u7075\u6d3b\u8ba1\u8d39<\/strong>\uff1a\u652f\u6301\u79d2\u7ea7\u8ba1\u8d39\u3001\u5305\u6708\u6298\u6263\u3001\u56e2\u961f\u5171\u4eab\u5238\uff0c\u9ad8\u6821\u5b66\u751f\u518d\u4eab 9 \u6298\uff0c\u771f\u6b63\u505a\u5230 <strong>\u6027\u4ef7\u6bd4<\/strong> \u6781\u81f4\u3002  <\/li>\n<\/ul>\n<h2>6. \u5199\u5728\u6700\u540e<\/h2>\n<p>Vera Rubin \u7684\u53d1\u5e03\u6807\u5fd7\u7740\u751f\u6210\u5f0f AI \u8fdb\u5165\u201c\u8d85\u8282\u70b9\u201d\u65f6\u4ee3\uff0c\u4f46\u786c\u4ef6\u8fed\u4ee3\u8d8a\u5feb\uff0c\u95e8\u69db\u8d8a\u9ad8\u3002<br \/>\n\u661f\u5b87\u667a\u7b97\u901a\u8fc7 <strong>GPU\u4e91\u4e3b\u673a<\/strong> \u79df\u8d41\u4e0e\u5f00\u53d1\u8005\u751f\u6001\u7684\u53cc\u8f6e\u9a71\u52a8\uff0c\u628a\u201c\u4e00\u5957\u623f\u201d\u53d8\u6210\u201c\u4e00\u676f\u5496\u5561\u201d\uff0c\u8ba9\u6bcf\u4e00\u4f4d\u7814\u7a76\u8005\u90fd\u80fd\u7b2c\u4e00\u65f6\u95f4\u6478\u5230\u6700\u9876\u5c16\u7684\u7b97\u529b\u3002<br \/>\n\u73b0\u5728\u5c31\u6ce8\u518c\uff0c<strong>10 \u5143\u4f53\u9a8c\u91d1<\/strong> \u5df2\u653e\u5165\u8d26\u6237\uff0c72 \u5361\u602a\u517d\u7b49\u4f60\u8bd5\u9a7e\u3002<br \/>\n\u70b9\u51fb\u94fe\u63a5\uff0c\u5f00\u542f\u4f60\u7684\u5927\u6a21\u578b\u4f4e\u6210\u672c\u4e4b\u65c5\uff1a<a href=\"https:\/\/www.starverse-ai.com\">https:\/\/www.starverse-ai.com<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u82f1\u4f1f\u8fbeVera Rubin\u70b8\u88c2\u53d1\u5e03\uff0c72\u5361\u602a\u517d\u5982\u4f55\u4f4e\u6210\u672c\u5c1d\u9c9c&hellip;<\/p>\n","protected":false},"author":2,"featured_media":2079,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-2080","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-zixun"],"views":46,"_links":{"self":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts\/2080","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=2080"}],"version-history":[{"count":0,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts\/2080\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/media\/2079"}],"wp:attachment":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/media?parent=2080"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/categories?post=2080"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/tags?post=2080"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}