{"id":1812,"date":"2023-10-08T09:56:10","date_gmt":"2023-10-08T01:56:10","guid":{"rendered":"http:\/\/shlj-tec.com\/?p=1812"},"modified":"2023-10-08T09:56:10","modified_gmt":"2023-10-08T01:56:10","slug":"nvidia-a100","status":"publish","type":"post","link":"http:\/\/shlj-tec.com\/index.php\/2023\/10\/08\/nvidia-a100\/","title":{"rendered":"NVIDIA A100"},"content":{"rendered":"<p><img decoding=\"async\" loading=\"lazy\" class=\"alignnone  wp-image-1813\" src=\"http:\/\/shlj-tec.com\/wp-content\/uploads\/2023\/10\/1696729322913-300x295.png\" alt=\"\" width=\"949\" height=\"933\" \/><\/p>\n<h2><strong><b>\u52a0\u901f\u5f53\u4eca\u65f6\u4ee3\u7684\u91cd\u8981\u5de5\u4f5c<\/b><\/strong><\/h2>\n<p>NVIDIA A100 Tensor Core GPU \u53ef\u9488\u5bf9 AI\u3001\u6570\u636e\u5206\u6790\u548c HPC \u5e94\u7528\u573a\u666f\uff0c\u5728\u4e0d\u540c\u89c4\u6a21\u4e0b\u5b9e\u73b0\u51fa\u8272\u7684\u52a0\u901f\uff0c\u6709\u6548\u52a9\u529b\u66f4\u9ad8\u6027\u80fd\u7684\u5f39\u6027\u6570\u636e\u4e2d\u5fc3\u3002A100 \u91c7\u7528 NVIDIA Ampere \u67b6\u6784\uff0c\u662f NVIDIA \u6570\u636e\u4e2d\u5fc3\u5e73\u53f0\u7684\u5f15\u64ce\u3002A100 \u7684\u6027\u80fd\u6bd4\u4e0a\u4e00\u4ee3\u4ea7\u54c1\u63d0\u5347\u9ad8\u8fbe 20 \u500d\uff0c\u5e76\u53ef\u5212\u5206\u4e3a\u4e03\u4e2a GPU \u5b9e\u4f8b\uff0c\u4ee5\u6839\u636e\u53d8\u5316\u7684\u9700\u6c42\u8fdb\u884c\u52a8\u6001\u8c03\u6574\u3002A100 \u63d0\u4f9b 40GB \u548c 80GB \u663e\u5b58\u4e24\u79cd\u7248\u672c\uff0cA100 80GB \u5c06 GPU \u663e\u5b58\u589e\u52a0\u4e86\u4e00\u500d\uff0c\u5e76\u63d0\u4f9b\u8d85\u5feb\u901f\u7684\u663e\u5b58\u5e26\u5bbd\uff08\u6bcf\u79d2\u8d85\u8fc7 2 \u4e07\u4ebf\u5b57\u8282 [TB\/s]\uff09\uff0c\u53ef\u5904\u7406\u8d85\u5927\u578b\u6a21\u578b\u548c\u6570\u636e\u96c6\u3002<\/p>\n<h2><strong><b>\u9002\u7528\u4e8e AI \u7684\u4f01\u4e1a\u5c31\u7eea\u8f6f\u4ef6<\/b><\/strong><\/h2>\n<p>NVIDIA EGX\u2122\u00a0\u5e73\u53f0\u4e2d\u5305\u62ec\u80fd\u591f\u5728\u6574\u4e2a\u57fa\u7840\u67b6\u6784\u4e2d\u5b9e\u73b0\u52a0\u901f\u8ba1\u7b97\u7684\u4f18\u5316\u8f6f\u4ef6\u3002\u5229\u7528 NVIDIA AI Enterprise \u8f6f\u4ef6\u5957\u4ef6\uff0c\u4f01\u4e1a\u53ef\u4ee5\u8bbf\u95ee\u7aef\u5230\u7aef\u7684\u4e91\u539f\u751f AI \u548c\u6570\u636e\u5206\u6790\u8f6f\u4ef6\uff0c\u8be5\u8f6f\u4ef6\u5957\u4ef6\u7ecf NVIDIA \u4f18\u5316\u3001\u8ba4\u8bc1\u548c\u652f\u6301\uff0c\u53ef\u5728\u57fa\u4e8e NVIDIA \u8ba4\u8bc1\u7cfb\u7edf\u2122 \u7684 VMware vSphere \u4e0a\u8fd0\u884c\u3002NVIDIA AI \u4f01\u4e1a\u5305\u62ec NVIDIA \u7684\u5173\u952e\u652f\u6301\u6280\u672f\uff0c\u7528\u4e8e\u5728\u73b0\u4ee3\u6df7\u5408\u4e91\u4e2d\u5feb\u901f\u90e8\u7f72\u3001\u7ba1\u7406\u548c\u6269\u5c55 AI \u5de5\u4f5c\u8d1f\u8f7d<\/p>\n<p>\u5f53\u4eca\u7684 AI \u6a21\u578b\u9762\u4e34\u7740\u5bf9\u8bdd\u5f0f AI \u7b49\u66f4\u9ad8\u5c42\u6b21\u7684\u6311\u6218\uff0c\u8fd9\u4fc3\u4f7f\u5176\u590d\u6742\u5ea6\u5448\u7206\u70b8\u5f0f\u589e\u957f\u3002\u8bad\u7ec3\u8fd9\u4e9b\u6a21\u578b\u9700\u8981\u5927\u89c4\u6a21\u7684\u8ba1\u7b97\u80fd\u529b\u548c\u53ef\u6269\u5c55\u6027\u3002<\/p>\n<p>NVIDIA A100\u00a0<a href=\"https:\/\/www.nvidia.cn\/data-center\/tensorcore\/\"><u>Tensor Core<\/u><\/a>\u00a0\u501f\u52a9 Tensor \u6d6e\u70b9\u8fd0\u7b97 (TF32) \u7cbe\u5ea6\uff0c\u53ef\u63d0\u4f9b\u6bd4 NVIDIA Volta \u9ad8 20 \u500d\u4e4b\u591a\u7684\u6027\u80fd\uff0c\u5e76\u4e14\u65e0\u9700\u66f4\u6539\u4ee3\u7801\uff1b\u82e5\u4f7f\u7528\u81ea\u52a8\u6df7\u5408\u7cbe\u5ea6\u548c FP16\uff0c\u6027\u80fd\u53ef\u8fdb\u4e00\u6b65\u63d0\u5347 2 \u500d\u3002\u4e0e NVIDIA\u00ae\u00a0<a href=\"https:\/\/www.nvidia.cn\/data-center\/nvlink\/\"><u>NVLink<\/u><u>\u00ae<\/u><\/a>\u3001NVIDIA\u00a0<a href=\"https:\/\/www.nvidia.cn\/data-center\/nvlink\/\"><u>NVSwitch<\/u><u>\u2122<\/u><\/a>\u3001PCIe 4.0\u3001NVIDIA\u00ae\u00a0InfiniBand\u00ae\u00a0\u548c\u00a0<a href=\"https:\/\/www.nvidia.cn\/data-center\/magnum-io\/\"><u>NVIDIA Magnum IO<\/u><u>\u2122<\/u><\/a>\u00a0SDK \u7ed3\u5408\u4f7f\u7528\u65f6\uff0c\u5b83\u80fd\u6269\u5c55\u5230\u6570\u5343\u4e2a A100 GPU\u3002<\/p>\n<p>2048 \u4e2a A100 GPU \u53ef\u5728\u4e00\u5206\u949f\u5185\u6210\u89c4\u6a21\u5730\u5904\u7406 BERT \u4e4b\u7c7b\u7684\u8bad\u7ec3\u5de5\u4f5c\u8d1f\u8f7d\uff0c\u8fd9\u662f\u975e\u5e38\u5feb\u901f\u7684\u89e3\u51b3\u95ee\u9898\u901f\u5ea6\u3002<\/p>\n<p>\u5bf9\u4e8e\u5177\u6709\u5e9e\u5927\u6570\u636e\u8868\u7684\u8d85\u5927\u578b\u6a21\u578b\uff08\u4f8b\u5982\u6df1\u5ea6\u5b66\u4e60\u63a8\u8350\u6a21\u578b [DLRM]\uff09\uff0cA100 80GB \u53ef\u4e3a\u6bcf\u4e2a\u8282\u70b9\u63d0\u4f9b\u9ad8\u8fbe 1.3TB \u7684\u7edf\u4e00\u663e\u5b58\uff0c\u800c\u4e14\u541e\u5410\u91cf\u6bd4 A100 40GB \u591a\u9ad8\u8fbe 3 \u500d\u3002<\/p>\n<p>NVIDIA \u7684\u9886\u5148\u5730\u4f4d\u5728\u00a0<a href=\"https:\/\/www.nvidia.cn\/data-center\/resources\/mlperf-benchmarks\/\"><u>MLPerf<\/u><\/a>\u00a0\u8fd9\u4e2a\u884c\u4e1a\u7ea7 AI \u8bad\u7ec3\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u5f97\u5230\u5370\u8bc1\uff0c\u521b\u4e0b\u591a\u9879\u6027\u80fd\u7eaa\u5f55\u3002<\/p>\n<h3><strong><b>\u6df1\u5ea6\u5b66\u4e60\u63a8\u7406<\/b><\/strong><\/h3>\n<p>A100 \u5f15\u5165\u4e86\u7a81\u7834\u6027\u7684\u529f\u80fd\u6765\u4f18\u5316\u63a8\u7406\u5de5\u4f5c\u8d1f\u8f7d\u3002\u5b83\u80fd\u5728\u4ece FP32 \u5230 INT4 \u7684\u6574\u4e2a\u7cbe\u5ea6\u8303\u56f4\u5185\u8fdb\u884c\u52a0\u901f\u3002\u591a\u5b9e\u4f8b GPU (<a href=\"https:\/\/www.nvidia.cn\/technologies\/multi-instance-gpu\/\"><u>MIG<\/u><\/a>) \u6280\u672f\u5141\u8bb8\u591a\u4e2a\u7f51\u7edc\u540c\u65f6\u57fa\u4e8e\u5355\u4e2a A100 \u8fd0\u884c\uff0c\u4ece\u800c\u4f18\u5316\u8ba1\u7b97\u8d44\u6e90\u7684\u5229\u7528\u7387\u3002\u5728 A100 \u5176\u4ed6\u63a8\u7406\u6027\u80fd\u589e\u76ca\u7684\u57fa\u7840\u4e4b\u4e0a\uff0c\u4ec5\u7ed3\u6784\u7a00\u758f\u652f\u6301\u4e00\u9879\u5c31\u80fd\u5e26\u6765\u9ad8\u8fbe\u4e24\u500d\u7684\u6027\u80fd\u63d0\u5347\u3002<\/p>\n<p>\u5728 BERT \u7b49\u5148\u8fdb\u7684\u5bf9\u8bdd\u5f0f AI \u6a21\u578b\u4e0a\uff0cA100 \u53ef\u5c06\u63a8\u7406\u541e\u5410\u91cf\u63d0\u5347\u5230\u9ad8\u8fbe CPU \u7684 249 \u500d\u3002<\/p>\n<p>\u5728\u53d7\u5230\u6279\u91cf\u5927\u5c0f\u9650\u5236\u7684\u6781\u590d\u6742\u6a21\u578b\uff08\u4f8b\u5982\u7528\u4e8e\u81ea\u52a8\u8bed\u97f3\u8bc6\u522b\u7528\u9014\u7684 RNN-T\uff09\u4e0a\uff0c\u663e\u5b58\u5bb9\u91cf\u6709\u6240\u589e\u52a0\u7684 A100 80GB \u80fd\u4f7f\u6bcf\u4e2a MIG \u7684\u5927\u5c0f\u589e\u52a0\u4e00\u500d\uff0c\u5e76\u63d0\u4f9b\u6bd4 A100 40GB \u9ad8 1.25 \u500d\u7684\u541e\u5410\u91cf\u3002<\/p>\n<p>NVIDIA \u4ea7\u54c1\u7684\u51fa\u8272\u6027\u80fd\u5728\u00a0<a href=\"https:\/\/www.nvidia.cn\/data-center\/resources\/mlperf-benchmarks\/\"><u>MLPerf \u63a8\u7406<\/u><\/a>\u6d4b\u8bd5\u4e2d\u5f97\u5230\u9a8c\u8bc1\u3002A100 \u518d\u5c06\u6027\u80fd\u63d0\u5347\u4e86 20 \u500d\uff0c\u8fdb\u4e00\u6b65\u6269\u5927\u4e86\u8fd9\u79cd\u6027\u80fd\u4f18\u52bf\u3002<\/p>\n<h3><strong><b>\u9ad8\u6027\u80fd\u8ba1\u7b97<\/b><\/strong><\/h3>\n<p>\u4e3a\u4e86\u83b7\u5f97\u65b0\u4e00\u4ee3\u7684\u53d1\u73b0\u6210\u679c\uff0c\u79d1\u5b66\u5bb6\u4eec\u5e0c\u671b\u901a\u8fc7\u4eff\u771f\u6765\u66f4\u597d\u5730\u4e86\u89e3\u6211\u4eec\u5468\u56f4\u7684\u4e16\u754c\u3002<\/p>\n<p>NVIDIA A100 \u91c7\u7528\u53cc\u7cbe\u5ea6 Tensor Core\uff0c\u5b9e\u73b0\u4e86\u81ea GPU \u63a8\u51fa\u4ee5\u6765\u9ad8\u6027\u80fd\u8ba1\u7b97\u6027\u80fd\u7684\u5de8\u5927\u98de\u8dc3\u3002\u7ed3\u5408 80GB \u7684\u8d85\u5feb GPU \u663e\u5b58\uff0c\u7814\u7a76\u4eba\u5458\u53ef\u4ee5\u5728 A100 \u4e0a\u5c06 10 \u5c0f\u65f6\u53cc\u7cbe\u5ea6\u4eff\u771f\u7f29\u77ed\u5230 4 \u5c0f\u65f6\u4ee5\u5185\u3002HPC \u5e94\u7528\u8fd8\u53ef\u4ee5\u5229\u7528 TF32 \u5c06\u5355\u7cbe\u5ea6\u3001\u5bc6\u96c6\u77e9\u9635\u4e58\u6cd5\u8fd0\u7b97\u7684\u541e\u5410\u91cf\u63d0\u9ad8\u9ad8\u8fbe 10 \u500d\u3002<\/p>\n<p>\u5bf9\u4e8e\u5177\u6709\u8d85\u5927\u6570\u636e\u96c6\u7684\u9ad8\u6027\u80fd\u8ba1\u7b97\u5e94\u7528\uff0c\u663e\u5b58\u5bb9\u91cf\u589e\u52a0\u7684 A100 80GB \u53ef\u5728\u8fd0\u884c\u6750\u6599\u4eff\u771f Quantum Espresso \u65f6\u5c06\u541e\u5410\u91cf\u63d0\u5347\u9ad8\u8fbe 2 \u500d\u3002\u6781\u5927\u7684\u663e\u5b58\u5bb9\u91cf\u548c\u8d85\u5feb\u901f\u7684\u663e\u5b58\u5e26\u5bbd\u4f7f A100 80GB \u975e\u5e38\u9002\u5408\u7528\u4f5c\u65b0\u4e00\u4ee3\u5de5\u4f5c\u8d1f\u8f7d\u7684\u5e73\u53f0\u3002<\/p>\n<p><img decoding=\"async\" loading=\"lazy\" class=\"alignnone  wp-image-1814\" src=\"http:\/\/shlj-tec.com\/wp-content\/uploads\/2023\/10\/1696729837892-193x300.png\" alt=\"\" width=\"942\" height=\"1464\" \/><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u52a0\u901f\u5f53\u4eca\u65f6\u4ee3\u7684\u91cd\u8981\u5de5\u4f5c NVIDIA A100 Tensor Core GPU \u53ef\u9488\u5bf9 AI\u3001\u6570\u636e\u5206\u6790\u548c HP [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[74],"tags":[],"_links":{"self":[{"href":"http:\/\/shlj-tec.com\/index.php\/wp-json\/wp\/v2\/posts\/1812"}],"collection":[{"href":"http:\/\/shlj-tec.com\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/shlj-tec.com\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/shlj-tec.com\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/shlj-tec.com\/index.php\/wp-json\/wp\/v2\/comments?post=1812"}],"version-history":[{"count":1,"href":"http:\/\/shlj-tec.com\/index.php\/wp-json\/wp\/v2\/posts\/1812\/revisions"}],"predecessor-version":[{"id":1815,"href":"http:\/\/shlj-tec.com\/index.php\/wp-json\/wp\/v2\/posts\/1812\/revisions\/1815"}],"wp:attachment":[{"href":"http:\/\/shlj-tec.com\/index.php\/wp-json\/wp\/v2\/media?parent=1812"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/shlj-tec.com\/index.php\/wp-json\/wp\/v2\/categories?post=1812"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/shlj-tec.com\/index.php\/wp-json\/wp\/v2\/tags?post=1812"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}