{"id":104033,"date":"2025-04-03T10:44:02","date_gmt":"2025-04-03T02:44:02","guid":{"rendered":"https:\/\/www.ioiotimes.com\/?p=104033"},"modified":"2025-04-03T10:58:57","modified_gmt":"2025-04-03T02:58:57","slug":"%e9%80%9f%e5%ba%a6%e5%87%ba%e7%9c%be%ef%bc%9anvidia-blackwell%e5%9c%a8%e6%9c%80%e6%96%b0mlperf%e6%8e%a8%e8%ab%96%e6%b8%ac%e8%a9%a6%e7%b5%90%e6%9e%9c%e4%b8%ad%e6%8b%94%e5%be%97%e9%a0%ad%e7%b1%8c","status":"publish","type":"post","link":"https:\/\/www.ioiotimes.com\/?p=104033","title":{"rendered":"\u901f\u5ea6\u51fa\u773e\uff1aNVIDIA Blackwell\u5728\u6700\u65b0MLPerf\u63a8\u8ad6\u6e2c\u8a66\u7d50\u679c\u4e2d\u62d4\u5f97\u982d\u7c4c"},"content":{"rendered":"\n

NVIDIA GB200 NVL72\u7cfb\u7d71\u5373\u4f7f\u5728\u6a21\u578b\u8b8a\u5f97\u66f4\u5927\u3001\u66f4\u8907\u96dc\u7684\u60c5\u6cc1\u4e0b\uff0c\u4f9d\u820a\u80fd\u5920\u4ee5\u51fa\u8272\u7684\u8f38\u9001\u91cf\u63d0\u5347AI\u5de5\u5ee0\u7684\u7372\u5229\u80fd\u529b<\/h3>\n\n\n\n

\u5728\u53cd\u6620\u90e8\u5206\u9ad8\u96e3\u5ea6\u63a8\u8ad6\u60c5\u5883\u7684\u6700\u65b0 MLPerf Inference V5.0 \u57fa\u6e96\u6e2c\u8a66\u4e2d\uff0cNVIDIA Blackwell<\/a> \u5e73\u53f0\u5275\u4e0b\u591a\u9805\u7d00\u9304\uff0c\u9019\u4e5f\u662f NVIDIA \u9996\u6b21\u4f7f\u7528\u5c08\u70ba\u4eba\u5de5\u667a\u6167\uff08AI\uff09\u63a8\u7406\u4f5c\u696d\u8a2d\u8a08\u7684\u6a5f\u67b6\u898f\u6a21\u89e3\u6c7a\u65b9\u6848 NVIDIA GB200 NVL72 \u7cfb\u7d71\u4f86\u63d0\u4ea4 MLPerf \u6e2c\u8a66\u5831\u544a\u3002<\/p>\n\n\n\n

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

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

\u70ba\u843d\u5be6\u6253\u9020\u5c16\u7aef AI \u7684\u627f\u8afe\uff0c\u9700\u8981\u7a31\u70ba AI \u5de5\u5ee0<\/a>\u7684\u65b0\u578b\u904b\u7b97\u57fa\u790e\u8a2d\u65bd\u3002\u8207\u50b3\u7d71\u7684\u8cc7\u6599\u4e2d\u5fc3\u4e0d\u4e00\u6a23\uff0cAI \u5de5\u5ee0\u4e0d\u53ea\u662f\u7528\u65bc\u5132\u5b58\u8207\u8655\u7406\u8cc7\u6599\uff0c\u800c\u662f\u900f\u904e\u5c07\u539f\u59cb\u8cc7\u6599\u8f49\u63db\u70ba\u5373\u6642\u6d1e\u5bdf\uff0c\u4f86\u5927\u898f\u6a21\u88fd\u9020\u667a\u6167\u3002AI \u5de5\u5ee0\u7684\u76ee\u6a19\u5f88\u7c21\u55ae\uff1a\u7528\u6700\u4f4e\u6210\u672c\uff0c\u76e1\u91cf\u70ba\u6700\u591a\u7684\u4f7f\u7528\u8005\u5feb\u901f\u63d0\u4f9b\u6e96\u78ba\u7684\u67e5\u8a62\u7b54\u6848\u3002<\/p>\n\n\n\n

\u8981\u505a\u5230\u9019\u500b\u76ee\u6a19\u7684\u8907\u96dc\u6027\u76f8\u7576\u9ad8\uff0c\u4e26\u4e14\u662f\u5728\u5e55\u5f8c\u9032\u884c\u3002\u96a8\u8457 AI \u6a21\u578b\u52a0\u5927\u5230\u6709\u8457\u6578\u5341\u5104\u751a\u81f3\u6578\u5146\u500b\u53c3\u6578\uff0c\u4ee5\u63d0\u4f9b\u66f4\u8070\u660e\u7684\u56de\u8986\u5167\u5bb9\uff0c\u7522\u751f\u6bcf\u500b\u8a5e\u5143\uff08token\uff09<\/a>\u6240\u9700\u7684\u904b\u7b97\u91cf\u4e5f\u96a8\u4e4b\u589e\u52a0\u3002\u9019\u9805\u9700\u6c42\u6703\u6e1b\u5c11 AI \u5de5\u5ee0\u53ef\u7522\u751f\u7684\u8a5e\u5143\u6578\u91cf\uff0c\u4e26\u4e14\u63d0\u9ad8\u6bcf\u500b\u8a5e\u5143\u7684\u6210\u672c\u3002\u5982\u679c\u60f3\u8981\u4fdd\u6301\u9ad8\u63a8\u8ad6\u8f38\u9001\u91cf\u548c\u4f4e\u55ae\u4f4d\u8a5e\u5143\u6210\u672c\uff0c\u5c31\u5fc5\u9808\u5728\u6676\u7247\u3001\u7db2\u8def\u7cfb\u7d71\u548c\u8edf\u9ad4\u7b49\u6280\u8853\u5806\u758a\u7684\u6bcf\u4e00\u5c64\u5feb\u901f\u9032\u884c\u5275\u65b0\u3002<\/p>\n\n\n\n

\u7d93\u540c\u696d\u5be9\u67e5\u7684\u63a8\u8ad6\u6548\u80fd\u696d\u754c\u57fa\u6e96 MLPerf Inference\uff0c\u5176\u6700\u65b0\u66f4\u65b0\u9805\u76ee\u5305\u62ec\u52a0\u5165 Llama 3.1 405B\uff0c\u9019\u662f\u898f\u6a21\u6700\u5927\u4e14\u6700\u96e3\u904b\u884c\u7684\u958b\u653e\u6b0a\u91cd\u6a21\u578b\u4e4b\u4e00\u3002\u8207\u539f\u672c\u7684 Llama 2 70B \u57fa\u6e96\u76f8\u6bd4\uff0c\u65b0\u7684 Llama 2 70B Interactive \u57fa\u6e96\u6709\u8457\u66f4\u56b4\u683c\u7684\u5ef6\u9072\u8981\u6c42\uff0c\u66f4\u80fd\u53cd\u6620\u751f\u7522\u90e8\u7f72\u74b0\u5883\u5728\u63d0\u4f9b\u6700\u4f73\u4f7f\u7528\u8005\u9ad4\u9a57\u65b9\u9762\u7684\u9650\u5236\u3002<\/p>\n\n\n\n

\u9664\u4e86 Blackwell \u5e73\u53f0\u4e4b\u5916\uff0c\u5728\u5168\u7aef\u6700\u4f73\u5316\u7684\u52a9\u529b\u4e0b\uff0cNVIDIA Hopper<\/a> \u5e73\u53f0\u4e5f\u5168\u9762\u5c55\u73fe\u51fa\u512a\u7570\u6548\u80fd\uff0c\u5728 Llama 2 70B \u4e0a\u7684\u6548\u80fd\u8f03\u53bb\u5e74\u986f\u8457\u63d0\u5347\u3002<\/p>\n\n\n\n

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

NVIDIA Blackwell\u5beb\u4e0b\u65b0\u7d00\u9304<\/strong><\/h3>\n\n\n\n

\u9023\u63a5 72 \u9846 NVIDIA Blackwell GPU \u4f5c\u70ba\u55ae\u4e00\u5927\u578b GPU \u7684 GB200 NVL72 \u7cfb\u7d71\uff0c\u5728 Llama 3.1 405B \u57fa\u6e96\u6e2c\u8a66\u4e2d\u7684\u8f38\u9001\u91cf\uff0c\u6bd4\u672c\u8f2a\u4f7f\u7528 NVIDIA H200 NVL8 \u63d0\u4ea4\u7684\u7d50\u679c\u9ad8\u51fa 30 \u500d\u3002\u9019\u9805\u58ef\u8209\u662f\u900f\u904e\u6bcf\u9846 GPU \u63d0\u9ad8\u4e09\u500d\u7684\u6548\u80fd\uff0c\u518d\u52a0\u4e0a\u9ad8\u51fa\u4e5d\u500d\u7684 NVIDIA NVLink<\/a> \u4e92\u9023\u7db2\u57df\u52a9\u529b\u4e0b\u4f86\u9054\u6210\u3002<\/p>\n\n\n\n

\"\"
\u25b2NVIDIA\u9996\u6b21\u4f7f\u7528\u5c08\u70baAI\u63a8\u7406\u4f5c\u696d\u8a2d\u8a08\u7684\u6a5f\u67b6\u898f\u6a21\u89e3\u6c7a\u65b9\u6848NVIDIA GB200 NVL72\u7cfb\u7d71\u4f86\u63d0\u4ea4MLPerf\u6e2c\u8a66\u5831\u544a<\/figcaption><\/figure>\n\n\n\n

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

\u96d6\u7136\u8a31\u591a\u516c\u53f8\u90fd\u5728\u5176\u786c\u9ad4\u7522\u54c1\u4e0a\u9032\u884c MLPerf \u57fa\u6e96\u6e2c\u8a66\u4ee5\u8a55\u4f30\u6548\u80fd\uff0c\u4f46\u53ea\u6709 NVIDIA \u53ca\u5176\u5408\u4f5c\u5925\u4f34\u63d0\u4ea4\u4e26\u516c\u958b Llama 3.1 405B \u57fa\u6e96\u6e2c\u8a66\u7684\u7d50\u679c\u3002<\/p>\n\n\n\n

\u751f\u7522\u63a8\u8ad6\u90e8\u7f72\u74b0\u5883\u901a\u5e38\u5728\u5169\u5927\u95dc\u9375\u6307\u6a19\u4e0a\u6709\u5ef6\u9072\u9650\u5236\u3002\u9996\u5148\u662f\u7522\u751f\u9996\u500b\u8a5e\u5143\u7684\u6642\u9593\uff08TTFT\uff09\uff0c\u4e5f\u5c31\u662f\u4f7f\u7528\u8005\u9700\u8981\u82b1\u591a\u5c11\u6642\u9593\u624d\u80fd\u958b\u59cb\u770b\u5230\u5927\u578b\u8a9e\u8a00\u6a21\u578b<\/a>\u5c0d\u67e5\u8a62\u7684\u56de\u61c9\u3002\u518d\u8005\u662f\u7522\u751f\u6bcf\u500b\u8f38\u51fa\u8a5e\u5143\u7684\u6642\u9593\uff08TPOT\uff09\uff0c\u4e5f\u5c31\u662f\u5c07\u8a5e\u5143\u50b3\u9001\u7d66\u4f7f\u7528\u8005\u7684\u901f\u5ea6\u3002<\/p>\n\n\n\n

\u5168\u65b0 Llama 2 70B Interactive \u57fa\u6e96\u6e2c\u8a66\u7684 TPOT \u7e2e\u77ed\u4e86\u4e94\u500d\uff0cTTFT \u5247\u964d\u4f4e\u4e86 4.4 \u500d\uff0c\u53ef\u4ee5\u5275\u9020\u51fa\u53cd\u61c9\u66f4\u8fc5\u901f\u7684\u4f7f\u7528\u8005\u9ad4\u9a57\u3002NVIDIA \u5728\u672c\u8f2a\u6e2c\u8a66\u6240\u63d0\u4ea4\u7684\u6e2c\u8a66\u5831\u544a\u4e2d\uff0c\u4f7f\u7528\u642d\u8f09 8 \u9846 Blackwell GPU \u7684 NVIDIA DGX B200 \u7cfb\u7d71\uff0c\u5176\u6548\u80fd\u8f03\u642d\u8f09 8 \u9846 NVIDIA H200 GPU \u7684\u7cfb\u7d71\u9ad8\u51fa\u4e09\u500d\uff0c\u70ba\u9019\u9805\u66f4\u5177\u6311\u6230\u6027\u7684 Llama 2 70B \u57fa\u6e96\u6a39\u7acb\u66f4\u9ad8\u6a19\u6e96\u3002<\/p>\n\n\n\n

\u7d50\u5408 Blackwell \u67b6\u69cb\u53ca\u5176\u7d93\u904e\u6700\u4f73\u5316\u7684\u8edf\u9ad4\u5806\u758a\uff0c\u53ef\u4ee5\u8b93\u63a8\u8ad6\u6548\u80fd\u66f4\u4e0a\u4e00\u5c64\u6a13\uff0c\u70ba AI \u5de5\u5ee0\u6253\u597d\u57fa\u790e\uff0c\u4ee5\u63d0\u4f9b\u66f4\u9ad8\u7684\u667a\u6167\u3001\u66f4\u9ad8\u7684\u8f38\u9001\u91cf\u8207\u66f4\u5feb\u7684\u8a5e\u5143\u751f\u6210\u7387\u3002<\/p>\n\n\n\n

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

\u63a1\u7528NVIDIA Hopper\u7684AI\u5de5\u5ee0\u50f9\u503c\u6301\u7e8c\u63d0\u5347<\/strong><\/h3>\n\n\n\n

NVIDIA Hopper \u67b6\u69cb\u65bc 2022 \u5e74\u63a8\u51fa\uff0c\u9a45\u52d5\u4e86\u73fe\u4eca\u8a31\u591a AI \u63a8\u8ad6\u5de5\u5ee0\uff0c\u4e5f\u6301\u7e8c\u652f\u63f4\u8a13\u7df4\u5404\u7a2e\u6a21\u578b\u3002\u900f\u904e\u6301\u7e8c\u7684\u8edf\u9ad4\u6700\u4f73\u5316\uff0cNVIDIA \u63d0\u5347\u4e86\u63a1\u7528 Hopper \u67b6\u69cb AI \u5de5\u5ee0\u7684\u8f38\u9001\u91cf\uff0c\u9032\u800c\u5275\u9020\u51fa\u66f4\u9f90\u5927\u7684\u50f9\u503c\u3002<\/p>\n\n\n\n

\u5728\u4e00\u5e74\u524d\u65bc MLPerf Inference v4.0 \u9996\u6b21\u63a8\u51fa\u7684 Llama 2 70B \u57fa\u6e96\u6e2c\u8a66\u4e2d\uff0cH100 GPU \u7684\u8f38\u9001\u91cf\u589e\u52a0\u4e86 1.5 \u500d\u3002H200 GPU \u4ee5\u76f8\u540c\u7684 Hopper GPU \u67b6\u69cb\u70ba\u57fa\u790e\uff0c\u642d\u914d\u66f4\u5927\u3001\u66f4\u5feb\u7684 GPU \u8a18\u61b6\u9ad4\uff0c\u53c8\u5c07\u8f38\u9001\u91cf\u52a0\u5927\u5230 1.6 \u500d\u3002<\/p>\n\n\n\n

Hopper \u67b6\u69cb\u4e5f\u57f7\u884c\u4e86\u6240\u6709\u7684\u57fa\u6e96\u6e2c\u8a66\uff0c\u5305\u62ec\u65b0\u52a0\u5165\u7684 Llama 3.1 405B\u3001Llama 2 70B Interactive \u53ca\u5716\u5f62\u795e\u7d93\u7db2\u8def<\/a>\u6e2c\u8a66\u3002\u9019\u7a2e\u591a\u6a23\u6027\u4ee3\u8868 Hopper \u53ef\u4ee5\u904b\u884c\u591a\u7a2e\u5de5\u4f5c\u8ca0\u8f09\uff0c\u4e26\u5728\u6a21\u578b\u548c\u4f7f\u7528\u60c5\u5883\u8b8a\u5f97\u66f4\u5177\u6311\u6230\u6027\u6642\u8ddf\u4e0a\u8173\u6b65\u3002<\/p>\n\n\n\n

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

\u4e0d\u53ef\u6216\u7f3a\u7684\u751f\u614b\u7cfb<\/strong><\/h3>\n\n\n\n

\u5728\u672c\u8f2a MLPerf \u6e2c\u8a66\u4e2d\uff0c\u5171\u6709 15 \u5bb6\u5408\u4f5c\u5925\u4f34\u63d0\u51fa\u4f7f\u7528 NVIDIA \u5e73\u53f0\u9032\u884c\u6e2c\u8a66\u7684\u512a\u79c0\u6210\u7e3e\uff0c\u5305\u62ec\u83ef\u78a9\u3001\u601d\u79d1\u3001CoreWeave\u3001\u6234\u723e\u79d1\u6280\u96c6\u5718\u3001\u5bcc\u58eb\u901a\u3001\u6280\u92fc\u79d1\u6280\uff08Giga Computing\uff09\u3001Google Cloud\u3001\u6167\u8207\u79d1\u6280\u3001Lambda\u3001\u806f\u60f3\u96c6\u5718\u3001Oracle Cloud Infrastructure\u3001\u96f2\u9054\u79d1\u6280\u3001\u7f8e\u8d85\u5fae\u3001Sustainable Metal Cloud \u548c VMware\u3002<\/p>\n\n\n\n

\u5f9e\u53c3\u8cfd\u4f5c\u54c1\u7684\u5ee3\u5ea6\u53ef\u4ee5\u770b\u51fa NVIDIA \u5e73\u53f0\u7684\u5f71\u97ff\u529b\uff0c\u5168\u7403\u6240\u6709\u96f2\u7aef\u670d\u52d9\u4f9b\u61c9\u5546\u53ca\u4f3a\u670d\u5668\u88fd\u9020\u5546\u7686\u53ef\u4f7f\u7528 NVIDIA \u5e73\u53f0\u3002<\/p>\n\n\n\n

MLCommons \u81f4\u529b\u65bc\u6301\u7e8c\u767c\u5c55 MLPerf Inference \u57fa\u6e96\u5957\u4ef6\uff0c\u4ee5\u8ddf\u4e0a AI \u9818\u57df\u7684\u6700\u65b0\u767c\u5c55\uff0c\u4e26\u70ba\u751f\u614b\u7cfb\u63d0\u4f9b\u56b4\u8b39\u4e14\u7d93\u540c\u696d\u5be9\u67e5\u7684\u6548\u80fd\u6578\u64da\uff0c\u9019\u5c0d\u5354\u52a9 IT \u6c7a\u7b56\u8005\u9078\u64c7\u6700\u4f73\u7684 AI \u57fa\u790e\u67b6\u69cb\u800c\u8a00\u81f3\u95dc\u91cd\u8981\u3002<\/p>\n\n\n\n

\u9032\u4e00\u6b65\u77ad\u89e3<\/em> <\/em>MLPerf<\/em><\/a>\u3002<\/em><\/p>\n\n\n\n

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

\u25bcFirst Look: NVIDIA DGX SuperPOD with DGX GB200 Systems<\/strong><\/p>\n\n\n\n

\n