{"id":948,"date":"2025-09-17T09:34:05","date_gmt":"2025-09-17T01:34:05","guid":{"rendered":"http:\/\/www.applicationsofllm.com\/?p=948"},"modified":"2025-09-17T09:34:05","modified_gmt":"2025-09-17T01:34:05","slug":"%e8%8b%b1%e4%bc%9f%e8%be%be%e4%b8%8b%e4%b8%80%e4%bb%a3gpu%e7%99%bb%e5%9c%ba%ef%bc%8crubin-cpx%e4%b8%80%e6%ac%a1%e6%8e%a8%e7%90%86%e6%95%b0%e7%99%be%e4%b8%87token%ef%bc%8c%e7%bd%91%e5%8f%8b%ef%bc%9a","status":"publish","type":"post","link":"http:\/\/www.applicationsofllm.com\/index.php\/2025\/09\/17\/%e8%8b%b1%e4%bc%9f%e8%be%be%e4%b8%8b%e4%b8%80%e4%bb%a3gpu%e7%99%bb%e5%9c%ba%ef%bc%8crubin-cpx%e4%b8%80%e6%ac%a1%e6%8e%a8%e7%90%86%e6%95%b0%e7%99%be%e4%b8%87token%ef%bc%8c%e7%bd%91%e5%8f%8b%ef%bc%9a\/","title":{"rendered":"\u82f1\u4f1f\u8fbe\u4e0b\u4e00\u4ee3GPU\u767b\u573a\uff0cRubin CPX\u4e00\u6b21\u63a8\u7406\u6570\u767e\u4e07Token\uff0c\u7f51\u53cb\uff1a\u8fd9\u662f\u5934\u91ce\u517d"},"content":{"rendered":"\n

\u5728 AI \u57fa\u7840\u8bbe\u65bd\u5cf0\u4f1a\u4e0a\uff0c\u82f1\u4f1f\u8fbe\u5ba3\u5e03\u63a8\u51fa\u4e00\u6b3e\u540d\u4e3a Rubin CPX<\/strong>\uff08Rubin Context GPUs\uff09 <\/strong>\u7684\u65b0 GPU\uff0c\u4e13\u4e3a\u8d85\u8fc7 100 \u4e07 token \u7684\u957f\u4e0a\u4e0b\u6587\u63a8\u7406\u800c\u8bbe\u8ba1\u3002<\/p>\n\n\n\n

\u5bf9\u7528\u6237\u800c\u8a00\uff0c\u8fd9\u610f\u5473\u7740\u4ed6\u4eec\u5728\u8f6f\u4ef6\u5f00\u53d1\u3001\u89c6\u9891\u751f\u6210\u7b49\u957f\u4e0a\u4e0b\u6587\u4efb\u52a1\u4e2d\u80fd\u591f\u83b7\u5f97\u66f4\u597d\u7684\u6027\u80fd\u3002<\/p>\n\n\n\n

\u4f8b\u5982\uff0c\u5728\u8f6f\u4ef6\u5f00\u53d1\u4e2d\uff0cAI \u7cfb\u7edf\u5fc5\u987b\u80fd\u591f\u5bf9\u6574\u4e2a\u4ee3\u7801\u5e93\u8fdb\u884c\u63a8\u7406\u3001\u7406\u89e3\u4ed3\u5e93\u7ea7\u4ee3\u7801\u7ed3\u6784\uff0c\u624d\u80fd\u66f4\u597d\u7684\u5e2e\u52a9\u5f00\u53d1\u8005\u3002\u540c\u6837\u5730\uff0c\u957f\u89c6\u9891\u548c\u7814\u7a76\u7c7b\u5e94\u7528\u4e5f\u8981\u6c42\u5728\u6570\u767e\u4e07 token \u8303\u56f4\u5185\u4fdd\u6301\u6301\u7eed\u7684\u8fde\u8d2f\u6027\u548c\u8bb0\u5fc6\u3002<\/p>\n\n\n\n

\u73b0\u5728\uff0c\u968f\u7740 Rubin CPX \u53d1\u5e03\uff0c\u8fd9\u4e9b\u95ee\u9898\u90fd\u80fd\u8fce\u5203\u800c\u89e3\u3002<\/p>\n\n\n\n

\u8fd9\u6b3e\u65b0\u578b GPU\uff08Rubin CPX\uff09  \u5c06\u4e0e NVIDIA Vera CPU \u548c Rubin GPU \u642d\u914d\u4f7f\u7528\uff0c\u5171\u540c\u7ec4\u6210\u5168\u65b0\u7684 NVIDIA Vera Rubin NVL144 CPX \u5e73\u53f0\u3002\u8fd9\u4e00\u96c6\u6210\u5f0f NVIDIA MGX \u7cfb\u7edf\u5728\u5355\u673a\u67b6\u5185\u53ef\u63d0\u4f9b 8 exaflops AI \u7b97\u529b\uff0c\u5176 AI \u6027\u80fd\u662f NVIDIA GB300 NVL72 \u7cfb\u7edf\u7684 7.5 \u500d\uff0c\u5e76\u914d\u5907 100TB \u9ad8\u901f\u5185\u5b58\u548c 1.7 PB\/s\uff08petabytes\uff09\u5185\u5b58\u5e26\u5bbd\u3002<\/p>\n\n\n\n

\u540c\u65f6\uff0cNVIDIA \u8fd8\u5c06\u4e3a\u5df2\u6709 Vera Rubin NVL144 \u7cfb\u7edf\u7684\u5ba2\u6237\u63d0\u4f9b Rubin CPX \u4e13\u7528\u8ba1\u7b97\u6258\u76d8\uff0c\u4ee5\u4fbf\u5728\u73b0\u6709\u5e73\u53f0\u57fa\u7840\u4e0a\u8fdb\u884c\u590d\u7528\u3002<\/p>\n\n\n\n

\"\u56fe\u7247\"\"\u56fe\u7247\"NVIDIA Vera Rubin NVL144 CPX \u673a\u67b6\u4e0e\u8ba1\u7b97\u6258\u76d8\uff0c\u914d\u5907 Rubin CPX\u3001Rubin GPU \u548c Vera CPU<\/sup><\/p>\n\n\n\n

NVIDIA \u521b\u59cb\u4eba\u517c\u9996\u5e2d\u6267\u884c\u5b98\u9ec4\u4ec1\u52cb\u8868\u793a\uff1aVera Rubin \u5e73\u53f0\u5c06\u6807\u5fd7\u7740 AI \u8ba1\u7b97\u524d\u6cbf\u7684\u53c8\u4e00\u6b21\u98de\u8dc3\uff0c\u5b83\u4e0d\u4ec5\u5f15\u5165\u4e86\u4e0b\u4e00\u4ee3 Rubin GPU\uff0c\u8fd8\u5e26\u6765\u4e86\u4e00\u7c7b\u5168\u65b0\u7684\u5904\u7406\u5668 CPX\u3002\u6b63\u5982 RTX \u66fe\u7ecf\u5f7b\u5e95\u6539\u53d8\u56fe\u5f62\u5904\u7406\u548c\u7269\u7406 AI \u4e00\u6837\uff0cRubin CPX \u662f\u9996\u6b3e\u4e13\u4e3a\u5927\u89c4\u6a21\u4e0a\u4e0b\u6587\u6253\u9020\u7684 CUDA GPU\uff0c\u4f7f\u6a21\u578b\u80fd\u591f\u4e00\u6b21\u6027\u5728\u6570\u767e\u4e07 token \u7684\u77e5\u8bc6\u8303\u56f4\u5185\u8fdb\u884c\u63a8\u7406<\/strong>\u3002<\/p>\n\n\n\n

\u5bf9\u4e8e\u8fd9\u6b3e\u65b0\u7684 GPU\uff0c\u5f88\u591a\u4eba\u8ba4\u4e3a\u8fd9\u5c06\u4f1a\u6539\u53d8\u521b\u4f5c\u8005\u7684\u6e38\u620f\u89c4\u5219\u3002<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n
\"\"<\/figure>\n\n\n\n

Rubin CPX \u6280\u672f\u7a81\u7834<\/strong><\/p>\n\n\n\n

\u4f17\u6240\u5468\u77e5\uff0c\u5927\u6a21\u578b\u6b63\u9010\u6b65\u6f14\u53d8\u4e3a\u5177\u5907\u591a\u6b65\u63a8\u7406\u3001\u6301\u4e45\u8bb0\u5fc6\u548c\u957f\u4e0a\u4e0b\u6587\u80fd\u529b\u7684\u667a\u80fd\u4f53\u7cfb\u7edf\uff0c\u4f7f\u5176\u80fd\u591f\u80dc\u4efb\u8f6f\u4ef6\u5f00\u53d1\u3001\u89c6\u9891\u751f\u6210\u4e0e\u6df1\u5ea6\u7814\u7a76\u7b49\u590d\u6742\u4efb\u52a1\u3002<\/p>\n\n\n\n

\u4f46\u8fd9\u4e9b\u5de5\u4f5c\u8d1f\u8f7d\u5bf9\u57fa\u7840\u8bbe\u65bd\u63d0\u51fa\u4e86\u524d\u6240\u672a\u6709\u7684\u8981\u6c42\uff0c\u5728\u8ba1\u7b97\u3001\u5b58\u50a8\u548c\u7f51\u7edc\u65b9\u9762\u5e26\u6765\u4e86\u5168\u65b0\u6311\u6218\uff0c\u8feb\u5207\u9700\u8981\u4ece\u6839\u672c\u4e0a\u91cd\u65b0\u601d\u8003\u63a8\u7406\u7684\u89c4\u6a21\u5316\u4e0e\u4f18\u5316\u65b9\u5f0f\u3002<\/p>\n\n\n\n

\u4e3a\u6b64\uff0c\u82f1\u4f1f\u8fbe SMART \u6846\u67b6\u63d0\u4f9b\u4e86\u4e00\u6761\u53ef\u884c\u65b9\u6848\u3002\u8be5\u6846\u67b6\u5f3a\u8c03\u91c7\u7528\u5168\u6808\u89e3\u8026\u5f0f\u57fa\u7840\u8bbe\u65bd\uff0c\u5b9e\u73b0\u8ba1\u7b97\u4e0e\u5185\u5b58\u8d44\u6e90\u7684\u9ad8\u6548\u8c03\u914d\u3002\u901a\u8fc7 Blackwell \u67b6\u6784 GB200 NVL72 \u5e73\u53f0\u4e0e NVFP4 \u4f4e\u7cbe\u5ea6\u63a8\u7406\u6280\u672f\u7684\u7ed3\u5408\uff0c\u8f85\u4ee5 TensorRT-LLM\u3001Dynamo \u7b49\u5f00\u6e90\u8f6f\u4ef6\uff0c\u4f7f\u5f97\u6574\u4e2a AI \u9886\u57df\u7684\u63a8\u7406\u6027\u80fd\u5927\u5927\u63d0\u5347\u3002<\/p>\n\n\n\n

\u53e6\u4e00\u65b9\u9762\uff0c\u63a8\u7406\u8fc7\u7a0b\u5305\u542b\u4e24\u4e2a\u622a\u7136\u4e0d\u540c\u7684\u9636\u6bb5\uff1a\u4e0a\u4e0b\u6587\u5904\u7406\u4e0e\u5185\u5bb9\u751f\u6210\uff0c\u6bcf\u4e2a\u9636\u6bb5\u5bf9\u57fa\u7840\u8bbe\u65bd\u7684\u9700\u6c42\u5b58\u5728\u672c\u8d28\u5dee\u5f02\u3002<\/p>\n\n\n\n

\u4e0a\u4e0b\u6587\u9636\u6bb5\u5c5e\u4e8e\u8ba1\u7b97\u5bc6\u96c6\u578b\uff0c\u9700\u8981\u9ad8\u541e\u5410\u91cf\u5904\u7406\u80fd\u529b\u6765\u6444\u5165\u548c\u5206\u6790\u6d77\u91cf\u8f93\u5165\u6570\u636e\uff0c\u4ee5\u4ea7\u751f\u9996\u4e2a\u8f93\u51fa token\u3002<\/p>\n\n\n\n

\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u751f\u6210\u9636\u6bb5\u5219\u53d7\u9650\u4e8e\u5185\u5b58\u5e26\u5bbd\uff0c\u4f9d\u8d56 NVLink \u7b49\u9ad8\u901f\u4e92\u8fde\u6280\u672f\u5b9e\u73b0\u5feb\u901f\u5185\u5b58\u4f20\u8f93\uff0c\u4ee5\u7ef4\u6301\u9010 token \u8f93\u51fa\u7684\u6027\u80fd\u8868\u73b0\u3002<\/p>\n\n\n\n

\u89e3\u8026\u5f0f\u63a8\u7406\u67b6\u6784\u4f7f\u8fd9\u4e24\u4e2a\u9636\u6bb5\u80fd\u591f\u72ec\u7acb\u5904\u7406\uff0c\u5b9e\u73b0\u5bf9\u8ba1\u7b97\u4e0e\u5185\u5b58\u8d44\u6e90\u7684\u7cbe\u51c6\u4f18\u5316\u3002\u8fd9\u79cd\u67b6\u6784\u53d8\u9769\u663e\u8457\u63d0\u5347\u4e86\u7cfb\u7edf\u541e\u5410\u91cf\uff0c\u964d\u4f4e\u4e86\u5ef6\u8fdf\uff0c\u5e76\u5168\u9762\u63d0\u9ad8\u8d44\u6e90\u5229\u7528\u7387\u3002<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

\u7136\u800c\uff0c\u89e3\u8026\u5f0f\u67b6\u6784\u4e5f\u5e26\u6765\u4e86\u65b0\u7684\u590d\u6742\u6027\uff0c\u9700\u8981\u5728\u4f4e\u5ef6\u8fdf KV \u7f13\u5b58\u4f20\u8f93\u3001LLM \u611f\u77e5\u8def\u7531\u4ee5\u53ca\u9ad8\u6548\u5185\u5b58\u7ba1\u7406\u4e4b\u95f4\u5b9e\u73b0\u7cbe\u786e\u534f\u8c03\u3002<\/p>\n\n\n\n

NVIDIA Dynamo\uff08\u4e00\u4e2a\u5f00\u6e90\u3001\u4f4e\u5ef6\u8fdf\u7684\u6a21\u5757\u5316\u63a8\u7406\u6846\u67b6\uff09 \u4f5c\u4e3a\u8fd9\u4e9b\u7ec4\u4ef6\u7684\u7f16\u6392\u5c42\uff0c\u5728\u6700\u65b0\u7684 MLPerf Inference \u7ed3\u679c\u4e2d\u53d1\u6325\u4e86\u5173\u952e\u4f5c\u7528\u3002\u501f\u52a9 Dynamo \u5728 GB200 NVL72 \u4e0a\u7684\u89e3\u8026\u63a8\u7406\uff0c\u521b\u9020\u4e86\u65b0\u7684\u6027\u80fd\u7eaa\u5f55\u3002<\/p>\n\n\n\n

\u4e3a\u4e86\u5145\u5206\u53d1\u6325\u89e3\u8026\u63a8\u7406\u7684\u4f18\u52bf\uff0c\u5c24\u5176\u662f\u5728\u7b97\u529b\u5bc6\u96c6\u7684\u4e0a\u4e0b\u6587\u9636\u6bb5\u3002<\/p>\n\n\n\n

\u6b63\u662f\u5728\u8fd9\u6837\u7684\u80cc\u666f\u4e0b\uff0cNVIDIA \u63a8\u51fa\u4e86 Rubin CPX GPU\uff1a\u4e13\u4e3a\u9ad8\u4ef7\u503c\u957f\u4e0a\u4e0b\u6587\u63a8\u7406\u5de5\u4f5c\u8d1f\u8f7d\u8bbe\u8ba1\u7684\u89e3\u51b3\u65b9\u6848<\/strong>\uff0c\u80fd\u591f\u63d0\u4f9b\u9ad8\u541e\u5410\u6027\u80fd\uff0c\u5e76\u4e0e\u89e3\u8026\u5f0f\u57fa\u7840\u8bbe\u65bd\u5b9e\u73b0\u65e0\u7f1d\u96c6\u6210\u3002<\/p>\n\n\n\n

\u5728\u6280\u672f\u4e0a\uff0c\u57fa\u4e8e NVIDIA Rubin \u67b6\u6784\uff0cRubin CPX GPU \u91c7\u7528\u4e86\u5177\u5907\u9ad8\u6027\u4ef7\u6bd4\u7684\u5355\u7247\u8bbe\u8ba1\uff0c\u5185\u7f6e\u5f3a\u5927\u7684 NVFP4 \u8ba1\u7b97\u8d44\u6e90\uff0c\u5e76\u9488\u5bf9 AI \u63a8\u7406\u4efb\u52a1\u8fdb\u884c\u4e86\u4f18\u5316\uff0c\u4ee5\u5b9e\u73b0\u6781\u9ad8\u7684\u6027\u80fd\u4e0e\u80fd\u6548\u8868\u73b0\u3002<\/p>\n\n\n\n

\u5177\u4f53\u800c\u8a00\uff0cRubin CPX \u91c7\u7528 NVFP4 <\/strong>\u7cbe\u5ea6\uff0c\u53ef\u63d0\u4f9b\u9ad8\u8fbe 30 petaflops<\/strong> \u7b97\u529b\uff0c\u5728\u6027\u80fd\u4e0e\u7cbe\u5ea6\u4e0a\u5747\u8fbe\u5230\u4e1a\u754c\u9886\u5148\u6c34\u5e73\uff0c\u5e76\u914d\u5907 128GB \u9ad8\u6027\u4ef7\u6bd4 GDDR7 <\/strong>\u5185\u5b58\uff0c\u80fd\u591f\u52a0\u901f\u6700\u82db\u523b\u7684\u4e0a\u4e0b\u6587\u8ba1\u7b97\u4efb\u52a1\u3002\u6b64\u5916\uff0c\u4e0e NVIDIA GB300 NVL72 \u7cfb\u7edf\u76f8\u6bd4\uff0cRubin CPX \u7684\u6ce8\u610f\u529b\u5904\u7406\u80fd\u529b\u63d0\u5347\u4e86 3 \u500d\uff0c\u663e\u8457\u589e\u5f3a\u4e86 AI \u6a21\u578b\u5728\u957f\u4e0a\u4e0b\u6587\u5e8f\u5217\u5904\u7406\u4e2d\u7684\u901f\u5ea6\u4e0e\u7a33\u5b9a\u6027\u3002<\/p>\n\n\n\n

\"\"<\/figure>\n","protected":false},"excerpt":{"rendered":"

\u5728 AI \u57fa\u7840\u8bbe\u65bd\u5cf0\u4f1a\u4e0a\uff0c\u82f1\u4f1f\u8fbe\u5ba3\u5e03\u63a8\u51fa\u4e00\u6b3e\u540d\u4e3a Rubin CPX\uff08Rubin Context G […]<\/p>\n","protected":false},"author":1,"featured_media":953,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[7],"tags":[],"class_list":["post-948","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-7"],"_links":{"self":[{"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/posts\/948","targetHints":{"allow":["GET"]}}],"collection":[{"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/comments?post=948"}],"version-history":[{"count":1,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/posts\/948\/revisions"}],"predecessor-version":[{"id":954,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/posts\/948\/revisions\/954"}],"wp:featuredmedia":[{"embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/media\/953"}],"wp:attachment":[{"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/media?parent=948"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/categories?post=948"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/tags?post=948"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}