{"id":89203,"date":"2024-11-14T13:16:21","date_gmt":"2024-11-14T05:16:21","guid":{"rendered":"https:\/\/www.ioiotimes.com\/?p=89203"},"modified":"2024-11-14T13:16:22","modified_gmt":"2024-11-14T05:16:22","slug":"%e5%b3%b0%e5%80%bc%e8%a8%93%e7%b7%b4%ef%bc%9ablackwell%e6%8f%90%e4%be%9b%e6%9b%b4%e9%ab%98%e6%b0%b4%e6%ba%96%e7%9a%84mlperf%e8%a8%93%e7%b7%b4%e6%95%88%e8%83%bd","status":"publish","type":"post","link":"https:\/\/www.ioiotimes.com\/?p=89203","title":{"rendered":"\u5cf0\u503c\u8a13\u7df4\uff1aBlackwell\u63d0\u4f9b\u66f4\u9ad8\u6c34\u6e96\u7684MLPerf\u8a13\u7df4\u6548\u80fd"},"content":{"rendered":"\n

\u5728\u5b8c\u6210\u6bcf\u9805MLPerf\u6e2c\u8a66\u5f8c\uff0cBlackwell\u63d0\u5347\u4e86AI\u9a45\u52d5\u7684\u61c9\u7528\u8a13\u7df4\u6548\u80fd\u8868\u73fe<\/h3>\n\n\n\n
\"\"<\/figure>\n\n\n\n

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

\u4f7f\u7528\u6587\u5b57\u3001\u96fb\u8166\u7a0b\u5f0f\u78bc\u3001\u86cb\u767d\u8cea\u93c8\u3001\u6458\u8981\u3001\u5f71\u7247\u751a\u81f3 3D \u5716\u5f62\u7684\u751f\u6210\u5f0f\u4eba\u5de5\u667a\u6167\uff08AI\uff09\u61c9\u7528\u9700\u8981\u8cc7\u6599\u4e2d\u5fc3\u898f\u6a21\u7684\u52a0\u901f\u904b\u7b97\uff0c\u624d\u80fd\u6709\u6548\u7387\u5730\u8a13\u7df4\u9a45\u52d5\u9019\u4e9b\u61c9\u7528\u7684\u5927\u578b\u8a9e\u8a00\u6a21\u578b\uff08LLM\uff09\u3002<\/p>\n\n\n\n

\u5728\u00a0MLPerf Training 4.1\u00a0\u57fa\u6e96\u6e2c\u8a66\u4e2d\uff0cNVIDIA Blackwell<\/a>\u00a0\u5e73\u53f0\u5728\u6240\u6709\u6e2c\u8a66\u4e2d\u7684\u5de5\u4f5c\u8ca0\u8f09\u65b9\u9762\u90fd\u53d6\u5f97\u4e86\u4ee4\u4eba\u5370\u8c61\u6df1\u523b\u7684\u7d50\u679c\uff0c\u4e26\u4e14\u5728\u5927\u578b\u8a9e\u8a00\u6a21\u578b\u57fa\u6e96\u6e2c\u8a66\u4e2d\uff0c\u6bcf\u500b GPU\u00a0\u7684\u6548\u80fd\u6700\u9ad8\u63d0\u5347\u4e86 2.2\u00a0\u500d\uff0c\u5305\u62ec\u00a0Llama 2 70B\u00a0\u548c\u00a0GPT-3 175B\u00a0\u9810\u8a13\u7df4\u3002<\/p>\n\n\n\n

\u6b64\u5916\uff0cNVIDIA \u5728 NVIDIA Hopper \u5e73\u53f0\u4e0a\u63d0\u4ea4\u7684\u5167\u5bb9\u7e7c\u7e8c\u4fdd\u6301\u6240\u6709\u57fa\u6e96\u6e2c\u8a66\u7684\u5927\u898f\u6a21\u8a18\u9304\uff0c\u5176\u4e2d\u5305\u62ec\u5728 GPT-3 175B \u57fa\u6e96\u6e2c\u8a66\u4e2d\u63d0\u4ea4\u7684 11,616 \u500b Hopper GPU\u3002<\/p>\n\n\n\n

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

Blackwell\u7684\u8de8\u8d8a\u5f0f\u9032\u5c55<\/strong><\/h3>\n\n\n\n

\u9996\u6b21\u5411\u00a0MLCommons\u00a0\u806f\u76df\u63d0\u4ea4\u7684 Blackwell \u8a13\u7df4\u7d50\u679c\uff0c\u51f8\u986f\u4e86\u8a72\u67b6\u69cb\u5982\u4f55\u63a8\u5347\u751f\u6210\u5f0f\u00a0AI\u00a0\u8a13\u7df4\u6548\u80fd\u3002MLCommons\u00a0\u806f\u76df\u70ba\u7522\u696d\u53c3\u8207\u8005\u5275\u5efa\u6a19\u6e96\u5316\u3001\u516c\u6b63\u4e14\u56b4\u683c\u7684\u540c\u696d\u5be9\u67e5\u6e2c\u8a66\u3002<\/p>\n\n\n\n

\u4f8b\u5982\uff0cBlackwell \u67b6\u69cb\u5305\u62ec\u53ef\u66f4\u9ad8\u6548\u5229\u7528 Tensor Core<\/a> \u7684\u65b0\u6838\u5fc3\u3002\u6838\u5fc3\u662f\u6700\u4f73\u5316\u7684\u3001\u5c08\u9580\u5efa\u69cb\u7684\u6578\u5b78\u904b\u7b97\uff0c\u4f8b\u5982\u77e9\u9663\u4e58\u6cd5\uff0c\u662f\u8a31\u591a\u6df1\u5ea6\u5b78\u7fd2\u6f14\u7b97\u6cd5\u7684\u6838\u5fc3\u3002<\/p>\n\n\n\n

Blackwell \u64c1\u6709\u66f4\u9ad8\u7684\u6bcf GPU \u904b\u7b97\u8f38\u9001\u91cf\uff0c\u4ee5\u53ca\u66f4\u5927\u3001\u66f4\u5feb\u7684\u9ad8\u983b\u5bec\u8a18\u61b6\u9ad4\uff0c\u4f7f\u5176\u80fd\u5920\u5728\u66f4\u5c11\u7684 GPU \u4e0a\u904b\u884c GPT-3 175B \u57fa\u6e96\u6e2c\u8a66\uff0c\u540c\u6642\u5be6\u73fe\u51fa\u8272\u7684\u6bcf GPU \u6548\u80fd\u3002<\/p>\n\n\n\n

\u5229\u7528\u66f4\u5927\u4e14\u983b\u5bec\u66f4\u9ad8\u7684 HBM3e \u8a18\u61b6\u9ad4\uff0c\u50c5\u9700 64 \u500b Blackwell GPU \u5c31\u80fd\u57f7\u884c GPT-3 \u5927\u578b\u8a9e\u8a00\u6a21\u578b\u7684\u57fa\u6e96\u6e2c\u8a66\uff0c\u4e14\u4e0d\u5f71\u97ff\u6bcf GPU \u7684\u6027\u80fd\u3002\u4f7f\u7528 Hopper \u57f7\u884c\u76f8\u540c\u7684\u57fa\u6e96\u6e2c\u8a66\uff0c\u5247\u9700\u8981 256 \u500b GPU\u3002<\/p>\n\n\n\n

Blackwell \u7684\u8a13\u7df4\u7d50\u679c\u5ef6\u7e8c\u5176\u5148\u524d\u5728 MLPerf Inference 4.1 \u7684\u63d0\u4ea4\u6210\u679c\uff0c\u5728\u8a72\u6e2c\u8a66\u4e2d\uff0cBlackwell \u7684\u5927\u578b\u8a9e\u8a00\u6a21\u578b\u63a8\u8ad6\u6548\u80fd\u6bd4 Hopper \u4e16\u4ee3\u9ad8\u51fa\u6700\u591a 4 \u500d\u3002\u5229\u7528 Blackwell \u67b6\u69cb\u7684 FP4 \u7cbe\u5ea6\u4ee5\u53ca NVIDIA QUASAR \u91cf\u5316\u7cfb\u7d71\u6240\u63d0\u4ea4\u7684\u8cc7\u6599\uff0c\u5728\u6eff\u8db3\u57fa\u6e96\u6e96\u78ba\u6027\u8981\u6c42\u7684\u540c\u6642\uff0c\u5c55\u793a\u4e86\u5f37\u5927\u7684\u6548\u80fd\u3002<\/p>\n\n\n\n

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

\u4e0d\u65b7\u6700\u4f73\u5316<\/strong><\/strong><\/h3>\n\n\n\n

NVIDIA \u5e73\u53f0\u7d93\u6b77\u6301\u7e8c\u7684\u8edf\u9ad4\u958b\u767c\uff0c\u63d0\u5347\u4e86\u591a\u7a2e\u6846\u67b6\u3001\u6a21\u578b\u548c\u61c9\u7528\u7684\u8a13\u7df4\u8207\u63a8\u7406\u6548\u80fd\u53ca\u529f\u80fd\u3002<\/p>\n\n\n\n

\u5728\u9019\u4e00\u8f2a\u00a0MLPerf\u00a0\u8a13\u7df4\u63d0\u4ea4\u4e2d\uff0cHopper\u00a0\u5728\u00a0GPT-3 175B\u00a0\u7684\u6bcf\u00a0GPU\u00a0\u8a13\u7df4\u6548\u80fd\u4e0a\uff0c\u63d0\u4f9b\u81ea\u57fa\u6e96\u63a8\u51fa\u4ee5\u4f86\u00a01.3\u00a0\u500d\u7684\u63d0\u5347\u3002NVIDIA\u00a0\u4e5f\u63d0\u4ea4\u4e86\u00a0GPT-3 175B\u00a0\u57fa\u6e96\u6e2c\u8a66\u7684\u5927\u898f\u6a21\u7d50\u679c\uff0c\u4f7f\u7528\u00a011,616\u00a0\u500b\u00a0Hopper GPU \u8207\u00a0NVIDIA NVLink\u00a0\u548c\u00a0NVSwitch<\/a>\u00a0\u9ad8\u983b\u5bec\u00a0GPU\u00a0\u5230\u00a0GPU\u00a0\u901a\u8a0a\uff0c\u4ee5\u53ca\u00a0NVIDIA Quantum-2 InfiniBand<\/a>\u00a0\u7db2\u8def\u9023\u63a5\u3002<\/p>\n\n\n\n

\u81ea\u53bb\u5e74\u4ee5\u4f86\uff0cNVIDIA Hopper GPU \u5728 GPT-3 175B \u57fa\u6e96\u6e2c\u8a66\u7684\u898f\u6a21\u548c\u6548\u80fd\u4e0a\u63d0\u5347\u4e86\u4e09\u500d\u4ee5\u4e0a\u3002\u6b64\u5916\uff0c\u5728 Llama 2 70B LoRA \u5fae\u8abf\u57fa\u6e96\u6e2c\u8a66\u4e2d\uff0cNVIDIA \u4f7f\u7528\u76f8\u540c\u6578\u91cf\u7684 Hopper GPU \u5c07\u6548\u80fd\u63d0\u5347\u4e86 26%\uff0c\u53cd\u6620\u51fa\u8edf\u9ad4\u7684\u6301\u7e8c\u5f37\u5316\u3002<\/p>\n\n\n\n

NVIDIA\u00a0\u6301\u7e8c\u81f4\u529b\u65bc\u6700\u4f73\u5316\u5176\u52a0\u901f\u904b\u7b97\u5e73\u53f0\uff0c\u5f9e\u800c\u4e0d\u65b7\u6539\u9032 MLPerf\u00a0\u6e2c\u8a66\u7d50\u679c\uff0c\u63d0\u9ad8\u5bb9\u5668\u5316\u8edf\u9ad4\u7684\u6548\u80fd\uff0c\u70ba\u73fe\u6709\u5e73\u53f0\u4e0a\u7684\u5408\u4f5c\u5925\u4f34\u548c\u5ba2\u6236\u5e36\u4f86\u66f4\u5f37\u5927\u7684\u904b\u7b97\u80fd\u529b\uff0c\u4e26\u63d0\u5347\u5176\u5e73\u53f0\u6295\u8cc7\u7684\u5831\u916c\u3002<\/p>\n\n\n\n

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

\u651c\u624b\u5408\u4f5c<\/strong><\/h3>\n\n\n\n

NVIDIA \u7684\u5408\u4f5c\u5925\u4f34\uff0c\u5305\u62ec\u83ef\u78a9\u3001Azure\u3001\u601d\u79d1\u3001\u6234\u723e\u79d1\u6280\u96c6\u5718\u3001\u5bcc\u58eb\u901a\u3001\u6280\u92fc\u79d1\u6280\u3001Lambda Labs\u3001\u806f\u60f3\u96c6\u5718\u3001Oracle Cloud\u3001\u96f2\u9054\u79d1\u6280\u548c\u7f8e\u8d85\u5fae\u7b49\u7cfb\u7d71\u88fd\u9020\u5546\u548c\u96f2\u7aef\u670d\u52d9\u4f9b\u61c9\u5546\uff0c\u4e5f\u5728\u672c\u8f2a MLPerf \u6e2c\u8a66\u4e2d\u63d0\u4ea4\u4e86\u51fa\u8272\u7684\u6210\u679c\u3002<\/p>\n\n\n\n

\u4f5c\u70ba\u00a0MLCommons\u00a0\u7684\u5275\u59cb\u6210\u54e1\uff0cNVIDIA\u00a0\u8a8d\u70ba\u696d\u754c\u6a19\u6e96\u57fa\u6e96\u6e2c\u8a66\u548c\u57fa\u6e96\u6e2c\u8a66\u6700\u4f73\u5be6\u8e10\u5728\u00a0AI\u00a0\u904b\u7b97\u4e2d\u7684\u89d2\u8272\u81f3\u95dc\u91cd\u8981\u3002\u900f\u904e\u5c0d AI \u548c\u9ad8\u6548\u80fd\u904b\u7b97\u5e73\u53f0\u9032\u884c\u540c\u696d\u5be9\u67e5\u3001\u7c21\u5316\u7684\u6bd4\u8f03\uff0c\u4f01\u696d\u80fd\u5920\u8ddf\u4e0a\u6700\u65b0\u7684\u00a0AI\u00a0\u904b\u7b97\u5275\u65b0\uff0c\u4e26\u7372\u53d6\u6709\u52a9\u65bc\u5f15\u5c0e\u91cd\u8981\u5e73\u53f0\u6295\u8cc7\u6c7a\u7b56\u7684\u95dc\u9375\u6578\u64da\u3002<\/p>\n\n\n\n

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

\u3000\u3000\u3000<\/p>\n\n\n\n

\ud83d\udfe6\u770bioioTIEMS \u5b78\u96d911\u63a1\u8cfc\u6280\u5de7 \u6436\u62ff\u8d85\u7d1a\u597d\u79ae!<\/a><\/strong>
\ud83d\udfe6\u73fe\u5728\u5c31\u52a0\u5165 ioioTIMES \u81c9\u66f8\u7c89\u7d72\u5718 \u66f4\u591a\u4e92\u52d5\u3001\u66f4\u591a\u597d\u5eb7\u650f\u62b5\u52a0!!
\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":"

\u5728\u5b8c\u6210\u6bcf\u9805MLPerf\u6e2c\u8a66\u5f8c\uff0cBla<\/p>\n","protected":false},"author":3,"featured_media":89216,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"rank_math_lock_modified_date":false,"footnotes":""},"categories":[13],"tags":[3988,580,6360,71],"class_list":["post-89203","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-news","tag-blackwell","tag-focus","tag-mlperf","tag-nvidia"],"_links":{"self":[{"href":"https:\/\/www.ioiotimes.com\/index.php?rest_route=\/wp\/v2\/posts\/89203"}],"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=89203"}],"version-history":[{"count":2,"href":"https:\/\/www.ioiotimes.com\/index.php?rest_route=\/wp\/v2\/posts\/89203\/revisions"}],"predecessor-version":[{"id":89217,"href":"https:\/\/www.ioiotimes.com\/index.php?rest_route=\/wp\/v2\/posts\/89203\/revisions\/89217"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.ioiotimes.com\/index.php?rest_route=\/wp\/v2\/media\/89216"}],"wp:attachment":[{"href":"https:\/\/www.ioiotimes.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=89203"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.ioiotimes.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=89203"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.ioiotimes.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=89203"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}