{"id":1498,"date":"2025-02-01T21:42:34","date_gmt":"2025-02-01T13:42:34","guid":{"rendered":"https:\/\/crab.cool\/?p=1498"},"modified":"2025-02-01T21:42:34","modified_gmt":"2025-02-01T13:42:34","slug":"musa%e6%9e%b6%e6%9e%84%e5%92%8ccuda%e7%9a%84%e5%85%b3%e7%b3%bb","status":"publish","type":"post","link":"https:\/\/crab.cool\/index.php\/2025\/02\/01\/musa%e6%9e%b6%e6%9e%84%e5%92%8ccuda%e7%9a%84%e5%85%b3%e7%b3%bb\/","title":{"rendered":"MUSA\u67b6\u6784\u548cCUDA\u7684\u5173\u7cfb"},"content":{"rendered":"<hr \/>\n<h3><strong>MUSA\u67b6\u6784\u4e0eCUDA\u6548\u7387\u5bf9\u6bd4\u5206\u6790<\/strong><\/h3>\n<p>\u6469\u5c14\u7ebf\u7a0b\u7684<strong>MUSA\uff08Moore Threads Unified System Architecture\uff09<\/strong>\u67b6\u6784\u65e8\u5728\u517c\u5bb9\u82f1\u4f1f\u8fbeCUDA\u751f\u6001\uff0c\u4f46\u5176\u8bbe\u8ba1\u76ee\u6807\u5e76\u975e\u5b8c\u5168\u590d\u5236CUDA\uff0c\u800c\u662f\u901a\u8fc7\u786c\u4ef6\u548c\u8f6f\u4ef6\u534f\u540c\u4f18\u5316\uff0c\u5728\u56fd\u4ea7GPU\u8d5b\u9053\u4e0a\u5bfb\u6c42\u5e73\u8861\u3002\u5173\u4e8e\u6548\u7387\u95ee\u9898\uff0c\u9700\u4ece\u591a\u4e2a\u7ef4\u5ea6\u7efc\u5408\u8bc4\u4f30\uff1a<\/p>\n<hr \/>\n<h3><strong>1. \u67b6\u6784\u517c\u5bb9\u6027\u4e0e\u6548\u7387\u635f\u5931\u7684\u6765\u6e90<\/strong><\/h3>\n<h4><strong>\uff081\uff09\u6307\u4ee4\u96c6\u4e0e\u786c\u4ef6\u5dee\u5f02<\/strong><\/h4>\n<ul>\n<li><strong>CUDA\u6838\u5fc3\u8bbe\u8ba1<\/strong>\uff1a\u82f1\u4f1f\u8fbeGPU\u7684SM\uff08Streaming Multiprocessor\uff09\u67b6\u6784\u7ecf\u8fc7\u591a\u5e74\u8fed\u4ee3\uff08\u5982Ampere\u3001Hopper\uff09\uff0c\u5176\u7ebf\u7a0b\u8c03\u5ea6\u3001\u7f13\u5b58\u673a\u5236\u3001Tensor Core\u7b49\u6a21\u5757\u9ad8\u5ea6\u4f18\u5316\uff0c\u4e0eCUDA\u6307\u4ee4\u96c6\u6df1\u5ea6\u7ed1\u5b9a\u3002<\/li>\n<li><strong>MUSA\u67b6\u6784\u8bbe\u8ba1<\/strong>\uff1a\u6469\u5c14\u7ebf\u7a0b\u7684MTT GPU\u786c\u4ef6\u867d\u517c\u5bb9CUDA\u7f16\u7a0b\u6a21\u578b\uff0c\u4f46\u7269\u7406\u5b9e\u73b0\uff08\u5982\u8ba1\u7b97\u5355\u5143\u6bd4\u4f8b\u3001\u5185\u5b58\u5e26\u5bbd\u5206\u914d\uff09\u53ef\u80fd\u4e0e\u82f1\u4f1f\u8fbe\u5b58\u5728\u5dee\u5f02\uff0c\u5bfc\u81f4\u76f8\u540c\u7b97\u6cd5\u5728\u786c\u4ef6\u6267\u884c\u65f6\u6548\u7387\u6ce2\u52a8\u3002<\/li>\n<\/ul>\n<h4><strong>\uff082\uff09\u8f6f\u4ef6\u6808\u5f00\u9500<\/strong><\/h4>\n<ul>\n<li><strong>\u7f16\u8bd1\u4e0e\u8fd0\u884c\u65f6\u8f6c\u6362<\/strong>\uff1aMUSA\u9700\u901a\u8fc7\u517c\u5bb9\u5c42\uff08\u5982MUSA SDK\uff09\u5c06CUDA\u4ee3\u7801\u8f6c\u6362\u4e3a\u81ea\u6709\u6307\u4ee4\u96c6\uff0c\u6b64\u8fc7\u7a0b\u53ef\u80fd\u5f15\u5165\u989d\u5916\u5f00\u9500\uff08\u5982\u6307\u4ee4\u6620\u5c04\u3001\u5185\u5b58\u7ba1\u7406\u4f18\u5316\u4e0d\u8db3\uff09\u3002<\/li>\n<li><strong>\u9a71\u52a8\u4e0e\u5de5\u5177\u94fe\u6210\u719f\u5ea6<\/strong>\uff1aCUDA\u751f\u6001\u7684\u9a71\u52a8\u3001\u5e93\uff08cuDNN\u3001cuBLAS\uff09\u7ecf\u8fc7\u957f\u671f\u4f18\u5316\uff0c\u800cMUSA\u7684\u8f6f\u4ef6\u751f\u6001\u5904\u4e8e\u65e9\u671f\u9636\u6bb5\uff0c\u6027\u80fd\u8c03\u4f18\u7a7a\u95f4\u8f83\u5927\u3002<\/li>\n<\/ul>\n<hr \/>\n<h3><strong>2. \u6548\u7387\u635f\u5931\u7684\u5177\u4f53\u573a\u666f<\/strong><\/h3>\n<h4><strong>\uff081\uff09\u901a\u7528\u8ba1\u7b97\u4efb\u52a1<\/strong><\/h4>\n<ul>\n<li><strong>\u7b80\u5355\u5e76\u884c\u4efb\u52a1<\/strong>\uff1a\u5bf9\u4e8e\u7eaf\u8ba1\u7b97\u5bc6\u96c6\u578b\u4efb\u52a1\uff08\u5982\u77e9\u9635\u4e58\u6cd5\uff09\uff0c\u82e5MUSA\u786c\u4ef6\u7b97\u529b\uff08FP32\/FP16 TOPS\uff09\u5bf9\u6807\u82f1\u4f1f\u8fbe\u540c\u7ea7\u4ea7\u54c1\uff0c\u7406\u8bba\u5cf0\u503c\u6027\u80fd\u53ef\u80fd\u63a5\u8fd1\uff0c\u4f46\u5b9e\u9645\u6548\u7387\u53d7\u5185\u5b58\u5e26\u5bbd\u548c\u8f6f\u4ef6\u4f18\u5316\u5f71\u54cd\u3002<\/li>\n<li><strong>\u6848\u4f8b\u5bf9\u6bd4<\/strong>\uff1a\n<ul>\n<li>\u82e5\u67d0CUDA\u7a0b\u5e8f\u5728\u82f1\u4f1f\u8fbeA100\u4e0a\u8fbe\u523080%\u786c\u4ef6\u5229\u7528\u7387\uff0c\u79fb\u690d\u5230MUSA GPU\u540e\u53ef\u80fd\u56e0\u6307\u4ee4\u8c03\u5ea6\u6216\u7f13\u5b58\u672a\u547d\u4e2d\u964d\u81f360-70%\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<h4><strong>\uff082\uff09\u590d\u6742AI\u6a21\u578b\u63a8\u7406<\/strong><\/h4>\n<ul>\n<li><strong>\u52a8\u6001\u8ba1\u7b97\u56fe\u652f\u6301<\/strong>\uff1aCUDA\u5bf9PyTorch\/TensorFlow\u7684\u52a8\u6001\u56fe\u4f18\u5316\u6210\u719f\uff0c\u800cMUSA\u53ef\u80fd\u56e0\u7f16\u8bd1\u5668\u9759\u6001\u4f18\u5316\u80fd\u529b\u4e0d\u8db3\uff0c\u5bfc\u81f4\u5b9e\u65f6\u63a8\u7406\u5ef6\u8fdf\u589e\u52a0\u3002<\/li>\n<li><strong>\u7b97\u5b50\u878d\u5408\u6548\u7387<\/strong>\uff1a\u82f1\u4f1f\u8fbe\u901a\u8fc7cuDNN\u5b9e\u73b0\u9ad8\u6548\u7b97\u5b50\u878d\u5408\uff0cMUSA\u9700\u81ea\u7814\u7c7b\u4f3c\u4f18\u5316\uff0c\u5426\u5219\u53ef\u80fd\u56e0\u591a\u6b21\u5185\u5b58\u8bfb\u5199\u589e\u52a0\u8017\u65f6\u3002<\/li>\n<\/ul>\n<h4><strong>\uff083\uff09\u56fe\u5f62\u6e32\u67d3\u4e0e\u6df7\u5408\u8d1f\u8f7d<\/strong><\/h4>\n<ul>\n<li><strong>\u56fe\u5f62\u4e0e\u8ba1\u7b97\u534f\u540c<\/strong>\uff1aCUDA\u5728\u5149\u7ebf\u8ffd\u8e2a\u3001\u7269\u7406\u4eff\u771f\u7b49\u6df7\u5408\u8d1f\u8f7d\u4e2d\u8c03\u5ea6\u4f18\u52bf\u660e\u663e\uff0cMUSA\u82e5\u786c\u4ef6\u8d44\u6e90\u5206\u914d\u7b56\u7565\u4e0d\u540c\uff0c\u53ef\u80fd\u5bfc\u81f4\u6548\u7387\u4e0b\u964d\u3002<\/li>\n<\/ul>\n<hr \/>\n<h3><strong>3. \u6469\u5c14\u7ebf\u7a0b\u7684\u4f18\u5316\u7b56\u7565<\/strong><\/h3>\n<h4><strong>\uff081\uff09\u786c\u4ef6\u5c42\u9762<\/strong><\/h4>\n<ul>\n<li><strong>\u5b9a\u5236\u5316\u8ba1\u7b97\u5355\u5143<\/strong>\uff1a\u9488\u5bf9\u5e38\u89c1CUDA\u64cd\u4f5c\uff08\u5982\u539f\u5b50\u64cd\u4f5c\u3001Warp\u8c03\u5ea6\uff09\u8bbe\u8ba1\u4e13\u7528\u786c\u4ef6\u903b\u8f91\uff0c\u51cf\u5c11\u6307\u4ee4\u8f6c\u6362\u5f00\u9500\u3002<\/li>\n<li><strong>\u5185\u5b58\u5b50\u7cfb\u7edf\u4f18\u5316<\/strong>\uff1a\u901a\u8fc7\u9ad8\u5e26\u5bbdHBM\u6216\u667a\u80fd\u7f13\u5b58\u5206\u533a\uff0c\u5f25\u8865\u517c\u5bb9\u6027\u5bfc\u81f4\u7684\u5185\u5b58\u8bbf\u95ee\u6548\u7387\u635f\u5931\u3002<\/li>\n<\/ul>\n<h4><strong>\uff082\uff09\u8f6f\u4ef6\u5c42\u9762<\/strong><\/h4>\n<ul>\n<li><strong>\u7f16\u8bd1\u5668\u6df1\u5ea6\u4f18\u5316<\/strong>\uff1aMUSA SDK\u901a\u8fc7\u9759\u6001\u7f16\u8bd1\u9884\u6d4b\u548c\u52a8\u6001\u8fd0\u884c\u65f6\u4f18\u5316\uff08\u5982JIT\u7f16\u8bd1\uff09\uff0c\u51cf\u5c11\u6307\u4ee4\u6620\u5c04\u5f00\u9500\u3002<\/li>\n<li><strong>\u751f\u6001\u5171\u5efa<\/strong>\uff1a\u4e0e\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\uff08PaddlePaddle\u3001MindSpore\uff09\u5408\u4f5c\uff0c\u76f4\u63a5\u96c6\u6210MUSA\u540e\u7aef\uff0c\u907f\u514d\u901a\u8fc7CUDA\u517c\u5bb9\u5c42\u4e2d\u8f6c\u3002<\/li>\n<\/ul>\n<hr \/>\n<h3><strong>4. \u5b9e\u6d4b\u6570\u636e\u4e0e\u884c\u4e1a\u53cd\u9988<\/strong><\/h3>\n<h4><strong>\uff081\uff09\u5b98\u65b9\u6027\u80fd\u5ba3\u79f0<\/strong><\/h4>\n<ul>\n<li>\u6469\u5c14\u7ebf\u7a0b\u5ba3\u79f0\u5176MTT S3000 GPU\u5728\u90e8\u5206AI\u63a8\u7406\u4efb\u52a1\uff08\u5982ResNet-50\uff09\u4e2d\u8fbe\u5230\u82f1\u4f1f\u8fbeA10\u7ea7\u6027\u80fd\uff0c\u4f46\u672a\u516c\u5f00\u5bf9\u6bd4CUDA\u539f\u751f\u73af\u5883\u4e0b\u7684\u6548\u7387\u635f\u8017\u7387\u3002<\/li>\n<li>\u5728\u56fe\u5f62\u6e32\u67d3\u573a\u666f\uff08\u5982Unity\u5f15\u64ce\uff09\uff0cMUSA\u9a71\u52a8\u4ecd\u5904\u4e8e\u529f\u80fd\u5b8c\u5584\u9636\u6bb5\uff0c\u5e27\u7387\u4e0e\u7a33\u5b9a\u6027\u8f83CUDA\u6709\u5dee\u8ddd\u3002<\/li>\n<\/ul>\n<h4><strong>\uff082\uff09\u7b2c\u4e09\u65b9\u6d4b\u8bd5<\/strong><\/h4>\n<ul>\n<li><strong>\u5b66\u672f\u673a\u6784\u6d4b\u8bd5<\/strong>\uff1a\u65e9\u671f\u5f00\u53d1\u8005\u53cd\u9988\uff0c\u79fb\u690dCUDA\u4ee3\u7801\u81f3MUSA\u65f6\uff0c\u9700\u624b\u52a8\u8c03\u6574\u5185\u5b58\u5206\u914d\u548c\u7ebf\u7a0b\u5757\u914d\u7f6e\uff0c\u4f18\u5316\u540e\u6027\u80fd\u53ef\u8fbeCUDA\u768470-85%\u3002<\/li>\n<li><strong>\u884c\u4e1a\u5e94\u7528\u6848\u4f8b<\/strong>\uff1a\u67d0\u56fd\u4ea7\u4e91\u8ba1\u7b97\u5e73\u53f0\u91c7\u7528MTT GPU\u90e8\u7f72NLP\u63a8\u7406\u670d\u52a1\uff0c\u5b9e\u6d4b\u541e\u5410\u91cf\u7ea6\u4e3a\u540c\u7b97\u529b\u82f1\u4f1f\u8fbeT4\u768465%\uff0c\u4f46\u6210\u672c\u4f4e30%\u3002<\/li>\n<\/ul>\n<hr \/>\n<h3><strong>5. \u6548\u7387\u635f\u5931\u7684\u6743\u8861\u4e0e\u957f\u671f\u5c55\u671b<\/strong><\/h3>\n<h4><strong>\uff081\uff09\u77ed\u671f\u52a3\u52bf<\/strong><\/h4>\n<ul>\n<li><strong>\u517c\u5bb9\u6027\u4ee3\u4ef7<\/strong>\uff1aMUSA\u4e3a\u517c\u5bb9CUDA\u751f\u6001\uff0c\u9700\u727a\u7272\u90e8\u5206\u786c\u4ef6\u8bbe\u8ba1\u81ea\u7531\u5ea6\uff0c\u5bfc\u81f4\u6548\u7387\u635f\u5931\uff1b  <\/li>\n<li><strong>\u751f\u6001\u77ed\u677f<\/strong>\uff1a\u7f3a\u4e4fCUDA\u7684\u6210\u719f\u5de5\u5177\u94fe\uff08\u5982Nsight\u3001NCCL\uff09\uff0c\u5f00\u53d1\u8005\u9700\u989d\u5916\u9002\u914d\u3002<\/li>\n<\/ul>\n<h4><strong>\uff082\uff09\u957f\u671f\u6f5c\u529b<\/strong><\/h4>\n<ul>\n<li><strong>\u56fd\u4ea7\u5316\u66ff\u4ee3\u7ea2\u5229<\/strong>\uff1a\u5728\u653f\u7b56\u9a71\u52a8\u4e0b\uff0cMUSA\u53ef\u901a\u8fc7\u5b9a\u5236\u5316\u4f18\u5316\uff08\u5982\u4e2d\u6587NLP\u3001\u653f\u52a1\u4e91\u573a\u666f\uff09\u63d0\u5347\u5b9e\u9645\u6548\u7387\uff1b  <\/li>\n<li><strong>\u5f02\u6784\u8ba1\u7b97\u6269\u5c55<\/strong>\uff1a\u82e5\u672a\u6765\u96c6\u6210\u4e13\u7528\u52a0\u901f\u5355\u5143\uff08\u5982LPU\u6a21\u5757\uff09\uff0c\u53ef\u80fd\u5b9e\u73b0\u201c\u901a\u7528GPU+\u9886\u57df\u52a0\u901f\u201d\u7684\u6df7\u5408\u67b6\u6784\uff0c\u90e8\u5206\u573a\u666f\u53cd\u8d85CUDA\u3002<\/li>\n<\/ul>\n<hr \/>\n<h3><strong>\u603b\u7ed3<\/strong><\/h3>\n<p><strong>MUSA\u67b6\u6784\u5728CUDA\u517c\u5bb9\u6027\u4e0a\u786e\u5b9e\u5b58\u5728\u6548\u7387\u635f\u5931<\/strong>\uff0c\u4e3b\u8981\u6e90\u4e8e\u786c\u4ef6\u5dee\u5f02\u3001\u8f6f\u4ef6\u751f\u6001\u6210\u719f\u5ea6\u53ca\u6307\u4ee4\u8f6c\u6362\u5f00\u9500\uff0c\u4f46\u5176\u8bbe\u8ba1\u76ee\u6807\u5e76\u975e\u5b8c\u5168\u66ff\u4ee3CUDA\uff0c\u800c\u662f\u5728\u56fd\u4ea7\u5316\u66ff\u4ee3\u4e0e\u81ea\u4e3b\u53ef\u63a7\u9700\u6c42\u4e2d\u5bfb\u627e\u5e73\u8861\u3002  <\/p>\n<ul>\n<li><strong>\u77ed\u671f\u5efa\u8bae<\/strong>\uff1a\u5728\u7b97\u529b\u9700\u6c42\u8f83\u4f4e\u6216\u653f\u7b56\u5bfc\u5411\u578b\u573a\u666f\uff08\u5982\u4fe1\u521b\u3001\u6559\u80b2\uff09\u53ef\u4f18\u5148\u91c7\u7528MUSA\uff0c\u501f\u52a9\u6210\u672c\u4f18\u52bf\u5f25\u8865\u6548\u7387\u635f\u5931\uff1b  <\/li>\n<li><strong>\u957f\u671f\u5c55\u671b<\/strong>\uff1a\u82e5\u6469\u5c14\u7ebf\u7a0b\u6301\u7eed\u4f18\u5316\u786c\u4ef6\u67b6\u6784\u5e76\u5b8c\u5584\u8f6f\u4ef6\u751f\u6001\uff0c\u90e8\u5206\u573a\u666f\u6548\u7387\u6709\u671b\u63a5\u8fd1\u751a\u81f3\u8d85\u8d8aCUDA\uff0c\u5c24\u5176\u5728\u56fd\u4ea7\u5b9a\u5236\u5316\u5e94\u7528\u4e2d\u3002  <\/li>\n<\/ul>\n<p>\u6700\u7ec8\u6548\u7387\u8868\u73b0\u9700\u7ed3\u5408\u5177\u4f53\u5de5\u4f5c\u8d1f\u8f7d\u548c\u4f18\u5316\u6295\u5165\u7efc\u5408\u8bc4\u4f30\uff0c\u5efa\u8bae\u901a\u8fc7\u5b9e\u9645\u4e1a\u52a1\u573a\u666f\u6d4b\u8bd5\u9a8c\u8bc1\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>MUSA\u67b6\u6784\u4e0eCUDA\u6548\u7387\u5bf9\u6bd4\u5206\u6790 \u6469\u5c14\u7ebf\u7a0b\u7684MUSA\uff08Moore Threads Unified Syste &hellip; <a href=\"https:\/\/crab.cool\/index.php\/2025\/02\/01\/musa%e6%9e%b6%e6%9e%84%e5%92%8ccuda%e7%9a%84%e5%85%b3%e7%b3%bb\/\" class=\"more-link\">\u7ee7\u7eed\u9605\u8bfb<span class=\"screen-reader-text\">MUSA\u67b6\u6784\u548cCUDA\u7684\u5173\u7cfb<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[4],"tags":[75,48],"class_list":["post-1498","post","type-post","status-publish","format-standard","hentry","category-world","tag-deepseek","tag-48"],"_links":{"self":[{"href":"https:\/\/crab.cool\/index.php\/wp-json\/wp\/v2\/posts\/1498","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/crab.cool\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/crab.cool\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/crab.cool\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/crab.cool\/index.php\/wp-json\/wp\/v2\/comments?post=1498"}],"version-history":[{"count":1,"href":"https:\/\/crab.cool\/index.php\/wp-json\/wp\/v2\/posts\/1498\/revisions"}],"predecessor-version":[{"id":1499,"href":"https:\/\/crab.cool\/index.php\/wp-json\/wp\/v2\/posts\/1498\/revisions\/1499"}],"wp:attachment":[{"href":"https:\/\/crab.cool\/index.php\/wp-json\/wp\/v2\/media?parent=1498"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/crab.cool\/index.php\/wp-json\/wp\/v2\/categories?post=1498"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/crab.cool\/index.php\/wp-json\/wp\/v2\/tags?post=1498"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}