{"id":89985,"date":"2024-11-21T10:55:46","date_gmt":"2024-11-21T02:55:46","guid":{"rendered":"https:\/\/www.ioiotimes.com\/?p=89985"},"modified":"2024-11-21T10:55:47","modified_gmt":"2024-11-21T02:55:47","slug":"%e8%bf%bd%e6%b1%82%e9%80%9f%e5%ba%a6%ef%bc%9anvidia%e5%8a%a0%e9%80%9f%e5%85%a8%e7%90%83%e5%a4%9a%e6%95%b8%e8%b6%85%e7%b4%9a%e9%9b%bb%e8%85%a6%ef%bc%8c%e4%bb%a5%e6%8e%a8%e5%8b%95%e7%a7%91%e5%ad%b8","status":"publish","type":"post","link":"https:\/\/www.ioiotimes.com\/?p=89985","title":{"rendered":"\u8ffd\u6c42\u901f\u5ea6\uff1aNVIDIA\u52a0\u901f\u5168\u7403\u591a\u6578\u8d85\u7d1a\u96fb\u8166\uff0c\u4ee5\u63a8\u52d5\u79d1\u5b78\u8207\u79d1\u6280\u9032\u6b65"},"content":{"rendered":"\n

\u6700\u65b0\u7684TOP500\u540d\u55ae\u986f\u793a\u6709384\u5957\u7cfb\u7d71\u63a1\u7528NVIDIA\u6280\u8853\uff0cGreen500\u540d\u55ae\u524d10\u540d\u4e2d\u76848\u540d\u4ea6\u63a1\u7528NVIDIA\u6280\u8853\uff1b\u6b64\u5916\uff0cNVIDIA\u65bcHPCwire\u8b80\u8005\u548c\u7de8\u8f2f\u7968\u9078\u734e\u56ca\u62ec11\u5ea7\u734e\u9805<\/h3>\n\n\n\n
\"\"<\/figure>\n\n\n\n

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

\u81ea\u00a02006\u00a0\u5e74\u767c\u5e03\u00a0CUDA\u00a0\u4ee5\u4f86\uff0cNVIDIA\u00a0\u4e00\u76f4\u5728\u63a8\u52d5\u4eba\u5de5\u667a\u6167\uff08AI\uff09\u548c\u52a0\u901f\u904b\u7b97\u9818\u57df\u7684\u9032\u6b65\uff0c\u800c\u6700\u65b0\u7684\u5168\u7403\u6700\u5f37\u5927\u8d85\u7d1a\u96fb\u8166\u00a0TOP500\u00a0\u540d\u55ae<\/a>\u7a81\u986f\u4e86 NVIDIA \u5728\u6b64\u9818\u57df\u6240\u53d6\u5f97\u7684\u5dd4\u5cf0\u6210\u5c31\u3002<\/p>\n\n\n\n

\u4eca\u5e74\uff0cTOP500 \u540d\u55ae\u4e0a\u7684 384 \u500b\u7cfb\u7d71\u5747\u63a1\u7528 NVIDIA \u6280\u8853\u3002\u5728\u65b0\u4e0a\u699c\u7684 53 \u500b\u7cfb\u7d71\u4e2d\uff0c87%\uff0846 \u500b\u7cfb\u7d71\uff09\u6709\u52a0\u901f\u3002\u5728\u9019\u4e9b\u52a0\u901f\u7684\u7cfb\u7d71\u4e2d\uff0c85% \u4f7f\u7528 NVIDIA Hopper GPU<\/a>\uff0c\u63a8\u52d5\u4e86\u6c23\u5019\u9810\u6e2c<\/a>\u3001\u85e5\u7269\u63a2\u7d22<\/a>\u548c\u91cf\u5b50\u6a21\u64ec<\/a>\u7b49\u9818\u57df\u7684\u9032\u6b65\u3002<\/p>\n\n\n\n

\u52a0\u901f\u904b\u7b97\u4e0d\u50c5\u50c5\u662f\u6bcf\u79d2\u6d6e\u9ede\u904b\u7b97\u6b21\u6578\uff08FLOPS\uff09\u3002\u5b83\u9700\u8981\u5168\u7aef\u3001\u7279\u5b9a\u65bc\u61c9\u7528\u7684\u6700\u4f73\u5316\u3002\u5728\u672c\u9031\u7684\u8d85\u7d1a\u904b\u7b97\u5927\u6703\uff08SC24\uff09\u4e0a\uff0cNVIDIA\u00a0\u5ba3\u5e03\u63a8\u51fa cuPyNumeric<\/a>\uff0c\u9019\u662f\u4e00\u500b\u00a0NVIDIA CUDA-X\u00a0\u51fd\u5f0f\u5eab\uff0c\u4f7f\u8d85\u904e\u00a0500\u00a0\u842c\u958b\u767c\u4eba\u54e1\u80fd\u5920\u7121\u7e2b\u64f4\u5c55\u5230\u5f37\u5927\u7684\u904b\u7b97\u53e2\u96c6\uff0c\u800c\u7121\u9700\u4fee\u6539\u5176\u00a0Python\u00a0\u7a0b\u5f0f\u78bc\u3002<\/p>\n\n\n\n

NVIDIA \u4e5f\u767c\u8868\u4e86 NVIDIA CUDA-Q<\/a> \u958b\u767c\u5e73\u53f0\u7684\u91cd\u5927\u66f4\u65b0\uff0c\u8a72\u5e73\u53f0\u4f7f\u91cf\u5b50\u7814\u7a76\u4eba\u54e1\u80fd\u5920\u4ee5\u4ee5\u524d\u8a8d\u70ba\u7121\u6cd5\u5be6\u73fe\u7684\u904b\u7b97\u898f\u6a21\u4f86\u6a21\u64ec\u91cf\u5b50\u8a2d\u5099\u3002<\/p>\n\n\n\n

\u6b64\u5916\uff0cNVIDIA\u00a0\u4ea6\u5728 HPCwire\u00a0\u8b80\u8005\u548c\u7de8\u8f2f\u7968\u9078\u734e\uff08Readers\u2019 and Editors\u2019 Choice Awards\uff09\u591a\u500b\u985e\u5225\u4e2d\u8d0f\u5f97\u8d85\u904e\u5341\u9805\u734e\u9805\uff0c\u9019\u662f\u00a0NVIDIA\u00a0\u9023\u7e8c\u00a020\u00a0\u5e74\u7372\u5f97\u80af\u5b9a\u3002<\/p>\n\n\n\n

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

<\/a>\u6df7\u5408\u7cbe\u5ea6\u548c<\/strong>AI<\/strong>\u7684\u79d1\u5b78\u767c\u73fe\u65b0\u6642\u4ee3<\/strong><\/strong><\/h3>\n\n\n\n

\u6df7\u5408\u7cbe\u5ea6\u6d6e\u9ede\u904b\u7b97\u548cAI\u5df2\u6210\u70ba\u7814\u7a76\u4eba\u54e1\u61c9\u5c0d\u73fe\u4ee3\u79d1\u5b78\u8907\u96dc\u6027\u7684\u9996\u9078\u5de5\u5177\u3002\u8207\u50b3\u7d71\u65b9\u6cd5\u76f8\u6bd4\uff0c\u5b83\u5011\u5177\u6709\u66f4\u9ad8\u7684\u901f\u5ea6\u3001\u6548\u7387\u548c\u9069\u61c9\u6027\uff0c\u800c\u4e14\u4e0d\u6703\u5f71\u97ff\u6e96\u78ba\u6027\u3002<\/p>\n\n\n\n

\u9019\u7a2e\u8f49\u8b8a\u4e0d\u50c5\u50c5\u662f\u7406\u8ad6\u4e0a\u7684\uff0c\u800c\u662f\u5df2\u7d93\u5728\u767c\u751f\u3002\u5728\u00a0SC24\u00a0\u4e0a\uff0c\u5169\u500b\u00a0Gordon Bell\u00a0\u5165\u570d\u8a08\u756b\u63ed\u793a\u5982\u4f55\u4f7f\u7528 AI \u548c\u6df7\u5408\u7cbe\u5ea6\u4f86\u5e6b\u52a9\u63a8\u9032\u57fa\u56e0\u7d44\u5b78\u548c\u86cb\u767d\u8cea\u8a2d\u8a08\u3002<\/p>\n\n\n\n

\u6c99\u70cf\u5730\u963f\u62c9\u4f2f\u963f\u5e03\u90fd\u62c9\u570b\u738b\u79d1\u6280\u5927\u5b78\uff08King Abdullah University of Science and Technology\uff09\u6559\u6388 David Keyes\u00a0\u5728\u5176\u984c\u70ba\u300c\u4f7f\u7528\u6df7\u5408\u7cbe\u5ea6\u65bc\u57fa\u56e0\u7d44\u5b78\u7814\u7a76\uff08Using Mixed Precision for Genomics\uff09\u300d\u7684\u8ad6\u6587\u4e2d\uff0c\u4f7f\u7528 0.8 exaflops\u00a0\u7684\u6df7\u5408\u7cbe\u5ea6\u4f86\u63a2\u7d22\u57fa\u56e0\u7d44\u8207\u5176\u5ee3\u7fa9\u57fa\u56e0\u578b\u4e4b\u9593\u7684\u95dc\u4fc2\uff0c\u9032\u4e00\u6b65\u63a2\u7d22\u8207\u75be\u75c5\u76db\u884c\u7387\u7684\u95dc\u806f\u6027\u3002<\/p>\n\n\n\n

\u540c\u6a23\u5730\uff0c\u4f86\u81ea\u7f8e\u570b\u963f\u8ca2\u570b\u5bb6\u5be6\u9a57\u5ba4\uff08Argonne National Laboratory\uff09\u7684\u904b\u7b97\u751f\u7269\u5b78\u5bb6 Arvind Ramanathan \u5728 NVIDIA Grace Hopper \u9a45\u52d5\u7684 Alps \u7cfb\u7d71\u4e0a\uff0c\u5229\u7528\u4e86 3 exaflops \u7684 AI \u6548\u80fd\u4f86\u52a0\u901f\u86cb\u767d\u8cea\u8a2d\u8a08\u3002<\/p>\n\n\n\n

\u70ba\u4e86\u9032\u4e00\u6b65\u63a8\u9032\u00a0AI\u00a0\u9a45\u52d5\u7684\u85e5\u7269\u63a2\u7d22\u548c\u6551\u547d\u7642\u6cd5\u7684\u958b\u767c\uff0c\u7814\u7a76\u4eba\u54e1\u53ef\u4ee5\u4f7f\u7528\u00a0NVIDIA BioNeMo<\/a>\uff0c\u9019\u662f\u5c08\u70ba\u88fd\u85e5\u61c9\u7528\u8a2d\u8a08\u7684\u5f37\u5927\u5de5\u5177\u3002BioNeMo\u00a0\u6846\u67b6<\/a>\u00a0\u73fe\u5df2\u958b\u6e90\uff0c\u53ef\u52a0\u901f\u7528\u65bc\u85e5\u7269\u63a2\u7d22\u548c\u5206\u5b50\u8a2d\u8a08\u7684 AI \u6a21\u578b\u5efa\u7acb\u3001\u5ba2\u88fd\u5316\u548c\u90e8\u7f72\u3002<\/p>\n\n\n\n

\u5728\u00a0TOP500\u00a0\u4e2d\uff0cAI \u548c\u6df7\u5408\u7cbe\u5ea6\u6d6e\u9ede\u904b\u7b97\u7684\u5ee3\u6cdb\u61c9\u7528\u53cd\u6620\u4e86\u5168\u7403\u904b\u7b97\u91cd\u9ede\u7684\u8f49\u8b8a\u3002\u73fe\u5728\uff0cTOP500\u00a0\u7cfb\u7d71\u53ef\u7372\u5f97\u7e3d\u5171\u00a0249 exaflops\u00a0\u7684\u00a0AI\u00a0\u6548\u80fd\uff0c\u63a8\u52d5\u4e86\u5404\u7522\u696d\u7684\u5275\u65b0\u548c\u767c\u73fe\u3002<\/p>\n\n\n

\n
\"\"
\u25b2TOP500\u00a0\u7d71\u8a08\u00a0AI\u3001FP32\u00a0\u548c\u00a0FP64 FLOPS\u00a0\u5e74\u5ea6\u8b8a\u5316<\/figcaption><\/figure><\/div>\n\n\n

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

NVIDIA \u52a0\u901f\u7684 TOP500 \u7cfb\u7d71\u5728 AI \u548c\u6df7\u5408\u7cbe\u5ea6\u7cfb\u7d71\u6548\u80fd\u7b49\u95dc\u9375\u6307\u6a19\u4e0a\u8868\u73fe\u5353\u8d8a\u3002NVIDIA \u7684\u52a0\u901f\u904b\u7b97\u5e73\u53f0\u4ee5\u8d85\u904e 190 exaflops \u7684 AI \u904b\u7b97\u6548\u80fd\u548c 17 exaflops \u7684\u55ae\u7cbe\u5ea6\uff08FP32\uff09\u6548\u80fd\uff0c\u6210\u70ba\u79d1\u5b78\u904b\u7b97\u7684\u65b0\u5f15\u64ce\u3002\u6b64\u5916\uff0cNVIDIA \u9084\u63d0\u4f9b 4 exaflops \u7684\u96d9\u7cbe\u5ea6\uff08FP64\uff09\u6548\u80fd\uff0c\u7528\u65bc\u4ecd\u9700\u8981\u9ad8\u7cbe\u5ea6\u7684\u7279\u5b9a\u79d1\u5b78\u8a08\u7b97\u3002<\/p>\n\n\n\n

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

\u52a0\u901f\u904b\u7b97\u5373\u6c38\u7e8c\u904b\u7b97<\/strong><\/strong><\/h3>\n\n\n\n

\u96a8\u8457\u5c0d\u904b\u7b97\u80fd\u529b\u9700\u6c42\u7684\u589e\u9577\uff0c\u6c38\u7e8c\u6027\u7684\u9700\u6c42\u4e5f\u8207\u65e5\u4ff1\u589e\u3002<\/p>\n\n\n\n

\u5728\u5168\u7403\u6700\u7bc0\u80fd\u8d85\u7d1a\u96fb\u8166 Green500 \u540d\u55ae\u4e2d\uff0c\u63a1\u7528 NVIDIA \u52a0\u901f\u904b\u7b97\u7684\u7cfb\u7d71\u540d\u5217\u524d 10 \u540d\u4e2d\u7684 8 \u540d\u3002\u4f8b\u5982\uff0cEuroHPC\/FZJ \u7684 JEDI \u7cfb\u7d71\u5be6\u73fe\u4e86\u9a5a\u4eba\u7684\u6bcf\u74e6 72.7 gigaflops\uff0c\u70ba\u6548\u80fd\u8207\u6c38\u7e8c\u6027\u7d50\u5408\u6a39\u7acb\u4e86\u65b0\u6a19\u687f\u3002<\/p>\n\n\n\n

\u91dd\u5c0d\u6c23\u5019\u9810\u6e2c\uff0cNVIDIA \u5728 SC24 \u5ba3\u5e03\u4e86\u5c08\u70ba NVIDIA Earth-2 \u8a2d\u8a08\u7684\u5169\u9805\u5168\u65b0 NIM \u5fae\u670d\u52d9<\/a>\u3002Earth-2 \u662f\u7528\u65bc\u6a21\u64ec\u548c\u8996\u89ba\u5316\u5929\u6c23\u8207\u6c23\u5019\u689d\u4ef6\u7684\u6578\u4f4d\u5b7f\u751f\u5e73\u53f0\u3002CorrDiff NIM \u548c FourCastNet NIM \u5fae\u670d\u52d9\u53ef\u5c07\u6c23\u5019\u8b8a\u9077\u5efa\u6a21\u548c\u6a21\u64ec\u7d50\u679c\u52a0\u901f\u9ad8\u9054 500 \u500d\u3002<\/p>\n\n\n\n

\u5728\u4e00\u500b\u74b0\u5883\u8db3\u8de1\u65e5\u76ca\u53d7\u5230\u95dc\u6ce8\u7684\u4e16\u754c\u4e2d\uff0cNVIDIA \u5728\u52a0\u901f\u904b\u7b97\u65b9\u9762\u7684\u5275\u65b0\uff0c\u517c\u9867\u4e86\u9ad8\u6548\u80fd\u8207\u80fd\u6e90\u6548\u7387\uff0c\u5354\u52a9\u5be6\u73fe\u66f4\u5149\u660e\u3001\u66f4\u6c38\u7e8c\u7684\u672a\u4f86\u3002<\/p>\n\n\n\n

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

\u8d85\u7d1a\u904b\u7b97\u793e\u7fa4\u64c1\u62b1NVIDIA<\/strong><\/h3>\n\n\n\n

NVIDIA \u7372\u5f97\u7684\u00a011\u00a0\u5ea7\u00a0HPCwire\u00a0\u8b80\u8005\u548c\u7de8\u8f2f\u7968\u9078\u734e\uff0c\u4ee3\u8868\u4e86\u6574\u500b\u79d1\u5b78\u754c\u7684\u5de5\u4f5c\u6210\u679c\uff0c\u6db5\u84cb\u5de5\u7a0b\u5e2b\u3001\u958b\u767c\u4eba\u54e1\u3001\u7814\u7a76\u4eba\u54e1\u3001\u5408\u4f5c\u5925\u4f34\u3001\u5ba2\u6236\u7b49\u3002<\/p>\n\n\n\n

\u9019\u4e9b\u734e\u9805\u5305\u542b\uff1a<\/p>\n\n\n\n