{"id":80123,"date":"2024-08-29T14:42:23","date_gmt":"2024-08-29T06:42:23","guid":{"rendered":"https:\/\/www.ioiotimes.com\/?p=80123"},"modified":"2024-08-29T14:43:59","modified_gmt":"2024-08-29T06:43:59","slug":"nvidia-blackwell%e5%9c%a8mlperf-inference%e7%9a%84%e5%9f%ba%e6%ba%96%e6%b8%ac%e8%a9%a6%e6%b4%bb%e5%8b%95%e4%b8%ad%e9%a6%96%e6%ac%a1%e4%ba%ae%e7%9b%b8%ef%bc%8c%e7%82%ba%e7%94%9f%e6%88%90%e5%bc%8fai","status":"publish","type":"post","link":"https:\/\/www.ioiotimes.com\/?p=80123","title":{"rendered":"NVIDIA Blackwell\u5728MLPerf Inference\u7684\u57fa\u6e96\u6e2c\u8a66\u6d3b\u52d5\u4e2d\u9996\u6b21\u4eae\u76f8\uff0c\u70ba\u751f\u6210\u5f0fAI\u6a39\u7acb\u5168\u65b0\u6a19\u6e96"},"content":{"rendered":"\n

\u4f7f\u7528NVIDIA Blackwell GPU\u9032\u884c\u7684\u9996\u6b21\u6e2c\u8a66\u7d50\u679c\uff0c\u5728Llama 2 70B\u4e0a\u5c07\u63a8\u8ad6\u6548\u80fd\u63d0\u9ad8\u56db\u500d\uff0c\u800cNVIDIA Hopper\u67b6\u69cb\u4e5f\u5728\u5404\u7522\u696dAI\u57fa\u6e96\u6e2c\u8a66\u4e2d\u6709\u8457\u66f4\u70ba\u512a\u7570\u7684\u8868\u73fe<\/h3>\n\n\n\n
\"\"<\/figure>\n\n\n\n

<\/p>\n\n\n\n

\u5728\u4f01\u696d\u722d\u76f8\u63a1\u7528\u751f\u6210\u5f0f\u4eba\u5de5\u667a\u6167\uff08AI\uff09\u8207\u9678\u7e8c\u63a8\u51fa\u5404\u9805\u65b0\u670d\u52d9\u7684\u60c5\u6cc1\u4e0b\uff0c\u5c0d\u8cc7\u6599\u4e2d\u5fc3\u57fa\u790e\u8a2d\u65bd\u7684\u9700\u6c42\u4e5f\u5275\u4e0b\u65b0\u9ad8\u3002\u8a13\u7df4\u5927\u578b\u8a9e\u8a00\u6a21\u578b<\/a>\uff08LLM\uff09\u662f\u500b\u96e3\u984c\uff0c\u4e0d\u904e\u5373\u6642\u63d0\u4f9b\u7531 LLM \u652f\u63f4\u7684\u670d\u52d9\u540c\u6a23\u4e0d\u6613\u3002<\/p>\n\n\n\n

\u5728\u6700\u65b0\u4e00\u8f2a\u7684 MLPerf \u7522\u696d\u57fa\u6e96\uff1aInference v4.1 \u4e2d\uff0cNVIDIA \u5e73\u53f0\u5728\u5404\u9805\u8cc7\u6599\u4e2d\u5fc3\u6e2c\u8a66\u9805\u76ee\u88e1\u7686\u5beb\u4e0b\u9818\u5148\u6210\u7e3e\u3002\u5373\u5c07\u63a8\u51fa\u7684 NVIDIA Blackwell<\/a> \u5e73\u53f0\u7531\u65bc\u4f7f\u7528\u7b2c\u4e8c\u4ee3 Transformer\u5f15\u64ce<\/a>\u8207 FP4 Tensor \u6838\u5fc3\uff0c\u5728\u8655\u7406 MLPerf \u6700\u5927\u7684 LLM \u5de5\u4f5c\u91cf Llama 2 70B \u4e4b\u969b\uff0c\u9996\u6b21\u6e2c\u8a66\u7684\u7d50\u679c\u986f\u793a\u5176\u6548\u80fd\u8f03 NVIDIA H100 Tensor Core GPU<\/a>\u67b6\u69cb\u9ad8\u51fa\u56db\u500d\u4e4b\u591a\u3002<\/p>\n\n\n\n

NVIDIA H200 Tensor \u6838\u5fc3 GPU<\/a> \u5728\u8cc7\u6599\u4e2d\u5fc3\u985e\u5225\u7684\u5404\u9805\u6e2c\u8a66\u88e1\u5747\u6709\u8457\u51fa\u8272\u8868\u73fe\uff0c\u5305\u62ec\u57fa\u6e96\u4e2d\u6700\u65b0\u52a0\u5165\u6709\u8457 467 \u5104\u500b\u53c3\u6578\u3001\u6bcf\u500b token \u6709 129 \u5104\u500b\u6d3b\u8e8d\u53c3\u6578\u7684 Mixtral 8x7B \u6df7\u5408\u5c08\u5bb6\uff08MoE\uff09LLM\u3002<\/p>\n\n\n\n

MoE \u6a21\u578b\u80fd\u5920\u5728\u55ae\u4e00\u90e8\u7f72\u4e2d\u56de\u7b54\u5404\u5f0f\u5404\u6a23\u7684\u554f\u984c\u8207\u57f7\u884c\u66f4\u591a\u4e0d\u540c\u7684\u4efb\u52d9\uff0c\u8b93\u4f7f\u7528\u8005\u6709\u66f4\u591a\u7a2e\u90e8\u7f72 LLM \u7684\u65b9\u5f0f\u800c\u5927\u53d7\u6b61\u8fce\u3002MoE \u7684\u6548\u7387\u4e5f\u66f4\u9ad8\uff0c\u56e0\u5176\u6bcf\u6b21\u9032\u884c\u63a8\u8ad6\u4f5c\u696d\u53ea\u8981\u555f\u52d5\u5e7e\u500b\u5c08\u5bb6\uff0c\u4ee3\u8868\u63d0\u4f9b\u7d50\u679c\u7684\u901f\u5ea6\u6bd4\u985e\u4f3c\u5927\u5c0f\u7684\u5bc6\u96c6\u6a21\u578b\u8981\u5feb\u4e0a\u4e0d\u5c11\u3002<\/p>\n\n\n\n

LLM \u6301\u7e8c\u6210\u9577\uff0c\u8b93\u4f7f\u7528\u8005\u9700\u8981\u6709\u66f4\u591a\u904b\u7b97\u80fd\u529b\u4f86\u8655\u7406\u63a8\u8ad6\u8acb\u6c42\u3002\u591a GPU \u904b\u7b97\u7684\u80fd\u529b\u5df2\u6210\u70ba\u5fc5\u9808\uff0c\u4ee5\u6eff\u8db3\u7576\u524d\u904b\u884c LLM \u7684\u5373\u6642\u5ef6\u9072\u9700\u6c42\uff0c\u9084\u6709\u76e1\u91cf\u8b93\u6700\u591a\u7684\u4f7f\u7528\u8005\u4eab\u53d7\u9019\u9805\u670d\u52d9\u3002NVIDIA NVLink \u548c NVSwitch \u53ef\u5728\u57fa\u65bc NVIDIA Hopper \u67b6\u69cb\u7684 GPU \u4e4b\u9593\u4ee5\u9ad8\u983b\u5bec\u6e9d\u901a\uff0c\u4e26\u70ba\u7576\u4eca\u5373\u6642\u3001\u5177\u6210\u672c\u6548\u76ca\u7684\u5927\u578b\u6a21\u578b\u63a8\u8ad6\u4f5c\u696d\u5e36\u4f86\u986f\u8457\u512a\u52e2\u3002\u800c Blackwell \u5e73\u53f0\u52a0\u4e0a\u6709\u8457 72 \u500b GPU \u66f4\u5927\u898f\u6a21\u7684 NVLink \u7db2\u57df\uff0c\u5c07\u9032\u4e00\u6b65\u64f4\u5c55 NVLink Switch \u7684\u529f\u80fd\u3002<\/p>\n\n\n\n

\u9664\u4e86 NVIDIA \u6709\u63d0\u51fa\u6e2c\u8a66\u8cc7\u6599\uff0c\u83ef\u78a9\u3001\u601d\u79d1\u3001\u6234\u723e\u79d1\u6280\u3001\u5bcc\u58eb\u901a\u3001\u6280\u92fc\u79d1\u6280\u3001\u6167\u8207\u79d1\u6280\u3001\u77bb\u535a\u7db2\u8def\u3001\u806f\u60f3\u3001\u96f2\u9054\u79d1\u6280\u548c\u7f8e\u8d85\u5fae\u7b49\u5341\u5bb6 NVIDIA \u7684\u5408\u4f5c\u5925\u4f34\uff0c\u540c\u6a23\u90fd\u6709\u63d0\u51fa\u53ef\u9760\u7684 MLPerf Inference \u6e2c\u8a66\u8cc7\u6599\uff0c\u7a81\u986f\u51fa NVIDIA \u5e73\u53f0\u5ee3\u6cdb\u7684\u53ef\u4f9b\u6027\u3002<\/p>\n\n\n\n

<\/p>\n\n\n\n

\u5805\u6301\u4e0d\u61c8\u9032\u884c\u8edf\u9ad4\u5275\u65b0<\/strong><\/strong><\/p>\n\n\n\n

NVIDIA \u5e73\u53f0\u6301\u7e8c\u958b\u767c\u8edf\u9ad4\uff0c\u6bcf\u500b\u6708\u90fd\u5728\u6548\u80fd\u8207\u529f\u80fd\u4e0a\u6709\u6240\u6539\u9032\u3002<\/p>\n\n\n\n

\u5728\u6700\u65b0\u4e00\u8f2a\u7684\u63a8\u8ad6\u6e2c\u8a66\u88e1\uff0c\u5305\u62ec NVIDIA Hopper \u67b6\u69cb\u3001NVIDIA Jetson<\/a> \u5e73\u53f0\u8207 NVIDIA Triton \u63a8\u8ad6\u4f3a\u670d\u5668<\/a>\u5728\u5167\u7684 NVIDIA \u7522\u54c1\uff0c\u5728\u6548\u80fd\u65b9\u9762\u5747\u5927\u5e45\u63d0\u5347\u3002<\/p>\n\n\n\n

\u8207\u4e0a\u4e00\u8f2a\u7684\u6210\u7e3e\u76f8\u6bd4\uff0cNVIDIA H200 GPU \u5728\u8655\u7406\u751f\u6210\u5f0f AI \u63a8\u8ad6\u6642\u7684\u6548\u80fd\u63d0\u9ad8\u4e86 27%\uff0c\u7a81\u986f\u51fa\u5ba2\u6236\u5728\u6295\u8cc7 NVIDIA \u5e73\u53f0\u65b9\u9762\u6240\u5e36\u4f86\u7684\u9577\u671f\u9644\u52a0\u50f9\u503c\u3002<\/p>\n\n\n\n

\u96b8\u5c6c\u65bc NVIDIA AI \u5e73\u53f0\u7684 Triton \u63a8\u8ad6\u4f3a\u670d\u5668\u53ef\u642d\u914d NVIDIA AI Enterprise \u8edf\u9ad4\u4f7f\u7528\uff0c\u662f\u4e00\u6b3e\u529f\u80fd\u9f4a\u5168\u7684\u958b\u6e90\u63a8\u8ad6\u4f3a\u670d\u5668\uff0c\u53ef\u5354\u52a9\u7d44\u7e54\u5c07\u7279\u5b9a\u6846\u67b6\u7684\u63a8\u8ad6\u4f3a\u670d\u5668\u6574\u5408\u81f3\u529f\u80fd\u4e00\u500b\u7d71\u4e00\u7684\u7684\u5e73\u53f0\u4e0a\u3002\u6b64\u8209\u6709\u52a9\u65bc\u964d\u4f4e\u5728\u751f\u7522\u74b0\u5883\u88e1\u904b\u884c AI \u6a21\u578b\u7684\u7e3d\u6301\u6709\u6210\u672c\uff0c\u4e14\u5c07\u90e8\u7f72\u6a21\u578b\u7684\u6642\u9593\u5f9e\u5e7e\u500b\u6708\u7e2e\u77ed\u5230\u5e7e\u5206\u9418\u3002<\/p>\n\n\n\n

\u5728\u9019\u4e00\u8f2a MLPerf \u6e2c\u8a66\u4e2d\uff0cTriton \u63a8\u8ad6\u4f3a\u670d\u5668\u7684\u6548\u80fd\u9054\u5230\u5e7e\u4e4e\u7b49\u540c\u65bc NVIDIA \u88f8\u6a5f\u7684\u6e2c\u8a66\u7d50\u679c\uff0c\u986f\u793a\u4f01\u696d\u7121\u9700\u518d\u65bc\u4f7f\u7528\u529f\u80fd\u8c50\u5bcc\u7684\u751f\u7522\u7d1a AI \u63a8\u8ad6\u4f3a\u670d\u5668\uff0c\u8207\u9054\u5230\u9ad8\u5cf0\u541e\u5410\u91cf\u6548\u80fd\u4e4b\u9593\u505a\u51fa\u9078\u64c7\u3002<\/p>\n\n\n\n

<\/p>\n\n\n\n

\u8d70\u5411\u908a\u7de3<\/strong><\/strong><\/p>\n\n\n\n

\u7576\u751f\u6210\u5f0f AI \u6a21\u578b\u90e8\u7f72\u5728\u908a\u7de3\u6642\u53ef\u4ee5\u5c07\u5f71\u50cf\u548c\u8996\u8a0a\u7b49\u611f\u6e2c\u5668\u8cc7\u6599\uff0c\u8f49\u63db\u6210\u5177\u6709\u5f37\u5927\u60c5\u5883\u610f\u8b58\u7684\u3001\u5373\u6642\u548c\u53ef\u57f7\u884c\u7684\u898b\u89e3\u5167\u5bb9\u3002\u7528\u65bc\u908a\u7de3 AI \u548c\u6a5f\u5668\u4eba\u7684 NVIDIA Jetson \u5e73\u53f0\u5177\u6709\u7368\u7279\u7684\u80fd\u529b\uff0c\u53ef\u4ee5\u904b\u884c\u4efb\u4f55\u985e\u578b\u7684\u672c\u5730\u7aef\u6a21\u578b\uff0c\u5305\u62ec LLM\u3001\u8996\u89ba Transformer \u6a21\u578b\u53ca Stable Diffusion\u3002<\/p>\n\n\n\n

\u5728\u9019\u4e00\u8f2a MLPerf \u57fa\u6e96\u6e2c\u8a66\u4e2d\uff0cNVIDIA Jetson AGX Orin<\/a> \u7cfb\u7d71\u6a21\u7d44\u5728\u8655\u7406 GPT-J LLM \u5de5\u4f5c\u8ca0\u8f09\u65b9\u9762\uff0c\u8868\u73fe\u8f03\u4e0a\u4e00\u8f2a\u6e2c\u8a66\u7684\u50b3\u8f38\u91cf\u63d0\u9ad8 6.2 \u500d\uff0c\u5ef6\u9072\u8868\u73fe\u5247\u6539\u5584 2.4 \u500d\u3002\u9019\u500b\u64c1\u6709 60 \u5104\u500b\u53c3\u6578\u7684\u6a21\u578b\u80fd\u5920\u5354\u52a9\u6539\u9020\u908a\u7de3\u7684\u751f\u6210\u5f0f AI\u3002\u4f7f\u7528\u8005\u73fe\u5728\u53ef\u4ee5\u85c9\u7531\u9019\u500b\u901a\u7528\u6a21\u578b\uff0c\u5728\u908a\u7de3\u8207\u4eba\u985e\u8a9e\u8a00\u6d41\u66a2\u9032\u884c\u4e92\u52d5\uff0c\u7121\u9700\u91dd\u5c0d\u7279\u5b9a\u7528\u4f8b\u9032\u884c\u958b\u767c\u3002<\/p>\n\n\n\n

<\/p>\n\n\n\n

\u6548\u80fd\u5168\u9762\u9818\u5148<\/strong><\/strong><\/p>\n\n\n\n

\u9019\u4e00\u8f2a MLPerf Inference \u6e2c\u8a66\u7d50\u679c\u986f\u793a NVIDIA \u5e73\u53f0\u5f9e\u8cc7\u6599\u4e2d\u5fc3\u5230\u908a\u7de3\uff0c\u90fd\u5c55\u793a\u51fa\u5176\u591a\u5143\u7528\u9014\u7279\u8272\u53ca\u9818\u5148\u7684\u6548\u80fd\u8868\u73fe\uff0c\u5728\u6240\u6709\u57fa\u6e96\u5de5\u4f5c\u8ca0\u8f09\u4e0a\u9a45\u52d5\u5404\u7a2e\u6700\u5275\u65b0\u7684 AI \u61c9\u7528\u8207\u670d\u52d9\u3002\u6b61\u8fce\u95b1\u8b80\u6280\u8853\u90e8\u843d\u683c<\/a>\uff0c\u4e86\u89e3\u9019\u6b21\u6e2c\u8a66\u7d50\u679c\u7684\u7d30\u7bc0\u3002<\/p>\n\n\n\n

\u642d\u8f09 H200 GPU \u7684\u7cfb\u7d71\u5373\u65e5\u8d77\u7531\u7b2c\u4e00\u5bb6\u5ba3\u5e03\u6b63\u5f0f\u63a8\u51fa\u7684\u96f2\u7aef\u670d\u52d9\u4f9b\u61c9\u5546 CoreWeave\uff0c\u4ee5\u53ca\u83ef\u78a9\u3001\u6234\u723e\u3001\u6167\u8207\u79d1\u6280\u3001\u96f2\u9054\u79d1\u6280\u548c\u7f8e\u8d85\u5fae\u7b49\u4f3a\u670d\u5668\u88fd\u9020\u5546\u63d0\u4f9b\u3002<\/p>\n\n\n\n

<\/p>\n\n\n\n

<\/p>\n\n\n\n

\ud83d\udfe6\u770b\u6d77\u97fb\u6700\u65b0Power\u6587\u7ae0\u6436\u62ff\u5e02\u50f95790\u5143\u7684PRIME PX-750W<\/a><\/strong>
\ud83d\udfe6ioioTIMES \u7238\u7238\u633a\u4f60\u6253\u96fb\u52d5 \u6436\u62ffWi-Fi 7\u7b49\u597d\u79ae<\/a><\/strong>
\ud83d\udfe6\u73fe\u5728\u5c31\u52a0\u5165 ioioTIMES \u81c9\u66f8\u7c89\u7d72\u5718<\/a> \u66f4\u591a\u4e92\u52d5\u3001\u66f4\u591a\u597d\u5eb7\u650f\u62b5\u52a0!!<\/strong>
\ud83d\udfe6\u6211\u5011\u6709LINE TODAY<\/a>\u983b\u9053\u4e86\uff0c\u5feb\u4f86\u8ffd\u8e2a\u6211\u5011\u5427!!–\u6700\u65b0\u79d1\u6280\u65b0\u805e \u76e1\u5728\u4f60\u624b<\/strong><\/h4>\n","protected":false},"excerpt":{"rendered":"

\u4f7f\u7528NVIDIA Blackwell<\/p>\n","protected":false},"author":3,"featured_media":80126,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"rank_math_lock_modified_date":false,"footnotes":""},"categories":[13],"tags":[3988,580,10805,71,5344],"class_list":["post-80123","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-news","tag-blackwell","tag-focus","tag-mlperf-inference","tag-nvidia","tag-ai"],"_links":{"self":[{"href":"https:\/\/www.ioiotimes.com\/index.php?rest_route=\/wp\/v2\/posts\/80123"}],"collection":[{"href":"https:\/\/www.ioiotimes.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.ioiotimes.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.ioiotimes.com\/index.php?rest_route=\/wp\/v2\/users\/3"}],"replies":[{"embeddable":true,"href":"https:\/\/www.ioiotimes.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=80123"}],"version-history":[{"count":4,"href":"https:\/\/www.ioiotimes.com\/index.php?rest_route=\/wp\/v2\/posts\/80123\/revisions"}],"predecessor-version":[{"id":80130,"href":"https:\/\/www.ioiotimes.com\/index.php?rest_route=\/wp\/v2\/posts\/80123\/revisions\/80130"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.ioiotimes.com\/index.php?rest_route=\/wp\/v2\/media\/80126"}],"wp:attachment":[{"href":"https:\/\/www.ioiotimes.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=80123"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.ioiotimes.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=80123"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.ioiotimes.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=80123"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}