{"id":72537,"date":"2025-12-18T17:03:36","date_gmt":"2025-12-18T09:03:36","guid":{"rendered":"https:\/\/aicats.wiki\/?p=72537"},"modified":"2025-12-18T17:03:44","modified_gmt":"2025-12-18T09:03:44","slug":"gemm-%e6%98%af%e4%bb%80%e9%ba%bc%ef%bc%9f-ai-%e7%94%a2%e6%a5%ad%e5%bf%85%e6%87%82%e7%9a%84%e9%ab%98%e6%95%88%e8%83%bd%e7%9f%a9%e9%99%a3%e4%b9%98%e6%b3%95%e5%8a%a0%e9%80%9f%e6%96%b9%e6%a1%88%e8%a9%b3","status":"publish","type":"post","link":"https:\/\/aicats.wiki\/tw\/2025\/12\/18\/72537-html","title":{"rendered":"gemm \u662f\u4ec0\u9ebc\uff1f AI \u7522\u696d\u5fc5\u61c2\u7684\u9ad8\u6548\u80fd\u77e9\u9663\u4e58\u6cd5\u52a0\u901f\u65b9\u6848\u8a73\u89e3"},"content":{"rendered":"<p><strong>GEMM<\/strong>\uff08\u901a\u7528\u77e9\u9663\u4e58\u6cd5\uff09\u662f<a class=\"external\" href=\"https:\/\/aicats.wiki\/tw\/tag\/ai\" title=\"\u67e5\u770b\u8207 AI \u76f8\u95dc\u7684\u6587\u7ae0\" target=\"_blank\">AI<\/a>\u3001\u79d1\u5b78\u8a08\u7b97\u9818\u57df\u7684<a href=\"https:\/\/aicats.wiki\/tw\/2025\/12\/05\/78889-html\/\" title=\"Tesla P40\u8a73\u89e3\uff1aAI\u958b\u767c\u8005\u5982\u4f55\u904b\u7528Tesla P40\u63d0\u5347\u6df1\u5ea6\u5b78\u7fd2\u8a13\u7df4\u6548\u7387\uff1f\">\u201c\u6027\u80fd\u5f15\u64ce\u201d<\/a>\uff0c<strong>\u6df1\u5ea6\u5b78\u7fd2\u8207\u5927\u6578\u64da\u6f14\u7b97\u6cd5\u7684\u7d55\u5927\u591a\u6578\u6838\u5fc3\u904b\u7b97\u90fd\u7e5e\u4e0d\u958bGEMM<\/strong>\u3002\u9ad8\u6548\u7387\u7684GEMM\u4e0d\u50c5\u652f\u6490\u8457\u6a21\u578b\u8a13\u7df4\u3001\u63a8\u7406\u3001\u5efa\u8b70\u7b49\u95dc\u9375\u4efb\u52d9\uff0c\u4e5f\u5c0dCPU\u3001GPU\u751a\u81f3AI\u6676\u7247\u7684\u67b6\u69cb\u8a2d\u8a08\u7522\u751f\u6df1\u9060\u5f71\u97ff\u3002\u672c\u6587\u8a73\u7d30\u6574\u7406\u4e86GEMM\u7684\u539f\u7406\u3001\u61c9\u7528\u3001\u4e3b\u6d41\u9ad8\u6548\u80fd\u5be6\u73fe\uff08OpenBLAS\u3001cuBLAS\u3001BLIS\u3001CUTLASS\u7b49\uff09\u53ca\u81ea\u52d5\u878d\u5408\u8abf\u5ea6\u5de5\u5177\uff0c\u8f14\u4ee5\u591a\u5e73\u53f0\u5be6\u6e2c\u3001\u512a\u5316\u8981\u9ede\u53ca\u958b\u767c\u8005FAQ\uff0c\u70baAI\u5de5\u7a0b\u5e2b\u3001\u6c7a\u7b56\u8005\u8207\u958b\u767c\u8005\u63d0\u4f9b\u4e00\u7ad9\u5f0f\u6027\u80fd\u9078\u578b\u6307\u5357\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img fetchpriority=\"high\" decoding=\"async\" width=\"1956\" height=\"1021\" src=\"https:\/\/aicats.wiki\/wp-content\/uploads\/2025\/12\/image-224.png\" alt=\"gemm \u662f\u4ec0\u9ebc\uff1f AI \u7522\u696d\u5fc5\u61c2\u7684\u9ad8\u6548\u80fd\u77e9\u9663\u4e58\u6cd5\u52a0\u901f\u65b9\u6848\u8a73\u89e3\" class=\"wp-image-84165\"\/><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>\u4ec0\u9ebc\u662fGEMM\uff1f\u70ba\u4ec0\u9ebc\u5b83\u662fAI\u548c\u79d1\u5b78\u8a08\u7b97\u7684\u57fa\u77f3<\/strong><\/h2>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>GEMM\u5b9a\u7fa9\u53ca\u5176\u57fa\u790e\u6578\u5b78\u5f62\u5f0f<\/strong><\/h3>\n\n\n\n<p>GEMM\uff08General Matrix-Matrix multiplication\uff0c\u901a\u7528\u77e9\u9663\u4e58\u6cd5\uff09\u662f\u7dda\u6027\u4ee3\u6578\u6700\u57fa\u790e\u3001\u6700\u5e38\u7528\u7684\u904b\u7b97\u4e4b\u4e00\uff0c\u5176\u901a\u7528\u5f62\u5f0f\u70ba\uff1a<br><code>C = \u03b1AB + \u03b2C<\/code><br>\u5176\u4e2d <strong>A\u3001B\u3001C<\/strong> \u5747\u70ba\u77e9\u9663\uff0c\u03b1\u3001\u03b2\u70ba\u6a19\u91cf\u3002\u5728\u591a\u6578AI\u5834\u666f\u4e0b\uff0c\u7c21\u5316\u70ba <code>C = AB<\/code>\u3002<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p><strong>GEMM<\/strong>\u5ee3\u6cdb\u61c9\u7528\u65bc\u6df1\u5ea6\u5b78\u7fd2\u5404\u985e\u7db2\u8def\uff08\u5982<code>Transformer<\/code>\u3001<code>CNN<\/code>\uff09\u3001\u5927\u898f\u6a21\u63a8\u85a6\u7cfb\u7d71\u3001\u79d1\u5b78\u6a21\u64ec\u3001\u7269\u7406\u5de5\u7a0b\u5efa\u6a21\u7b49\uff0c\u662fAI\u6f14\u7b97\u6cd5\u7684\u771f\u6b63\u7b97\u529b\u74f6\u9838\u3002<\/p>\n<\/blockquote>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u6642\u9593\u8907\u96dc\u5ea6<\/strong>\uff1a<code>O(MNK)<\/code><\/li>\n\n\n\n<li><strong>\u7a7a\u9593\u8907\u96dc\u5ea6<\/strong>\uff1a\u7d04\u7b49\u65bc<code>O(MK + KN + MN)<\/code><\/li>\n<\/ul>\n\n\n\n<p>\u7576\u77e9\u9663\u7dad\u5ea6\u57281024\u53ca\u4ee5\u4e0a\u6642\uff0c\u904b\u7b97\u898f\u6a21\u53ef\u9054<strong>\u5341\u5104\u7d1a\u6d6e\u9ede\u64cd\u4f5c<\/strong>\uff0c\u6545AI\u6676\u7247\u3001\u4f3a\u670d\u5668\u90fd\u5c08\u9580\u570d\u7e5eGEMM\u6253\u9020\u52a0\u901f\u55ae\u5143\u3002<\/p>\n\n\n\n<p>\u63a8\u85a6\u95b1\u8b80\uff1a<a href=\"https:\/\/zhuanlan.zhihu.com\/p\/1896699250961797686\" target=\"_blank\"  rel=\"nofollow noopener\"  class=\"external\" >\u8a73\u76e1\u6f14\u7b97\u6cd5\u5256\u6790-GEMM\u6700\u4f73\u5316\u539f\u7406<\/a><\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>GEMM\u5728AI\u7522\u696d\u7684\u6838\u5fc3\u5730\u4f4d<\/strong><\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u6df1\u5ea6\u5b78\u7fd2\u6846\u67b6\uff08PyTorch\u3001TensorFlow\uff09\u5e95\u5c64\u77e9\u9663\u64cd\u4f5c\u57fa\u672c\u4e0a\u90fd\u6703\u547c\u53eb\u9ad8\u6548\u80fdGEMM<\/li>\n\n\n\n<li><strong>NLP\u3001CV\u6a21\u578b\u6bcf\u5929\u6210\u5343\u4e0a\u842c\u6b21\u5927\u77e9\u9663\u904b\u7b97<\/strong><\/li>\n\n\n\n<li>AI\u6676\u7247\u3001FPGA\u3001\u8d85\u7b97\u90fd\u5c08\u9580\u8a2d\u8a08GEMM\u786c\u9ad4\u52a0\u901f\u6a21\u7d44<\/li>\n\n\n\n<li>GEMM\u8868\u73fe\u597d\u58de\uff0c\u76f4\u63a5\u6c7a\u5b9a\u6a21\u578b\u8a13\u7df4\u8207\u63a8\u7406\u6548\u7387<\/li>\n<\/ul>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>\u4e3b\u6d41\u9ad8\u6548\u80fdGEMM\u65b9\u6848\u8207\u5de5\u5177\u5927\u76e4\u9ede<\/strong><\/h2>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><tbody><tr><th>\u540d\u7a31<\/th><th>\u985e\u578b<\/th><th>\u9069\u7528\u786c\u9ad4<\/th><th>\u6700\u4f73\u5316\u7279\u9ede<\/th><th>\u5b98\u7db2\/\u9023\u7d50<\/th><\/tr><tr><td><strong>OpenBLAS<\/strong><\/td><td>\u958b\u6e90\u51fd\u5f0f\u5eab<\/td><td>CPU\/X86\/ARM<\/td><td>\u591a\u7dda\u7a0b\/\u77e9\u9663\u5206\u584a<\/td><td><a href=\"https:\/\/www.openblas.net\/\" target=\"_blank\"  rel=\"nofollow noopener\"  class=\"external\" >OpenBLAS<\/a><\/td><\/tr><tr><td><strong>Intel MKL<\/strong><\/td><td>\u5546\u7528\u5eab<\/td><td>Intel CPU\/X86<\/td><td>\u77e2\u91cf\u5316\/SIMD<\/td><td><a href=\"https:\/\/www.intel.com\/content\/www\/us\/en\/developer\/tools\/oneapi\/onemkl-download.html\" target=\"_blank\"  rel=\"nofollow noopener\"  class=\"external\" >oneAPI MKL<\/a><\/td><\/tr><tr><td><strong>cuBLAS<\/strong><\/td><td>\u5546\u7528\u5eab<\/td><td>NVIDIA GPU<\/td><td>Tensor Core\/FMA<\/td><td><a href=\"https:\/\/developer.nvidia.com\/cublas\" target=\"_blank\"  rel=\"nofollow noopener\"  class=\"external\" >cuBLAS<\/a><\/td><\/tr><tr><td><strong>CUTLASS<\/strong><\/td><td>\u958b\u6e90\u5de5\u5177\u5eab<\/td><td>NVIDIA GPU<\/td><td>\u9ad8\u5ea6\u53ef\u914d\u7f6e\/AI\u512a\u5316<\/td><td><a href=\"https:\/\/github.com\/NVIDIA\/cutlass\" target=\"_blank\"  rel=\"nofollow noopener\"  class=\"external\" >CUTLASS<\/a><\/td><\/tr><tr><td><strong>BLIS<\/strong><\/td><td>\u958b\u6e90\u51fd\u5f0f\u5eab<\/td><td>CPU\/X86\/ARM<\/td><td>\u591a\u5c64\u5206\u584a\/\u53ef\u64f4\u5c55<\/td><td><a href=\"https:\/\/github.com\/flame\/blis\" target=\"_blank\"  rel=\"nofollow noopener\"  class=\"external\" >BLIS<\/a><\/td><\/tr><tr><td><strong>Eigen<\/strong><\/td><td>C++\u51fd\u5f0f\u5eab<\/td><td>\u8de8\u5e73\u53f0CPU<\/td><td>\u6a21\u677f\u5316\/\u81ea\u52d5\u5411\u91cf\u5316<\/td><td><a href=\"https:\/\/eigen.tuxfamily.org\/\" target=\"_blank\"  rel=\"nofollow noopener\"  class=\"external\" >Eigen<\/a><\/td><\/tr><tr><td><strong>cblas\/sgemm\/dgemm<\/strong><\/td><td>BLAS\u6a19\u6e96<\/td><td>\u901a\u7528CPU<\/td><td>\u6a19\u6e96API<\/td><td><a href=\"https:\/\/www.netlib.org\/blas\/\" target=\"_blank\"  rel=\"nofollow noopener\"  class=\"external\" >BLAS\u4ecb\u7d39<\/a><\/td><\/tr><tr><td><strong>TVM\/ONNX Runtime\/TensorRT<\/strong><\/td><td>\u6846\u67b6<\/td><td>CPU+GPU+AI Chip<\/td><td>\u8abf\u5ea6\u81ea\u52d5\u641c\u5c0b\u878d\u5408<\/td><td><a href=\"https:\/\/tvm.apache.org\/\" target=\"_blank\"  rel=\"nofollow noopener\"  class=\"external\" >TVM<\/a><\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/aicats.wiki\/wp-content\/uploads\/2025\/11\/my_prefix_1763797218.jpg\" alt=\"OpenBLAS\u5b98\u65b9\u7db2\u7ad9\u622a\u5716\" class=\"wp-image-51824\"\/><figcaption class=\"wp-element-caption\">\u5716\uff0f<a href=\"http:\/\/www.openmathlib.org\/OpenBLAS\/\" title=\"\" target=\"_blank\"  rel=\"nofollow noopener\"  class=\"external\" >OpenBLAS\u5b98\u65b9\u7db2\u7ad9\u622a\u5716<\/a><\/figcaption><\/figure>\n\n\n\n<p><strong>\u8a3b\uff1a<\/strong> \u76ee\u524d\u4e3b\u6d41AI\u6846\u67b6\u7684\u5e95\u5c64GEMM\u90fd\u9ad8\u5ea6\u4f9d\u8cf4\u4e0a\u8ff0\u51fd\u5f0f\u5eab\u3002 PyTorch\u3001TensorFlow\u901a\u5e38\u505a\u4e86\u4e8c\u6b21\u5c01\u88dd\uff0c\u6bcf\u7576\u51fd\u5f0f\u5eab\u6548\u80fd\u5347\u7d1a\uff0cAI\u6a21\u578b\u76f4\u63a5\u53d7\u76ca\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>CPU\u5e73\u53f0\u4e0b\u7684\u9ad8\u6027\u80fdGEMM\u5be6\u73fe\u8207\u6700\u4f73\u5316<\/strong><\/h2>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>BLAS\u6a19\u6e96API\u8207\u5e38\u7528\u4ecb\u9762<\/strong><\/h3>\n\n\n\n<!--wp-compress-html--><!--wp-compress-html no compression-->\n<pre class=\"wp-block-code\"><code>\/\/ sgemm\/dgemm\u7d93\u5178\u547c\u53eb\u7bc4\u4f8b(C\/C++) void cblas_sgemm(const CBLAS_LAYOUT Layout, const CBLAS_TRANSPOSE TransA, const CBLAS_TRANSPOSE TransB, const int M, const int N, const int K, const flost. ldb, const float beta, float *C, const int ldc);<\/code><\/pre>\n<!--wp-compress-html no compression--><!--wp-compress-html-->\n\n\n\n<ul class=\"wp-block-list\">\n<li><code>sgemm<\/code> : \u55ae\u7cbe\u5ea6(float)<\/li>\n\n\n\n<li><code>dgemm<\/code> : \u96d9\u7cbe\u5ea6(double)<\/li>\n<\/ul>\n\n\n\n<p><strong>\u6240\u6709\u4e3b\u6d41AI\u8207\u79d1\u5b78\u8a08\u7b97\u8edf\u9ad4\u5747\u76f8\u5bb9\u65bc\u6b64\u4ecb\u9762\u3002<\/strong><\/p>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>\u4e3b\u6d41CPU GEMM\u51fd\u5f0f\u5eab\u6a6b\u5411\u5c0d\u6bd4<\/strong><\/h3>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><tbody><tr><th>\u5e93<\/th><th>\u512a\u5316\u624b\u6bb5<\/th><th>\u591a\u57f7\u884c\u7dd2<\/th><th>\u5411\u91cf\u6307\u4ee4<\/th><th>\u5099\u8a3b<\/th><\/tr><tr><td><strong>OpenBLAS<\/strong><\/td><td>Cache\u5206\u584a+SIMD+\u7dda\u7a0b<\/td><td>\u221a<\/td><td>AVX512\/AVX2\/NEON<\/td><td>\u793e\u5340\u6d3b\u8e8d<\/td><\/tr><tr><td><strong>Intel MKL<\/strong><\/td><td>AVX512\u624b\u5de5\u512a\u5316<\/td><td>\u221a<\/td><td>AVX512<\/td><td>Intel\u5e73\u53f0\u6700\u512a<\/td><\/tr><tr><td><strong>BLIS<\/strong><\/td><td>\u53ef\u63d2\u62d4\/\u591a\u6838<\/td><td>\u221a<\/td><td>AVX\/NEON<\/td><td>AMD\/ARM\u4e5f\u9ad8\u6548<\/td><\/tr><tr><td><strong>MathNet(.NET)<\/strong><\/td><td>\u8a17\u7ba1\u7a0b\u5f0f\u78bc+\u57fa\u672cSIMD<\/td><td>\u6709\u9650<\/td><td>.NET SIMD<\/td><td>\u8de8\u5e73\u53f0\uff0c\u901f\u5ea6\u4e2d\u7b49<\/td><\/tr><tr><td><strong>Eigen<\/strong><\/td><td>\u6a21\u677f\u81ea\u52d5\u5411\u91cf\u5316<\/td><td>\u00d7<\/td><td>\u652f\u6301\u81ea\u52d5SIMD<\/td><td>C++\u76f8\u5bb9\u512a\u5148\uff0c\u6975\u5927\u898f\u6a21\u7565\u6162<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>\u5be6\u6e2c\u57fa\u6e96\u8cc7\u6599\uff081024\u00d71024 SGEMM\uff0c\u55ae\u7bc0\u9ede\uff09<\/strong><\/h3>\n\n\n\n<figure class=\"wp-block-image size-full\"><img decoding=\"async\" width=\"1948\" height=\"1023\" src=\"https:\/\/aicats.wiki\/wp-content\/uploads\/2025\/12\/image-223.png\" alt=\"C# .NET\u985e\u5225\u5eab\u9ad8\u6548\u80fd\u5c0d\u6bd4\" class=\"wp-image-84160\"\/><figcaption class=\"wp-element-caption\">\u5716\uff0f<a href=\"https:\/\/www.cnblogs.com\/zyl910\/p\/19067040\/MatrixBenchmarkCs_gemm_simd\" target=\"_blank\"  rel=\"nofollow noopener\"  class=\"external\" >C# .NET\u985e\u5225\u5eab\u9ad8\u6548\u80fd\u5c0d\u6bd4<\/a><\/figcaption><\/figure>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><tbody><tr><th>\u5be6\u73fe<\/th><th>\u8017\u6642(ns)<\/th><th>\u76f8\u5c0d\u6027\u80fd<\/th><th>GFLOPS<\/th><\/tr><tr><td>\u4e09\u91cd\u5faa\u74b0C++<\/td><td>4,712,905,103<\/td><td>1x<\/td><td>0.42<\/td><\/tr><tr><td>OpenBLAS<\/td><td>2,932,070<\/td><td>1607x<\/td><td>682<\/td><\/tr><tr><td>Intel MKL<\/td><td>4,379,927<\/td><td>1076x<\/td><td>456<\/td><\/tr><tr><td>MathNet<\/td><td>53,205,723<\/td><td>88x<\/td><td>37.5<\/td><\/tr><tr><td>SIMD+\u5206\u584a\u4e26\u884c(C#)<\/td><td>4,363,112<\/td><td>1080x<\/td><td>458<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p>\u66f4\u591a\u8cc7\u6599\u4f86\u6e90\uff1a<a href=\"https:\/\/www.cnblogs.com\/zyl910\/p\/19067040\/MatrixBenchmarkCs_gemm_simd\" target=\"_blank\"  rel=\"nofollow noopener\"  class=\"external\" >C# .NET\u985e\u5225\u5eab\u9ad8\u6548\u80fd\u5c0d\u6bd4<\/a><\/p>\n\n\n\n<h4 class=\"wp-block-heading\"><strong>CPU\u512a\u5316\u4e3b\u6d41\u6280\u5de7<\/strong><\/h4>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5faa\u74b0\u5c55\u958b\u3001reorder\uff0c\u63d0\u5347Cache\/VPU\u5c40\u90e8\u6027<\/li>\n\n\n\n<li>Cache\u5206\u584a\uff08L2\/L3\uff09\uff0c\u964d\u4f4e\u8a18\u61b6\u9ad4\u983b\u5bec\u58d3\u529b<\/li>\n\n\n\n<li>SIMD\u6307\u4ee4\u512a\u5316\uff08SSE\/AVX\/NEON\uff09<\/li>\n\n\n\n<li>\u591a\u6838\u5fc3\u4e26\u884c\uff08OpenMP\/PThread\uff09<\/li>\n\n\n\n<li>\u878d\u5408\u4e58\u52a0FMA\u6307\u4ee4<\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-image size-full\"><img decoding=\"async\" width=\"1947\" height=\"1023\" src=\"https:\/\/aicats.wiki\/wp-content\/uploads\/2025\/12\/image-225.png\" alt=\"MathNet.Numerics\" class=\"wp-image-84168\"\/><figcaption class=\"wp-element-caption\">\u5716\uff0f<a href=\"https:\/\/www.nuget.org\/packages\/MathNet.Numerics\/\" target=\"_blank\"  rel=\"nofollow noopener\"  class=\"external\" >MathNet.Numerics<\/a><\/figcaption><\/figure>\n\n\n\n<h4 class=\"wp-block-heading\"><strong>\u4e3b\u6d41CPU GEMM\u5eab\u4f4d\u5740\u6e05\u55ae\uff1a<\/strong><\/h4>\n\n\n\n<ul class=\"wp-block-list\">\n<li><a href=\"https:\/\/github.com\/xianyi\/OpenBLAS\" target=\"_blank\"  rel=\"nofollow noopener\"  class=\"external\" >OpenBLAS<\/a><\/li>\n\n\n\n<li><a href=\"https:\/\/www.intel.com\/content\/www\/us\/en\/developer\/tools\/oneapi\/onemkl-download.html\" target=\"_blank\"  rel=\"nofollow noopener\"  class=\"external\" >Intel MKL<\/a><\/li>\n\n\n\n<li><a href=\"https:\/\/github.com\/flame\/blis\" target=\"_blank\"  rel=\"nofollow noopener\"  class=\"external\" >BLIS<\/a><\/li>\n\n\n\n<li><a href=\"https:\/\/gitlab.com\/libeigen\/eigen\" target=\"_blank\"  rel=\"nofollow noopener\"  class=\"external\" >Eigen<\/a><\/li>\n\n\n\n<li><a href=\"https:\/\/www.nuget.org\/packages\/MathNet.Numerics\/\" target=\"_blank\"  rel=\"nofollow noopener\"  class=\"external\" >MathNet.Numerics<\/a><\/li>\n<\/ul>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>GPU\u8207AI\u6676\u7247\u4e0a\u7684\u9ad8\u6548\u80fdGEMM<\/strong><\/h2>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>NVIDIA\/AMD GPU\u4e3b\u6d41\u51fd\u5f0f\u5eab<\/strong><\/h3>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><tbody><tr><th>\u5eab\u540d<\/th><th>\u5e73\u53f0<\/th><th>\u6700\u4f73\u5316<\/th><th>\u7528\u9014<\/th><th>\u5b98\u7db2<\/th><\/tr><tr><td>cuBLAS<\/td><td>NVIDIA<\/td><td>Tensor Core\/FMA<\/td><td>\u8a13\u7df4\/\u63a8\u7406<\/td><td><a href=\"https:\/\/developer.nvidia.com\/cublas\" target=\"_blank\"  rel=\"nofollow noopener\"  class=\"external\" >cuBLAS<\/a><\/td><\/tr><tr><td>CUTLASS<\/td><td>NVIDIA<\/td><td>\u6a21\u677f\u5316\u5b9a\u5236<\/td><td>\u81ea\u7814Kernels<\/td><td><a href=\"https:\/\/github.com\/NVIDIA\/cutlass\" target=\"_blank\"  rel=\"nofollow noopener\"  class=\"external\" >CUTLASS<\/a><\/td><\/tr><tr><td>ROCm BLAS<\/td><td>AMD<\/td><td>\u5e73\u53f0\u91dd\u5c0d<\/td><td>AMD\u8a13\u7df4\/\u63a8\u7406<\/td><td><a href=\"https:\/\/github.com\/ROCmSoftwarePlatform\/rocBLAS\" target=\"_blank\"  rel=\"nofollow noopener\"  class=\"external\" >ROCm BLAS<\/a><\/td><\/tr><tr><td>TensorRT\/ONNX<\/td><td>\u591a\u5e73\u53f0<\/td><td>\u81ea\u52d5\u878d\u5408\/INT8<\/td><td>\u63a8\u7406\u5f15\u64ce<\/td><td><a href=\"https:\/\/developer.nvidia.com\/tensorrt\" target=\"_blank\"  rel=\"nofollow noopener\"  class=\"external\" >TensorRT<\/a><\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>cuBLAS\/ROCm<\/strong>\u70ba\u4e3b\u6d41\u8a13\u7df4\u5f15\u64ce\uff0c\u652f\u63f4FP32\u3001FP16\u3001BF16\u6df7\u5408\u7cbe\u6e96\u5ea6GEMM\u3002<\/li>\n\n\n\n<li>NVIDIA Tensor Core\u3001AI\u6676\u7247\u5747\u6709\u5c08\u7528GEMM\u96fb\u8def\uff0c\u6975\u81f4\u6548\u80fd\u3002<\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/aicats.wiki\/wp-content\/uploads\/2025\/11\/my_prefix_1763797225.jpg\" alt=\"cuBLAS\u5b98\u7db2\u4ecb\u9762\" class=\"wp-image-51824\"\/><figcaption class=\"wp-element-caption\">\u5716\uff0fcuBLAS\u5b98\u7db2\u4ecb\u9762<\/figcaption><\/figure>\n\n\n\n<p>\u63a8\u85a6\u53c3\u8003\uff1a<a href=\"https:\/\/docs.nvidia.com\/cuda\/cublas\/index.html\" target=\"_blank\"  rel=\"nofollow noopener\"  class=\"external\" >NVIDIA cuBLAS\u6587\u6a94<\/a> | <a href=\"https:\/\/cutlass.github.io\/\" target=\"_blank\"  rel=\"nofollow noopener\"  class=\"external\" >CUTLASS\u6559\u7a0b<\/a><\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>AI \u6846\u67b6\u8207\u81ea\u52d5GEMM\u878d\u5408\u8abf\u5ea6<\/strong><\/h2>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>\u6846\u67b6\u539f\u751fGEMM\u652f\u63f4\u5217\u8868<\/strong><\/h3>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><tbody><tr><th>\u6846\u67b6<\/th><th>\u4f9d\u8cf4<\/th><th>\u81ea\u52d5\u878d\u5408<\/th><th>\u786c\u9ad4<\/th><\/tr><tr><td>PyTorch<\/td><td>MKL\/OpenBLAS\/cuBLAS<\/td><td>\u52d5\u614b\u8abf\u5ea6<\/td><td>CPU\/GPU\/AI\u6676\u7247<\/td><\/tr><tr><td>TensorFlow<\/td><td>MKL\/OpenBLAS\/cuBLAS<\/td><td>\u81ea\u52d5\u878d\u5408<\/td><td>CPU\/GPU\/TPU<\/td><\/tr><tr><td>ONNX Runtime<\/td><td>\u591a\u5e73\u53f0<\/td><td>TVM\u8abf\u5ea6<\/td><td>\u591a\u7a2e\u786c\u9ad4<\/td><\/tr><tr><td>PaddlePaddle<\/td><td>MKL\/\u81ea\u7814<\/td><td>CUDA\/AI\u6676\u7247\u652f\u63f4<\/td><td>\u591a\u5e73\u53f0<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>TensorFlow\/Torch \u8f49ONNX \u7d93TensorRT\/ONNX Runtime\u53ef\u81ea\u52d5\u9078\u53d6\u6700\u4f73GEMM\u5be6\u73fe<\/strong>\uff0c\u958b\u767c\u8005\u7121\u9700\u95dc\u6ce8\u5e95\u5c64\u7d30\u7bc0\u3002<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>\u4e3b\u6d41\u81ea\u52d5\u878d\u5408\u5de5\u5177<\/strong><\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li><a href=\"https:\/\/tvm.apache.org\/\" target=\"_blank\"  rel=\"nofollow noopener\"  class=\"external\" >TVM<\/a>\uff1a\u81ea\u52d5Kernel\u641c\u5c0b\u6700\u512aGEMM<\/li>\n\n\n\n<li><a href=\"https:\/\/github.com\/oneapi-src\/onednn\" target=\"_blank\"  rel=\"nofollow noopener\"  class=\"external\" >OneDNN<\/a>\uff1aIntel\u81ea\u52d5\u6700\u4f73\u5316\u878d\u5408<\/li>\n\n\n\n<li><a href=\"https:\/\/pytorch.org\/blog\/torchdynamo\/\" target=\"_blank\"  rel=\"nofollow noopener\"  class=\"external\" >TorchDynamo<\/a>\uff1aPytorch\u52d5\u614bKernel<\/li>\n\n\n\n<li><a href=\"https:\/\/developer.nvidia.com\/tensorrt\" target=\"_blank\"  rel=\"nofollow noopener\"  class=\"external\" >TensorRT<\/a>\uff1aNVIDIA\u63a8\u7406\u5168\u81ea\u52d5\u512a\u5316<\/li>\n<\/ul>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>GEMM\u6548\u80fd\u512a\u5316\u5de5\u7a0b\u5be6\u4f8b<\/strong><\/h2>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>\u4ea4\u63db\u5faa\u74b0\u9806\u5e8f\u8207\u5206\u584a(C++\u5be6\u8e10)<\/strong><\/h3>\n\n\n\n<!--wp-compress-html--><!--wp-compress-html no compression-->\n<pre class=\"wp-block-code\"><code>\/\/ \u6a38\u7d20\u7248for (int i = 0; i &lt; N; ++i) for (int j = 0; j &lt; N; ++j) for (int k = 0; k &lt; N; ++k) C[i][j] += A[i][k]*B[k][j]; \/\/ \u91cd\u6392\u6700\u4f73; float a = A[i][k]; for (int j = 0; j &lt; N; ++j) C[i][j] += a * B[k][j]; }<\/code><\/pre>\n<!--wp-compress-html no compression--><!--wp-compress-html-->\n\n\n\n<p><strong>\u91cd\u6392\u5f8cB\/C\u7684\u5b58\u53d6\u5c40\u90e8\u6027\u63d0\u5347\uff0cCache\u53caSIMD\u52a0\u901f\u5e45\u5ea6\u660e\u986f\u3002<\/strong><\/p>\n\n\n\n<h4 class=\"wp-block-heading\"><strong>\u5206\u584a\u7b56\u7565\u8207\u5e73\u53f0\u9069\u7528\u6027\uff08\u793a\u610f\u8868\uff09<\/strong><\/h4>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><tbody><tr><th>\u5206\u584a\u7b56\u7565<\/th><th>\u512a\u9ede<\/th><th>\u9069\u7528\u5e73\u53f0<\/th><\/tr><tr><td>L2\/L3\u5206\u584a<\/td><td>\u7de9\u5b58\u9ad8\u547d\u4e2d<\/td><td>CPU<\/td><\/tr><tr><td>Tile+SIMD<\/td><td>\u5c0d\u9f4a\u5411\u91cf\uff0c\u6d41\u6c34\u7dda<\/td><td>CPU\/GPU<\/td><\/tr><tr><td>Thread\u5206\u584a<\/td><td>\u5e73\u884c<\/td><td>\u591a\u6838\u5fc3CPU\/GPU<\/td><\/tr><tr><td>TensorCore\u7279\u5b9a\u5206\u584a<\/td><td>AI\u6676\u7247\u6975\u81f4\u6027\u80fd<\/td><td>NVIDIA GPU\/AI\u6676\u7247<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<h4 class=\"wp-block-heading\"><strong>FPGA\u3001AI\u6676\u7247GEMM\u52a0\u901f\u65b9\u6848\u7c21\u8981<\/strong><\/h4>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5927\u898f\u6a21GEMM\u9663\u5217\uff08Mult\/Accumulate\u55ae\u5143\uff09<\/li>\n\n\n\n<li>\u652f\u63f4\u53ef\u8b8a\u7cbe\u5ea6\uff08INT8\/BF16\/F16\uff09\uff0c\u7b97\u529b\u6210\u767e\u4e0a\u5343TFLOPS<\/li>\n\n\n\n<li>\u570b\u7522\u6607\u9a30\u7b49AI\u6676\u7247\u6709\u5c08\u7528GEMM\u6838<\/li>\n<\/ul>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>\u958b\u767c\u8005FAQ\u8207\u9078\u578b\u5efa\u8b70<\/strong><\/h2>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>\u5982\u4f55\u9078\u4e0d\u540c\u5e73\u53f0GEMM\u51fd\u5f0f\u5eab\uff1f<\/strong><\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>CPU\uff08x86\/AMD\uff09<\/strong>\uff1a\u9996\u9078OpenBLAS\u3001BLIS\uff0cIntel\u5e73\u53f0\u53ef\u9078MKL<\/li>\n\n\n\n<li><strong>GPU<\/strong>\uff1aNVIDIA\u7528cuBLAS\uff0cAMD\u7528rocmBLAS<\/li>\n\n\n\n<li><strong>\u5d4c\u5165\u5f0f\/ARM<\/strong>\uff1a\u5efa\u8b70\u7528Eigen\u7b49\u8f15\u91cf\u7d1a\u5be6\u73fe<\/li>\n\n\n\n<li><strong>.NET\/Java<\/strong>\uff1a\u512a\u5148\u5916\u90e8\u547c\u53ebBLAS\u51fd\u5f0f\u5eab<\/li>\n\n\n\n<li><strong>AI\u6676\u7247\/FPGA<\/strong>\uff1a\u7528\u5b98\u65b9SDK\u63a8\u85a6\u65b9\u6848<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>\u6027\u80fd\u662f\u5426\u9054\u6a19\uff1f<\/strong><\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li>1024\u00d71024\u5927\u77e9\u9663\u55ae\u7dda\u7a0b&gt;20 GFLOPS\uff0c\u591a\u7dda\u7a0b\u6216GPU&gt;1000 GFLOPS<\/li>\n\n\n\n<li>\u5584\u7528perf\/nvprof\/ncu\/VTune\u5206\u6790\u74f6\u9838<\/li>\n\n\n\n<li>\u8cc7\u6599\u6392\u5217\u683c\u5f0f\u5f71\u97ff\u6975\u5927\uff0c\u5efa\u8b70\u5339\u914d\u5eab\u9810\u8a2d\u683c\u5f0f<\/li>\n<\/ul>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>\u7d50\u8a9e<\/strong><\/h2>\n\n\n\n<p><strong>GEMM\u662fAI\u6280\u8853\u5806\u758a\u7684\u52d5\u529b\u4f86\u6e90<\/strong>\uff0c\u5f71\u97ff\u6df1\u9060\u3002<strong>\u9ad8\u6548\u80fdGEMM\u65b9\u6848\u5982OpenBLAS\u3001cuBLAS\u3001BLIS\u7b49\u4e0d\u65b7\u5275\u65b0\u9032\u5316\uff0c\u65b0\u4e00\u4ee3AI\u6676\u7247\u4e5f\u5728\u9019\u6a21\u7d44\u6df1\u8015\u52a0\u901f\uff0c\u5f15\u9818AI\u7522\u696d\u767c\u5c55\u65b0\u968e\u68af\u3002<\/strong> <br>\u4e86\u89e3\u4e26\u6d3b\u7528\u6700\u65b0GEMM\u512a\u5316\u65b9\u6848\uff0c\u4e0d\u50c5\u4fdd\u969c\u7cfb\u7d71\u6548\u80fd\uff0c\u9084\u80fd\u5728AI\u696d\u52d9\u5275\u65b0\u4e0a\u8d0f\u5f97\u5148\u6a5f\u3002<\/p>\n\n\n\n<p><em>\u9032\u4e00\u6b65\u6280\u8853\u53c3\u8003\uff1a<a href=\"https:\/\/www.netlib.org\/blas\/\" target=\"_blank\"  rel=\"nofollow noopener\"  class=\"external\" >BLAS\u5b98\u65b9<\/a> | <a href=\"https:\/\/www.intel.com\/content\/www\/us\/en\/developer\/tools\/oneapi\/onemkl-download.html\" target=\"_blank\"  rel=\"nofollow noopener\"  class=\"external\" >oneMKL<\/a> | <a href=\"https:\/\/developer.nvidia.com\/cublas\" target=\"_blank\"  rel=\"nofollow noopener\"  class=\"external\" >cuBLAS<\/a> | <a href=\"https:\/\/www.openblas.net\/\" target=\"_blank\"  rel=\"nofollow noopener\"  class=\"external\" >OpenBLAS<\/a> <\/em><\/p>","protected":false},"excerpt":{"rendered":"<p>GEMM\uff08\u901a\u7528\u77e9\u9635\u4e58\u6cd5\uff09\u662fAI\u3001\u79d1\u5b66\u8ba1\u7b97\u9886\u57df\u7684\u201c\u6027\u80fd\u5f15\u64ce\u201d\uff0c\u6df1\u5ea6\u5b66\u4e60\u4e0e\u5927\u6570\u636e\u7b97\u6cd5\u7684\u7edd\u5927\u591a\u6570\u6838\u5fc3\u8fd0\u7b97\u90fd\u7ed5\u4e0d\u5f00GE [&hellip;]<\/p>\n","protected":false},"author":3,"featured_media":0,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_crsspst_to_aicatswiki":true,"_monsterinsights_skip_tracking":false,"_monsterinsights_sitenote_active":false,"_monsterinsights_sitenote_note":"","_monsterinsights_sitenote_category":0,"footnotes":""},"categories":[305],"tags":[247,352,284],"content_visibility":[262],"class_list":["post-72537","post","type-post","status-publish","format-standard","hentry","category-ai-tools-platforms","tag-ai"],"aioseo_notices":[],"_links":{"self":[{"href":"https:\/\/aicats.wiki\/tw\/wp-json\/wp\/v2\/posts\/72537","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/aicats.wiki\/tw\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/aicats.wiki\/tw\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/aicats.wiki\/tw\/wp-json\/wp\/v2\/users\/3"}],"replies":[{"embeddable":true,"href":"https:\/\/aicats.wiki\/tw\/wp-json\/wp\/v2\/comments?post=72537"}],"version-history":[{"count":1,"href":"https:\/\/aicats.wiki\/tw\/wp-json\/wp\/v2\/posts\/72537\/revisions"}],"predecessor-version":[{"id":84173,"href":"https:\/\/aicats.wiki\/tw\/wp-json\/wp\/v2\/posts\/72537\/revisions\/84173"}],"wp:attachment":[{"href":"https:\/\/aicats.wiki\/tw\/wp-json\/wp\/v2\/media?parent=72537"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/aicats.wiki\/tw\/wp-json\/wp\/v2\/categories?post=72537"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/aicats.wiki\/tw\/wp-json\/wp\/v2\/tags?post=72537"},{"taxonomy":"content_visibility","embeddable":true,"href":"https:\/\/aicats.wiki\/tw\/wp-json\/wp\/v2\/content_visibility?post=72537"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}