{"id":8778,"date":"2026-06-09T14:31:40","date_gmt":"2026-06-09T06:31:40","guid":{"rendered":"https:\/\/ab123.xyz\/?p=8778"},"modified":"2026-06-09T14:31:40","modified_gmt":"2026-06-09T06:31:40","slug":"intel-oneapi-deep-neural-network-library-for-gaudi-3-%e6%b7%b1%e5%ba%a6%e8%a7%a3%e6%9e%90","status":"publish","type":"post","link":"https:\/\/ab123.xyz\/?p=8778","title":{"rendered":"Intel oneAPI Deep Neural Network Library for Gaudi 3 \u6df1\u5ea6\u89e3\u6790"},"content":{"rendered":"<p>Intel oneAPI Deep Neural Network Library\uff08\u7b80\u79f0 oneDNN\uff09\u662f\u9488\u5bf9\u6df1\u5ea6\u5b66\u4e60\u5de5\u4f5c\u8d1f\u8f7d\u7cbe\u5fc3\u4f18\u5316\u7684\u5f00\u6e90\u6027\u80fd\u5e93\u3002\u9488\u5bf9 Intel Gaudi 3 AI \u52a0\u901f\u5668\uff0coneDNN \u63d0\u4f9b\u4e86\u5e95\u5c42\u7b97\u5b50\u7ea7\u52a0\u901f\uff0c\u8986\u76d6\u5377\u79ef\u3001\u5f52\u4e00\u5316\u3001\u6c60\u5316\u3001\u5faa\u73af\u795e\u7ecf\u7f51\u7edc\u7b49\u6838\u5fc3\u8fd0\u7b97\u3002\u8be5\u5e93\u878d\u5408\u4e86 Intel \u591a\u5e74\u7684\u82af\u7247\u4f18\u5316\u7ecf\u9a8c\uff0c\u53ef\u81ea\u52a8\u5229\u7528 Gaudi 3 \u7684\u77e9\u9635\u5f15\u64ce\u548c\u9ad8\u5e26\u5bbd\u5185\u5b58\uff0c\u663e\u8457\u63d0\u5347\u8bad\u7ec3\u4e0e\u63a8\u7406\u541e\u5410\u91cf\u3002<\/p>\n<p><a href=\"https:\/\/www.intel.com\/content\/www\/us\/en\/developer\/tools\/oneapi\/onednn.html\" target=\"_blank\">\u5b98\u65b9\u7f51\u7ad9<\/a><\/p>\n<h2>\u6838\u5fc3\u529f\u80fd\u4e0e\u4f18\u52bf<\/h2>\n<h3>\u7b97\u5b50\u7ea7\u6781\u81f4\u52a0\u901f<\/h3>\n<p>oneDNN \u9488\u5bf9 Gaudi 3 \u7684\u7279\u5b9a\u6307\u4ee4\u96c6\u8fdb\u884c\u624b\u5199\u6c47\u7f16\u7ea7\u8c03\u4f18\uff0c\u5728\u5377\u79ef\u3001GEMM\uff08\u901a\u7528\u77e9\u9635\u4e58\u6cd5\uff09\u7b49\u5173\u952e\u7b97\u5b50\u4e0a\u5b9e\u73b0 2-3 \u500d\u6027\u80fd\u63d0\u5347\u3002\u5e93\u5185\u7f6e\u81ea\u52a8\u8c03\u4f18\u5f15\u64ce\uff0c\u53ef\u6839\u636e\u8f93\u5165\u5f62\u72b6\u81ea\u52a8\u9009\u62e9\u6700\u4f18\u5185\u6838\u3002<\/p>\n<h3>\u591a\u6846\u67b6\u65e0\u7f1d\u96c6\u6210<\/h3>\n<p>\u8be5\u5e93\u63d0\u4f9b\u6807\u51c6 C++ \u548c SYCL \u63a5\u53e3\uff0c\u539f\u751f\u652f\u6301 TensorFlow\u3001PyTorch\u3001ONNX Runtime \u7b49\u4e3b\u6d41\u6846\u67b6\u3002\u5f00\u53d1\u8005\u65e0\u9700\u4fee\u6539\u6a21\u578b\u4ee3\u7801\uff0c\u4ec5\u9700\u901a\u8fc7\u73af\u5883\u53d8\u91cf\u6216 API \u8c03\u7528\u5373\u53ef\u5c06\u8fd0\u7b97\u5378\u8f7d\u81f3 Gaudi 3\u3002<\/p>\n<ul>\n<li>\u652f\u6301\u52a8\u6001\u5f62\u72b6\u4e0e\u7a00\u758f\u8ba1\u7b97\uff0c\u51cf\u5c11\u5197\u4f59\u5185\u5b58\u8bbf\u95ee\u3002<\/li>\n<li>\u5185\u7f6e FP16\u3001BF16\u3001INT8 \u6df7\u5408\u7cbe\u5ea6\u652f\u6301\uff0c\u964d\u4f4e\u663e\u5b58\u5360\u7528\u3002<\/li>\n<li>\u63d0\u4f9b\u56fe\u7ea7\u4f18\u5316\u7f16\u8bd1\u5668\uff0c\u878d\u5408\u591a\u4e2a\u7b97\u5b50\u4ee5\u51cf\u5c11\u6838\u542f\u52a8\u5f00\u9500\u3002<\/li>\n<\/ul>\n<h2>\u5178\u578b\u5e94\u7528\u573a\u666f<\/h2>\n<h3>\u5927\u89c4\u6a21\u8bed\u8a00\u6a21\u578b\u8bad\u7ec3<\/h3>\n<p>\u5728\u5343\u4ebf\u53c2\u6570 LLM \u8bad\u7ec3\u4e2d\uff0coneDNN \u5229\u7528 Gaudi 3 \u7684 Tile \u67b6\u6784\u548c\u6ce2\u524d\u8c03\u5ea6\u6280\u672f\uff0c\u4f7f\u6ce8\u610f\u529b\u673a\u5236\u548c\u524d\u9988\u7f51\u7edc\u7684\u8ba1\u7b97\u6548\u7387\u63d0\u5347 40% \u4ee5\u4e0a\u3002\u914d\u5408\u5206\u5e03\u5f0f\u6570\u636e\u5e76\u884c\uff0c\u53ef\u7ebf\u6027\u6269\u5c55\u81f3\u6570\u767e\u4e2a\u52a0\u901f\u5668\u3002<\/p>\n<h3>\u5b9e\u65f6\u63a8\u7406\u670d\u52a1<\/h3>\n<p>\u5bf9\u4e8e\u7535\u5546\u63a8\u8350\u3001\u77ed\u89c6\u9891\u5206\u7c7b\u7b49\u5ef6\u65f6\u654f\u611f\u573a\u666f\uff0coneDNN \u7684\u63a8\u7406\u540e\u7aef\u652f\u6301\u52a8\u6001\u6279\u5904\u7406\u4e0e\u79bb\u7ebf\u6743\u91cd\u538b\u7f29\uff0c\u5728 Gaudi 3 \u4e0a\u53ef\u5b9e\u73b0 &lt;5ms \u7684\u5355\u6837\u672c\u63a8\u7406\u5ef6\u8fdf\u3002<\/p>\n<h3>\u8ba1\u7b97\u673a\u89c6\u89c9\u4e0e\u79d1\u5b66\u8ba1\u7b97<\/h3>\n<p>\u56fe\u50cf\u5206\u7c7b\u3001\u76ee\u6807\u68c0\u6d4b\u3001\u5206\u5b50\u52a8\u529b\u5b66\u6a21\u62df\u7b49\u4efb\u52a1\u5747\u53ef\u53d7\u76ca\u3002\u5e93\u63d0\u4f9b\u7684\u91cf\u5316\u548c\u4fee\u526a\u5de5\u5177\u80fd\u5c06\u6a21\u578b\u4f53\u79ef\u7f29\u51cf 70% \u4ee5\u4e0a\uff0c\u540c\u65f6\u4fdd\u6301\u7cbe\u5ea6\u635f\u5931\u5728 0.5% \u4ee5\u5185\u3002<\/p>\n<h2>\u5feb\u901f\u5165\u95e8\u6307\u5357<\/h2>\n<p>\u5f00\u53d1\u8005\u53ef\u901a\u8fc7 Intel \u5b98\u65b9\u5bb9\u5668\u955c\u50cf\u83b7\u53d6\u9884\u7f16\u8bd1\u73af\u5883\uff0c\u6216\u4ece GitHub \u4ed3\u5e93\u6e90\u7801\u7f16\u8bd1\u3002\u4ee5\u4e0b\u4e3a\u5178\u578b\u5b89\u88c5\u6b65\u9aa4\uff1a<\/p>\n<ul>\n<li>\u901a\u8fc7 pip \u5b89\u88c5 oneDNN Python \u7ed1\u5b9a\uff1a<code>pip install intel-extension-for-pytorch<\/code><\/li>\n<li>\u5728 PyTorch \u811a\u672c\u4e2d\u6dfb\u52a0 <code>import oneccl_bindings_for_pytorch<\/code> \u4ee5\u542f\u7528 Gaudi 3 \u901a\u4fe1\u4f18\u5316\u3002<\/li>\n<li>\u8fd0\u884c\u5b98\u65b9\u63d0\u4f9b\u7684 ResNet-50 \u57fa\u51c6\u6d4b\u8bd5\u9a8c\u8bc1\u6027\u80fd\uff1a<code>python benchmarks\/benchmark.py --model resnet50 --device hpu<\/code><\/li>\n<\/ul>\n<p>Intel \u8fd8\u63d0\u4f9b\u8be6\u5c3d\u7684\u4f18\u5316\u6307\u5357\u548c\u8c03\u4f18\u5de5\u5177\uff0c\u5e2e\u52a9\u5f00\u53d1\u8005\u9488\u5bf9\u7279\u5b9a\u6a21\u578b\u8fdb\u4e00\u6b65\u6316\u6f5c\u3002<\/p>\n<hr \/>\n<p><strong>\u3010\u70ed\u70b9\u65b0\u95fb\u3011<\/strong><\/p>\n<p><strong>\u3010\u6807\u9898\u3011<\/strong>\u56fd\u4ea7\u5927\u6a21\u578b DeepSeek \u53d1\u5e03\u65b0\u4e00\u4ee3\u63a8\u7406\u6a21\u578b\uff0c\u591a\u9879\u57fa\u51c6\u8d85\u8d8a GPT-4o<br \/><strong>\u3010\u5206\u7c7b\u3011<\/strong>\u79d1\u6280<br \/><strong>\u3010\u6b63\u6587\u3011<\/strong>\u8fd1\u65e5\uff0c\u56fd\u4ea7\u5927\u6a21\u578b\u516c\u53f8 DeepSeek \u6b63\u5f0f\u53d1\u5e03\u65b0\u4e00\u4ee3\u63a8\u7406\u6a21\u578b DeepSeek-R2\uff0c\u5728\u6570\u5b66\u7ade\u8d5b\u3001\u4ee3\u7801\u751f\u6210\u3001\u591a\u8f6e\u5bf9\u8bdd\u7b49\u591a\u9879\u6743\u5a01\u8bc4\u6d4b\u4e2d\u5747\u8fbe\u5230\u56fd\u9645\u9886\u5148\u6c34\u5e73\u3002\u8be5\u6a21\u578b\u91c7\u7528 MoE-Transformer \u6df7\u5408\u67b6\u6784\uff0c\u6fc0\u6d3b\u53c2\u6570\u4ec5 37B\uff0c\u5374\u5b9e\u73b0\u4e86\u4e0e 200B \u7ea7\u522b\u6a21\u578b\u76f8\u5f53\u7684\u63a8\u7406\u80fd\u529b\u3002\u4e1a\u5185\u4e13\u5bb6\u8868\u793a\uff0cDeepSeek-R2 \u7684\u53d1\u5e03\u6807\u5fd7\u7740\u4e2d\u56fd AI \u5728\u5e95\u5c42\u7b97\u6cd5\u548c\u7b97\u529b\u4f18\u5316\u4e0a\u53d6\u5f97\u91cd\u5927\u7a81\u7834\uff0c\u5c06\u4e3a\u91d1\u878d\u3001\u533b\u7597\u7b49\u9886\u57df\u7684\u667a\u80fd\u5316\u5347\u7ea7\u63d0\u4f9b\u66f4\u81ea\u4e3b\u53ef\u63a7\u7684\u57fa\u5ea7\u3002<br \/><strong>\u3010\u6765\u6e90\u3011<\/strong><a href=\"https:\/\/www.ithome.com\/0\/123456.htm\" target=\"_blank\">IT\u4e4b\u5bb6<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>Intel oneAPI Deep Neural Network Library\uff08\u7b80\u79f0 oneDNN\uff09\u662f\u9488\u5bf9\u6df1 [&hellip;]<\/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":[9447,10971,10966,2993,10972],"class_list":["post-8778","post","type-post","status-publish","format-standard","hentry","category-4","tag-ai","tag-gaudi-3","tag-intel-oneapi","tag-2993","tag-10972"],"_links":{"self":[{"href":"https:\/\/ab123.xyz\/index.php?rest_route=\/wp\/v2\/posts\/8778","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/ab123.xyz\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/ab123.xyz\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/ab123.xyz\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/ab123.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=8778"}],"version-history":[{"count":1,"href":"https:\/\/ab123.xyz\/index.php?rest_route=\/wp\/v2\/posts\/8778\/revisions"}],"predecessor-version":[{"id":8779,"href":"https:\/\/ab123.xyz\/index.php?rest_route=\/wp\/v2\/posts\/8778\/revisions\/8779"}],"wp:attachment":[{"href":"https:\/\/ab123.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=8778"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/ab123.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=8778"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/ab123.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=8778"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}