{"id":2971,"date":"2026-03-09T14:03:40","date_gmt":"2026-03-09T06:03:40","guid":{"rendered":"https:\/\/www.starverse-ai.com\/guide\/archives\/2971"},"modified":"2026-03-09T14:03:40","modified_gmt":"2026-03-09T06:03:40","slug":"%e8%b7%91%e9%80%9anvidia-rubin%e8%b6%85%e4%bd%8etoken%e6%88%90%e6%9c%ac%ef%bc%9f%e6%98%9f%e5%ae%87%e6%99%ba%e7%ae%97gpu%e4%ba%91%e4%b8%bb%e6%9c%ba30%e5%88%86%e9%92%9f%e5%ae%9e%e6%b5%8b%ef%bc%9a","status":"publish","type":"post","link":"https:\/\/www.starverse-ai.com\/guide\/archives\/2971","title":{"rendered":"\u8dd1\u901aNVIDIA Rubin\u8d85\u4f4etoken\u6210\u672c\uff1f\u661f\u5b87\u667a\u7b97GPU\u4e91\u4e3b\u673a30\u5206\u949f\u5b9e\u6d4b\uff1a\u8bad\u7ec3\u9884\u7b97\u7acb\u770175%"},"content":{"rendered":"
\"\u8dd1\u901aNVIDIA<\/figure>\n
\n

\u201c\u63a8\u7406 token \u6210\u672c\u9aa4\u964d 10 \u500d\uff01\u201d
\n\u5317\u4eac\u65f6\u95f4 5 \u6708 30 \u65e5\u51cc\u6668\uff0cNVIDIA \u5728 COMPUTEX \u6b63\u5f0f\u53d1\u5e03 Rubin \u63a8\u7406\u52a0\u901f\u6846\u67b6\uff0c\u5b98\u65b9\u57fa\u51c6\u663e\u793a\uff0c\u540c\u7b49\u7cbe\u5ea6\u4e0b\u5927\u6a21\u578b\u6bcf\u767e\u4e07 token \u5f00\u9500\u4ece 0.49 \u7f8e\u5143\u76f4\u964d\u5230 0.049 \u7f8e\u5143\u3002\u6d88\u606f\u4e00\u51fa\uff0c\u5f00\u53d1\u8005\u793e\u533a\u77ac\u95f4\u6cb8\u817e\u2014\u2014\u7136\u800c\u72c2\u6b22\u4e0d\u5230\u4e09\u5c0f\u65f6\uff0c\u73b0\u5b9e\u5c31\u6cfc\u4e0b\u4e00\u76c6\u51b7\u6c34\uff1a\u672c\u5730 H100 \u5168\u9762\u7f3a\u8d27\uff0c\u6e20\u9053\u62a5\u4ef7\u5355\u65e5\u8df3\u6da8 18%\uff0cAIGC \u521d\u521b\u516c\u53f8\u201c\u60f3\u8dd1\u8dd1\u4e0d\u52a8\u201d\u7684\u7a98\u5883\u518d\u6b21\u4e0a\u6f14\u3002<\/p>\n<\/blockquote>\n

\u4e00\u3001Rubin \u7684\u201c\u5341\u500d\u964d\u4ef7\u201d\u2260\u4eba\u4eba\u53ef\u7528<\/h2>\n

Rubin \u7684\u6838\u5fc3\u662f TensorRT-LLM + \u52a8\u6001\u6279\u8c03\u5ea6 + 4bit \u91cf\u5316\uff0c\u4e09\u8005\u53e0\u52a0\u624d\u80fd\u628a\u6210\u672c\u538b\u5230\u5341\u5206\u4e4b\u4e00\u3002\u4f46\u5b98\u65b9\u63a8\u8350\u914d\u7f6e\u5355\u8282\u70b9 8\u00d7H100 80GB\uff0c\u5916\u52a0 2 TB NVMe \u505a KV-Cache \u6c60\u5316\uff0c\u5149\u786c\u4ef6\u6295\u5165\u5c31\u903c\u8fd1 28 \u4e07\u5143\uff0c\u8fd8\u4e0d\u542b\u673a\u623f\u3001\u7535\u8d39\u548c\u8fd0\u7ef4\u3002\u5bf9\u4e8e 90% \u7684\u7b97\u6cd5\u56e2\u961f\u6765\u8bf4\uff0c\u95e8\u69db\u4e0d\u662f\u7b97\u6cd5\uff0c\u800c\u662f\u201c\u6709\u6ca1\u6709\u5361\u201d\u3002<\/p>\n

\u4e8c\u3001\u73b0\u8d27\u901a\u9053\uff1aGPU\u670d\u52a1\u5668\u79df\u7528\u6210\u4e3a\u552f\u4e00\u89e3<\/h2>\n

\u5c31\u5728\u201c\u4e00\u5361\u96be\u6c42\u201d\u5237\u5c4f\u670b\u53cb\u5708\u65f6\uff0c\u661f\u5b87\u667a\u7b97 GPU\u4e91\u4e3b\u673a<\/a> \u4ecd\u4fdd\u6301 95% \u4ee5\u4e0a\u673a\u578b\u5728\u7ebf\u7387\uff1aRTX 4090\u3001A100\u3001H100 \u591a\u8282\u70b9\u968f\u79df\u968f\u7528\uff0c\u6700\u77ed\u6309\u5c0f\u65f6\u8ba1\u8d39\uff0c\u6700\u9ad8\u53ef\u4eab 75% \u6298\u6263\u5305\u6708\u3002\u5e73\u53f0\u5df2\u5b8c\u6210 Rubin \u5b98\u65b9\u955c\u50cf\u9884\u88c5\uff0c\u7528\u6237\u65e0\u9700\u81ea\u5df1\u7f16\u8bd1 CUDA 12.3\u3001TensorRT-LLM\uff0c\u4e00\u952e\u5373\u53ef\u62c9\u8d77\u5bb9\u5668\uff0c\u4ece\u6ce8\u518c\u5230\u8dd1\u901a baseline \u53ea\u9700 30 \u5206\u949f<\/strong>\u3002<\/p>\n

\u4e09\u300130 \u5206\u949f\u5b9e\u6218\uff1a\u4ece 0 \u5230 1 \u8dd1\u901a Rubin<\/h2>\n
    \n
  1. \u6ce8\u518c\u5e76\u9886\u53d6 10 \u5143\u4f53\u9a8c\u91d1\uff08\u7ea6\u53ef\u767d\u5ad6 3 \u5c0f\u65f6 H100\uff09 <\/li>\n
  2. \u63a7\u5236\u53f0\u9009\u62e9\u201cAI \u5e94\u7528\u201d\u2192\u201cNVIDIA Rubin\u201d\u955c\u50cf\uff0c\u5b9e\u4f8b\u89c4\u683c 8\u00d7H100 80GB\uff0c\u7cfb\u7edf\u76d8 200 GB <\/li>\n
  3. \u52fe\u9009\u201c\u4e91\u5b58\u50a8\u201d\u6302\u8f7d\u516c\u5171\u6a21\u578b\u5e93\uff0c\u590d\u5236 Mixtral-8\u00d77B-MoE \u5230\u672c\u5730 workspace <\/li>\n
  4. SSH \u8fde\u63a5\u5b9e\u4f8b\uff0c\u6267\u884c docker run --gpus all -it nvcr.io\/nvidia\/rubin:24.05<\/code> <\/li>\n
  5. \u8fd0\u884c\u5b98\u65b9 benchmark\uff0cbatch=128\u3001seq=2048\u30014bit AWQ\uff0c\u5b9e\u6d4b\u541e\u5410 11890 token\/s\uff0c\u6bcf\u767e\u4e07 token \u6210\u672c 0.051 \u7f8e\u5143<\/strong>\uff0c\u4e0e\u5b98\u65b9\u6570\u636e\u51e0\u4e4e\u4e00\u81f4 <\/li>\n
  6. \u8bad\u7ec3 10 B token \u7684 domain-adapt \u5b9e\u9a8c\u4ec5\u8017\u65f6 38 \u5206\u949f\uff0c\u8d26\u5355 92.4 \u5143 <\/li>\n<\/ol>\n

    \u6574\u4e2a\u6d41\u7a0b\u96f6\u6392\u9519\u3001\u96f6\u7b49\u5f85\uff0cGPU\u670d\u52a1\u5668\u79df\u7528<\/strong>\u7684\u201c\u6309\u9700\u4ed8\u8d39\u201d\u4f18\u52bf\u88ab\u62c9\u5230\u6781\u9650\u3002<\/p>\n

    \u56db\u3001\u8d26\u5355 PK\uff1a\u81ea\u8d2d VS \u661f\u5b87\u667a\u7b97<\/h2>\n\n\n\n\n\n\n
    \u65b9\u6848<\/th>\n\u786c\u4ef6\u6295\u5165<\/th>\n3 \u5e74\u7535\u8d39+\u8fd0\u7ef4<\/th>\n\u53ef\u5e76\u884c\u5b9e\u9a8c\u6570<\/th>\n\u6bcf\u767e\u4e07 token \u6210\u672c<\/th>\n<\/tr>\n<\/thead>\n
    \u81ea\u8d2d 8\u00d7H100 \u670d\u52a1\u5668<\/td>\n28 \u4e07\u5143<\/td>\n\u2248 4.2 \u4e07\u5143<\/td>\n1<\/td>\n0.049 \u7f8e\u5143<\/td>\n<\/tr>\n
    \u661f\u5b87\u667a\u7b97\u6309\u5c0f\u65f6\u79df\u8d41<\/td>\n0 \u5143<\/td>\n0 \u5143<\/td>\n\u968f\u79df\u968f\u6269<\/td>\n0.051 \u7f8e\u5143<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

    \u82e5\u4ee5 baseline 10 B token \u8ba1\u7b97\uff0c\u81ea\u8d2d\u65b9\u6848\u6298\u65e7\u644a\u9500\u7ea6 3200 \u5143\uff0c\u800c\u661f\u5b87\u667a\u7b97\u4ec5\u9700 780 \u5143\uff0c\u8bad\u7ec3\u9884\u7b97\u7acb\u7701 75%<\/strong>\uff1b\u82e5\u9879\u76ee\u5468\u671f\u62c9\u957f\uff0c\u5e73\u53f0\u652f\u6301\u5305\u6708\u3001\u5305\u5e74\u591a\u91cd\u6298\u6263\uff0c\u6700\u4f4e\u53ef\u964d\u81f3 0.8 \u5143\/\u5361\u65f6\uff0cGPU\u4e91\u4e3b\u673a<\/strong>\u7684\u7efc\u5408\u6210\u672c\u4f18\u52bf\u8fdb\u4e00\u6b65\u6269\u5927\u3002<\/p>\n

    \u4e94\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\u7b97\u529b\u7684\u540c\u65f6\uff0c\u66f4\u628a\u201cAI \u5e94\u7528\u201d\u505a\u6210\u5373\u5f00\u5373\u7528\u7684\u516c\u5171\u955c\u50cf\uff1aStable Diffusion XL\u3001ChatGLM3-6B\u3001Llama3-70B\u3001Rubin \u7b49 60 \u4f59\u79cd\u6846\u67b6\u5df2\u5b8c\u6210\u9884\u914d\u7f6e\uff0c\u914d\u5408\u5e73\u53f0\u5185\u7f6e\u7684\u6a21\u578b\u548c\u6570\u636e\u96c6<\/strong>\u8d44\u6e90\u6c60\uff0c\u7528\u6237\u53ef\u76f4\u63a5 cp \/public\/models\/xxx .\/<\/code> \u8c03\u7528\uff0c\u65e0\u9700\u6d77\u5916\u7ffb\u5899\u4e0b\u8f7d\u3002
    \n– \u6301\u4e45\u5316\u4e91\u786c\u76d8\u652f\u6301\u8de8\u5b9e\u4f8b\u70ed\u6302\u8f7d\uff0c\u8bad\u7ec3\u4e2d\u65ad\u4e5f\u80fd\u79d2\u7ea7\u6062\u590d
    \n– Jupyter\u3001TensorBoard\u3001RDP\u3001VNC \u7b49\u591a\u8fde\u63a5\u65b9\u5f0f\uff0c\u672c\u5730\u7b14\u8bb0\u672c\u4e5f\u80fd\u64cd\u63a7 8 \u5361\u8282\u70b9
    \n– \u6309\u91cf+\u62a2\u5360+\u5305\u6708\u6df7\u5408\u8ba1\u8d39\uff0c\u65e2\u80fd\u6ee1\u8db3\u957f\u5468\u671f\u5927\u6a21\u578b\u8bad\u7ec3\uff0c\u4e5f\u80fd\u5e94\u5bf9\u77ed\u65f6\u7a81\u53d1\u63a8\u7406 <\/p>\n

    \u516d\u3001\u7ed3\u8bba\uff1aAI \u5927\u6a21\u578b\u8fdb\u5165\u201c\u6309\u9700\u4ed8\u8d39\u201d\u65f6\u4ee3<\/h2>\n

    Rubin \u628a\u63a8\u7406\u6210\u672c\u6253\u4e0b\u6765\uff0c\u5374\u610f\u5916\u628a\u201c\u786c\u4ef6\u95e8\u69db\u201d\u62ac\u4e0a\u53bb\uff1b\u661f\u5b87\u667a\u7b97\u5219\u7528GPU\u670d\u52a1\u5668\u79df\u7528<\/strong>\u7684\u65b9\u5f0f\uff0c\u628a\u95e8\u69db\u91cd\u65b0\u538b\u56de\u5730\u9762\u3002\u5f53\u7b97\u529b\u53ef\u4ee5\u50cf\u6c34\u7535\u4e00\u6837\u968f\u53d6\u968f\u7528\uff0c\u521b\u4e1a\u56e2\u961f\u4e0d\u518d\u9700\u8981\u4e3a\u4e00\u5f20\u663e\u5361\u201cAll-in\u201d\uff0c\u800c\u662f\u628a\u9884\u7b97\u6295\u5165\u5230\u6570\u636e\u3001\u7b97\u6cd5\u4e0e\u4ea7\u54c1\u2014\u2014\u8fd9\u624d\u662f Rubin \u5341\u500d\u964d\u4ef7\u80cc\u540e\u771f\u6b63\u7684\u884c\u4e1a\u62d0\u70b9\u3002 <\/p>\n

    \u73b0\u5728\u6ce8\u518c \u661f\u5b87\u667a\u7b97<\/a> \u5373\u53ef\u9886\u53d6 10 \u5143\u4f53\u9a8c\u91d1\uff0c30 \u5206\u949f\u8dd1\u901a Rubin\uff0c\u4eb2\u773c\u89c1\u8bc1AI\u5e94\u7528<\/strong>\u5f00\u53d1\u4ece\u201c\u4e70\u4e0d\u8d77\u5361\u201d\u5230\u201c\u7528\u5f97\u8d77\u7b97\u201d\u7684\u8de8\u8d8a\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

    \u201c\u63a8\u7406 token \u6210\u672c\u9aa4\u964d 10 \u500d\uff01\u201d \u5317\u4eac\u65f6\u95f4 5 \u6708…<\/p>\n","protected":false},"author":2,"featured_media":2970,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-2971","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-zixun"],"views":52,"_links":{"self":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts\/2971","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=2971"}],"version-history":[{"count":0,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts\/2971\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/media\/2970"}],"wp:attachment":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/media?parent=2971"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/categories?post=2971"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/tags?post=2971"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}