{"id":4154,"date":"2023-10-09T08:20:27","date_gmt":"2023-10-09T00:20:27","guid":{"rendered":""},"modified":"2023-10-09T08:20:27","modified_gmt":"2023-10-09T00:20:27","slug":"","status":"publish","type":"post","link":"http:\/\/www.idea2003.com\/dongtai\/4154.html","title":{"rendered":"AMD \u84c4\u52bf\u5f85\u53d1\u51c6\u5907\u6311\u6218\u82f1\u4f1f\u8fbe\u7684\u4eba\u5de5\u667a\u80fd\u9738\u4e3b\u5730\u4f4d"},"content":{"rendered":"

\u5fae\u65b0\u521b\u60f3(idea2003.com) 10\u67088\u65e5\u6d88\u606f:<\/strong>AMD<\/span> \u6b63\u5728\u52a0\u7d27\u6311\u6218\u82f1\u4f1f\u8fbe<\/span>\u5728 AI \u5e02\u573a\u8fd1\u4e4e\u5784\u65ad\u7684\u5730\u4f4d\u3002\u4ee5 GPU \u4e3b\u5bfc\u5730\u4f4d\u800c\u95fb\u540d\u7684\u82f1\u4f1f\u8fbe\u4e00\u76f4\u4ee5\u5176\u4e13\u6709\u7684 CUDA \u8f6f\u4ef6\u548c PyTorch \u65e5\u76ca\u589e\u957f\u7684\u6d41\u884c\u5ea6\u6765\u5de9\u56fa\u5176\u7edf\u6cbb\u5730\u4f4d\uff0c\u5176\u4e2d 92% \u7684\u4eba\u5de5\u667a\u80fd<\/span>\u6a21\u578b\u90fd\u662f PyTorch \u72ec\u5360\u7684\u3002<\/span><\/p>\n

\n

\u7136\u800c\uff0cAMD \u51ed\u501f\u5176\u4e0e PyTorch \u7684\u5408\u4f5c\u4ee5\u53ca\u5fae\u8f6f\u7684\u652f\u6301\uff0c\u6709\u671b\u98a0\u8986\u73b0\u72b6<\/strong>\u3002PyTorch \u7684\u80fd\u529b\u4f7f\u7528\u6237\u514d\u53d7\u5e95\u5c42 GPU \u67b6\u6784\u7684\u5f71\u54cd\uff0c\u5e76\u652f\u6301 AMD GPU\uff0c\u5df2\u7ecf\u5f00\u59cb\u4e3a AMD \u548c\u5176\u4ed6\u4f9b\u5e94\u5546\u5728\u82f1\u4f1f\u8fbe\u7684\u5f3a\u52bf\u5730\u4f4d\u4e2d\u5f00\u8f9f\u4e86\u9053\u8def\u3002AMD \u7684\u79d8\u5bc6\u6b66\u5668\u662f\u5373\u5c06\u63a8\u51fa\u7684 Instinct MI300A \u5904\u7406\u5668<\/strong>\uff0c\u65e8\u5728\u4e0e\u82f1\u4f1f\u8fbe\u7684 Grace-Hopper \u8d85\u7ea7<\/span>\u82af\u7247\u7ade\u4e89\u3002<\/p>\n

\u8fdb\u4e00\u6b65\u4e86\u89e3\u82f1\u4f1f\u8fbe\u7684\u4e3b\u5bfc\u5730\u4f4d<\/strong><\/p>\n

\u5728\u4eba\u5de5\u667a\u80fd\u9886\u57df\uff0c\u82f1\u4f1f\u8fbe\u7684\u7edf\u6cbb\u5730\u4f4d\u663e\u800c\u6613\u89c1\u3002\u672f\u8bed\u300cGPU\u300d\u51e0\u4e4e\u7b49\u540c\u4e8e\u82f1\u4f1f\u8fbe\uff0c\u8fd9\u4e3b\u8981\u5f52\u529f\u4e8e\u56f4\u7ed5\u5176\u786c\u4ef6\u6784\u5efa\u7684\u7cbe\u5fc3\u6784\u5efa\u7684\u8f6f\u4ef6\u300c\u62a4\u57ce\u6cb3\u300d\u3002\u8fd9\u4e2a\u300c\u62a4\u57ce\u6cb3\u300d\u5efa\u7acb\u5728\u8bf8\u5982 CUDA \u4e4b\u7c7b\u7684\u5de5\u5177\u5468\u56f4\uff0cCUDA \u662f\u82f1\u4f1f\u8fbe\u521b\u5efa\u7684\u4e00\u4e2a API\uff0c\u5141\u8bb8\u8f6f\u4ef6\u5f00\u53d1\u4eba\u5458\u4f7f\u7528\u652f\u6301 CUDA \u7684\u56fe\u5f62\u5904\u7406\u5355\u5143\uff08GPU\uff09\u8fdb\u884c\u901a\u7528\u5904\u7406<\/strong>\u3002<\/p>\n

CUDA \u7684\u6d41\u884c\u5ea6\u5f97\u5230\u4e86 PyTorch \u7684\u5d1b\u8d77\u7684\u63a8\u52a8\uff0cPyTorch \u662f\u4e00\u4e2a\u5f00\u6e90\u7684\u673a\u5668\u5b66\u4e60\u6846\u67b6\uff0c\u5df2\u7ecf\u6210\u4e3a GenAI \u7684\u9996\u9009<\/span>\u5de5\u5177\uff0c\u6709\u8d85\u8fc7 92% \u7684\u53ef\u7528\u6a21\u578b\u662f PyTorch \u72ec\u5360\u7684<\/strong>\u3002<\/p>\n

\u82f1\u4f1f\u8fbe\u7684\u7edf\u6cbb\u5730\u4f4d\u8fdb\u4e00\u6b65\u4f53\u73b0\u5728\u5176\u5728 GPU \u6570\u636e\u4e2d\u5fc3\u5e02\u573a\u4e0a\u7684\u63a7\u5236\uff0c\u5e02\u573a\u4efd\u989d\u9ad8\u8fbe 95% \u4ee5\u4e0a\u3002\u82f1\u4f1f\u8fbe\u6700\u65b0<\/span>\u53d1\u5e03\u7684 DGX H100 \u7cfb\u7edf\u5df2\u7ecf\u5f00\u59cb\u5728\u5168\u7403\u8303\u56f4\u5185\u53d1\u8d27\uff0c\u914d\u5907\u4e86 8 \u9897\u300cHopper H100\u300d\u6570\u636e\u4e2d\u5fc3\u82af\u7247\uff0c\u62e5\u6709\u60ca\u4eba\u7684 4\uff0c000 \u4e07\u4ebf\u6b21\u7684\u8ba1\u7b97\u80fd\u529b\u3002\u8fd9\u4e9b\u7cfb\u7edf\u6b63\u5728\u5404\u4e2a\u884c\u4e1a\u90e8\u7f72\uff0c\u4ece\u533b\u7597\u4fdd\u5065\u5230\u91d1\u878d\uff0c\u52a0\u5f3a\u4e86\u82f1\u4f1f\u8fbe\u5728\u4eba\u5de5\u667a\u80fd\u9886\u57df\u65e0\u4e0e\u4f26\u6bd4\u7684\u5f71\u54cd\u529b\u3002<\/p>\n

AMD \u96c4\u5fc3\u52c3\u52c3\u5730\u8fdb\u519b\u4eba\u5de5\u667a\u80fd\u9886\u57df<\/strong><\/p>\n

\u5c3d\u7ba1\u82f1\u4f1f\u8fbe\u7684\u5730\u4f4d\u770b\u4f3c\u575a\u4e0d\u53ef\u6467\uff0c\u4f46 AMD \u4e00\u76f4\u5728\u9ed8\u9ed8\u53d6\u5f97\u8fdb\u5c55<\/strong>\u3002\u8be5\u516c\u53f8\u5f00\u53d1\u4e86\u4e00\u6b3e\u540d\u4e3a HIP \u7684 CUDA \u8f6c\u6362\u5de5\u5177\uff0c\u800c\u5373\u5c06\u63a8\u51fa\u7684 Instinct MI300A \u5904\u7406\u5668\u88ab\u89c6\u4e3a\u4e0e\u82f1\u4f1f\u8fbe\u7684\u8d85\u7ea7<\/span>\u82af\u7247\u5f3a\u6709\u7ade\u4e89\u529b\u7684\u4ea7\u54c1\u3002<\/strong>\u8be5\u5904\u7406\u5668\u4e13\u4e3a\u63a8\u7406\u5e94\u7528\u8bbe\u8ba1\uff0c\u8fd9\u662f GenAI \u4e2d\u9700\u8981\u52a0\u901f\u8ba1\u7b97\u7684\u5173\u952e\u4efb\u52a1\uff0c\u5e76\u65e8\u5728\u786e\u7acb AMD \u4f5c\u4e3a\u63a8\u7406\u89e3\u51b3\u65b9\u6848\u9886\u57df\u7684\u884c\u4e1a\u9886\u8896<\/span>\u3002<\/p>\n

AMD \u7684\u52aa\u529b\u5f97\u5230\u4e86\u4e0e PyTorch \u7684\u5408\u4f5c\u548c\u79d1\u6280\u5de8\u5934\u5fae\u8f6f\u7684\u652f\u6301\u3002\u4f5c\u4e3a PyTorch \u57fa\u91d1\u4f1a\u7684\u521b\u59cb\u6210\u5458\u548c\u8463\u4e8b\u4f1a\u4ee3\u8868\uff0cAMD \u4e3a ROCm \u5f00\u6e90\u8f6f\u4ef6\u751f\u6001\u7cfb\u7edf\u63d0\u4f9b\u4e86\u66f4\u65b0\uff0c\u4e3a PyTorch \u63d0\u4f9b\u4e86\u5bf9 AMD Instinct \u52a0\u901f\u5668\u548c Radeon GPU \u7684\u7a33\u5b9a\u652f\u6301\u3002ROCm \u662f CUDA \u7684\u66ff\u4ee3\u54c1\u3002\u8fd9\u79cd\u5408\u4f5c\u5bfc\u81f4\u4e86 AI \u6a21\u578b\u6027\u80fd\u548c\u53ef\u6269\u5c55\u6027\u7684\u63d0\u9ad8\uff0c\u4f7f\u5f00\u53d1\u4eba\u5458\u80fd\u591f\u5229\u7528 AMD GPU \u52a0\u901f\u5668\u548c ROCm \u8f6f\u4ef6\u751f\u6001\u7cfb\u7edf\u6784\u5efa AI \u89e3\u51b3\u65b9\u6848\u3002<\/p>\n

\u5fae\u8f6f\u4e0e AMD \u7ed3\u6210\u6218\u7565\u8054\u76df<\/strong><\/p>\n

\u5fae\u8f6f\u4e0e AMD \u7684\u5408\u4f5c\u8868\u660e\u4e86\u4ed6\u4eec\u8981\u5728 AI \u82af\u7247\u5e02\u573a\u4e2d\u83b7\u5f97\u66f4\u591a\u7684 AI \u80fd\u529b\u7ec4\u4ef6\u5e76\u63d0\u4f9b\u82f1\u4f1f\u8fbe\u7edf\u6cbb\u5730\u4f4d\u7684\u66ff\u4ee3\u65b9\u6848\u7684\u6218\u7565\u4e3e\u63aa\u3002\u8fd9\u4e24\u5bb6\u516c\u53f8\u6b63\u5728\u5171\u540c\u5f00\u53d1\u4e00\u6b3e\u9762\u5411 AI \u5de5\u4f5c\u8d1f\u8f7d\u7684\u5fae\u8f6f\u5904\u7406\u5668\uff0c\u4ee3\u53f7\u4e3a Athena\uff0c\u8868\u660e\u4e86\u4ed6\u4eec\u6253\u51fb\u82f1\u4f1f\u8fbe\u7edf\u6cbb\u5730\u4f4d\u7684\u591a\u7ba1\u9f50\u4e0b\u7b56\u7565<\/strong>\u3002<\/p>\n

\u5c3d\u7ba1 AMD \u7684 GPU \u6ca1\u6709\u50cf\u82f1\u4f1f\u8fbe\u7684\u90a3\u6837\u88ab\u5e7f\u6cdb\u4f7f\u7528\uff0c\u4f46\u5fae\u8f6f\u7684\u652f\u6301\u548c\u5de5\u7a0b\u8d44\u6e90\u53ef\u80fd\u4f1a\u63d0\u9ad8 AMD \u4ea7\u54c1\u5728 AI \u5de5\u4f5c\u8d1f\u8f7d\u4e2d\u7684\u6027\u80fd\u3002\u8fd9\u79cd\u5408\u4f5c\u53ef\u80fd\u4f1a\u4e3a\u52a0\u901f AI \u786c\u4ef6\u521b\u9020\u4e00\u4e2a\u66f4\u5177\u7ade\u4e89\u529b\u7684\u5e02\u573a\uff0c\u6f5c\u5728\u964d\u4f4e\u5fae\u8f6f\u7684\u670d\u52a1\u5668\u6210\u672c\u3002<\/p>\n

AMD \u98a0\u8986\u4eba\u5de5\u667a\u80fd\u9886\u57df\u7684\u6f5c\u529b<\/strong><\/p>\n

\u6700\u8fd1\u7684\u57fa\u51c6\u6d4b\u8bd5\u663e\u793a\uff0cAMD \u7684 MI250 GPU \u7684\u6027\u80fd\u7ea6\u4e3a\u82f1\u4f1f\u8fbe\u7684 A100 \u7684 80%<\/strong>\uff0c\u4eba\u4eec\u5bf9\u8fd9\u4e24\u5bb6\u79d1\u6280\u5de8\u5934\u5728\u8bad\u7ec3\u5927\u89c4\u6a21\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u9886\u57df\u7684\u6f5c\u5728\u7ade\u4e89\u611f\u5230\u5174\u594b\u3002\u4e0b\u4e00\u8f6e\u7ade\u4e89\u53ef\u80fd\u4f1a\u53d1\u751f\u5728 AMD \u5373\u5c06\u63a8\u51fa\u7684 MI300 \u548c\u82f1\u4f1f\u8fbe\u7684 H100 \u4e4b\u95f4\uff0c\u5176\u4e2d MI300 \u7684\u66f4\u5927\u5185\u5b58\u5bb9\u91cf\u548c\u5e26\u5bbd\u53ef\u80fd\u4f1a\u88ab H100 \u7684 Transformer \u5f15\u64ce\u548c\u5bf9\u66f4\u591a HBM \u7684\u652f\u6301\u6240\u62b5\u6d88\u3002<\/strong><\/p>\n

AMD \u7684\u6218\u7565\u5408\u4f5c\u4f19\u4f34\u5173\u7cfb\uff0c\u52a0\u4e0a\u4ed6\u4eec\u4e3a\u5373\u5c06\u63a8\u51fa\u7684\u5904\u7406\u5668\u6240\u505a\u7684\u67b6\u6784\u9009\u62e9\uff0c\u6709\u53ef\u80fd\u5728\u4eba\u5de5\u667a\u80fd\u9886\u57df\u98a0\u8986\u82f1\u4f1f\u8fbe\u7684\u5784\u65ad\u5730\u4f4d\u3002\u7136\u800c\uff0c\u4eba\u5de5\u667a\u80fd\u884c\u4e1a\u4ecd\u5904\u4e8e\u65e9\u671f\u9636\u6bb5\uff0c\u5728\u8fd9\u4e2a\u5e02\u573a\u4e0a\u8d62\u5f97\u786c\u4ef6\u7ade\u4e89\u5c06\u53d6\u51b3\u4e8e\u6027\u80fd\u3001\u53ef\u79fb\u690d\u6027\u548c\u53ef\u7528\u6027<\/strong>\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

\u5fae\u65b0\u521b\u60f3(idea2003.com) 10\u67088\u65e5\u6d88\u606f:AMD \u6b63\u5728\u52a0...<\/p>\n","protected":false},"author":1,"featured_media":831,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[19],"tags":[],"class_list":["post-4154","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-dongtai"],"_links":{"self":[{"href":"http:\/\/www.idea2003.com\/wp-json\/wp\/v2\/posts\/4154","targetHints":{"allow":["GET"]}}],"collection":[{"href":"http:\/\/www.idea2003.com\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/www.idea2003.com\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/www.idea2003.com\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/www.idea2003.com\/wp-json\/wp\/v2\/comments?post=4154"}],"version-history":[{"count":0,"href":"http:\/\/www.idea2003.com\/wp-json\/wp\/v2\/posts\/4154\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"http:\/\/www.idea2003.com\/wp-json\/wp\/v2\/media\/831"}],"wp:attachment":[{"href":"http:\/\/www.idea2003.com\/wp-json\/wp\/v2\/media?parent=4154"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/www.idea2003.com\/wp-json\/wp\/v2\/categories?post=4154"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/www.idea2003.com\/wp-json\/wp\/v2\/tags?post=4154"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}