{"id":1152393,"date":"2025-10-16T20:33:03","date_gmt":"2025-10-17T03:33:03","guid":{"rendered":"https:\/\/www.microsoft.com\/en-us\/research\/?post_type=msr-blog-post&p=1152393"},"modified":"2025-10-16T20:33:04","modified_gmt":"2025-10-17T03:33:04","slug":"pipethreader","status":"publish","type":"msr-blog-post","link":"https:\/\/www.microsoft.com\/en-us\/research\/articles\/pipethreader\/","title":{"rendered":"PipeThreader\uff1a\u8ba9AI\u82af\u7247\u201c\u5404\u53f8\u5176\u804c\u201d\u7684\u667a\u80fd\u6d41\u6c34\u7ebf\u7f16\u8bd1\u5668"},"content":{"rendered":"\n

\u7f16\u8005\u6309\uff1a\u5f53\u5927\u6a21\u578b\u7684\u89c4\u6a21\u4e0d\u65ad\u7a81\u7834\u6781\u9650\uff0cAI \u8bad\u7ec3\u7684\u7b97\u529b\u9700\u6c42\u4e5f\u5728\u4ee5\u6307\u6570\u7ea7\u6500\u5347\u3002\u4f20\u7edf GPU \u5df2\u96be\u4ee5\u517c\u987e\u6027\u80fd\u4e0e\u80fd\u6548\uff0c\u5f02\u6784 AI \u82af\u7247\u9010\u6e10\u6210\u4e3a\u65b0\u8d8b\u52bf\u3002\u4f46\u5982\u4f55\u8ba9\u4e0d\u540c\u8ba1\u7b97\u5355\u5143\u201c\u534f\u540c\u4f5c\u6218\u201d\uff0c\u4ecd\u662f\u7f16\u8bd1\u9886\u57df\u7684\u4e00\u5927\u96be\u9898\u3002\u5317\u4eac\u5927\u5b66\u3001\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u3001\u5e1d\u56fd\u7406\u5de5\u5b66\u9662\u4e0e\u4e0a\u6d77\u4ea4\u901a\u5927\u5b66\u8054\u5408\u56e2\u961f\u7814\u53d1\u7684 PipeThreader \u7cfb\u7edf\uff0c\u65e8\u5728\u89e3\u51b3\u8fd9\u4e00\u6838\u5fc3\u6311\u6218\u3002PipeThreader \u9996\u6b21\u63d0\u51fa\u201c\u6d41\u6c34\u7ebf\u5bfc\u5411\u7f16\u8bd1\u201d\u7406\u5ff5\uff0c\u8ba9\u7f16\u8bd1\u5668\u80fd\u591f\u50cf\u201c\u6d41\u6c34\u7ebf\u8bbe\u8ba1\u5e08\u201d\u4e00\u6837\uff0c\u4e3a\u5f02\u6784\u82af\u7247\u81ea\u52a8\u89c4\u5212\u6700\u4f18\u6267\u884c\u65b9\u6848\u3002\u5b9e\u9a8c\u8868\u660e\uff0cPipeThreader \u5728 Transformer\u3001FlashAttention\u3001Mamba \u7b49\u6a21\u578b\u4e0a\u5747\u53d6\u5f97\u663e\u8457\u52a0\u901f\u6548\u679c\uff0c\u6700\u9ad8\u53ef\u5b9e\u73b0\u8fd1\u4e24\u500d\u6027\u80fd\u63d0\u5347\uff0c\u5e76\u80fd\u8f7b\u677e\u8fc1\u79fb\u81f3\u4e0d\u540c\u786c\u4ef6\u5e73\u53f0\u3002\u8fd9\u4e00\u6210\u679c\u4e0d\u4ec5\u7a81\u7834\u4e86\u4f20\u7edf\u624b\u5de5\u8c03\u5ea6\u7684\u74f6\u9888\uff0c\u4e5f\u9884\u793a\u7740\u667a\u80fd\u7f16\u8bd1\u7684\u65b0\u65b9\u5411\u2014\u2014\u8ba9\u7f16\u8bd1\u5668\u7406\u89e3\u786c\u4ef6\u3001\u81ea\u52a8\u4f18\u5316\u7b97\u529b\u3002\u76f8\u5173\u8bba\u6587\u5df2\u88ab OSDI 2025 \u63a5\u6536\u3002<\/p>\n\n\n\n


\n\n\n\n

\u968f\u7740\u5927\u6a21\u578b\u89c4\u6a21\u7684\u6025\u5267\u81a8\u80c0\uff0cAI \u8bad\u7ec3\u7684\u8ba1\u7b97\u9700\u6c42\u5448\u6307\u6570\u7ea7\u4e0a\u5347\u3002\u4f20\u7edf\u7684 GPU \u52a0\u901f\u65b9\u6848\u5df2\u7ecf\u96be\u4ee5\u6ee1\u8db3\u6548\u7387\u4e0e\u80fd\u8017\u7684\u53cc\u91cd\u8981\u6c42\u3002\u5982\u4eca\u8d8a\u6765\u8d8a\u591a\u7684\u5382\u5546\u5f00\u59cb\u8bbe\u8ba1\u5f02\u6784 AI \u82af\u7247<\/strong>\uff0c\u5b83\u4eec\u96c6\u6210\u4e86\u4e0d\u540c\u7c7b\u578b\u7684\u8ba1\u7b97\u5355\u5143\uff0c\u5982\u77e9\u9635\u4e58\u6cd5\u52a0\u901f\u5668\u3001\u5411\u91cf\u5355\u5143\u3001\u901a\u7528\u6838\u3001DMA \u5f15\u64ce\u7b49\uff0c\u4ee5\u6c42\u5728\u6027\u80fd\u4e0e\u80fd\u6548\u4e4b\u95f4\u53d6\u5f97\u5e73\u8861\u3002<\/p>\n\n\n\n

\u7136\u800c\uff0c\u65b0\u7684\u95ee\u9898\u968f\u4e4b\u800c\u6765\u3002\u8fd9\u4e9b\u6267\u884c\u5355\u5143\u7684\u8ba1\u7b97\u7279\u6027\u548c\u5e26\u5bbd\u7ea6\u675f\u5404\u4e0d\u76f8\u540c\uff0c\u5982\u4f55\u8ba9\u5b83\u4eec\u534f\u540c\u5de5\u4f5c\u3001\u5b9e\u73b0\u9ad8\u6548\u6d41\u6c34\u7ebf\u5e76\u884c<\/strong>\uff0c\u6210\u4e3a\u00a0AI\u00a0\u7f16\u8bd1\u5668\u9762\u4e34\u7684\u6700\u5927\u6311\u6218\u4e4b\u4e00\u3002<\/p>\n\n\n\n

\u624b\u5de5\u6392\u5e03\u7b97\u5b50\uff08operator pipeline\uff09\u51e0\u4e4e\u65e0\u6cd5\u5e94\u5bf9\u73b0\u4ee3\u6a21\u578b\u7684\u590d\u6742\u5ea6\uff0c\u6bcf\u4e00\u5c42\u3001\u6bcf\u4e00\u4e2a\u7b97\u5b50\uff0c\u90fd\u53ef\u80fd\u8de8\u8d8a\u591a\u4e2a\u5f02\u6784\u5355\u5143\u6267\u884c\uff0c\u7a0d\u6709\u4e0d\u614e\uff0c\u5c31\u53ef\u80fd\u5bfc\u81f4\u8d44\u6e90\u6d6a\u8d39\u6216\u6d41\u6c34\u7ebf\u963b\u585e\u3002\u76ee\u524d\u5e7f\u6cdb\u4f7f\u7528\u7684\u00a0FlashAttention-3 \u00a0\u7b97\u5b50\u5373\u4e3a\u4e00\u4e2a\u5178\u578b\u4f8b\u8bc1\u3002\u5b83\u901a\u8fc7\u7cbe\u5fc3\u8bbe\u8ba1\u7684\u591a\u9636\u6bb5\u6d41\u6c34\u7ebf\uff0c\u5728\u663e\u5b58\u8bbf\u95ee\u4e0e\u8ba1\u7b97\u4e4b\u95f4\u5b9e\u73b0\u4e86\u6781\u9ad8\u7684\u5e76\u884c\u6548\u7387\uff0c\u4f46\u8fd9\u79cd\u201c\u624b\u5de5\u8c03\u5ea6\u201d\u7684\u65b9\u5f0f\u5f80\u5f80\u9700\u8981\u7ecf\u9a8c\u4e30\u5bcc\u7684\u4e13\u5bb6\uff0c\u82b1\u8d39\u5927\u91cf\u65f6\u95f4\u53cd\u590d\u8bd5\u9a8c\u4e0e\u8c03\u4f18\uff0c\u4e0d\u4ec5\u5f00\u53d1\u5468\u671f\u957f\u3001\u53ef\u79fb\u690d\u6027\u5dee\uff0c\u8fd8\u6781\u6613\u51fa\u9519\u3002<\/p>\n\n\n\n

\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e00\u95ee\u9898\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u4e0e\u6765\u81ea\u5317\u4eac\u5927\u5b66\u3001\u5e1d\u56fd\u7406\u5de5\u5b66\u9662\u4ee5\u53ca\u4e0a\u6d77\u4ea4\u901a\u5927\u5b66\u7684\u79d1\u7814\u4eba\u5458\u4eec\u5408\u4f5c\uff0c\u63d0\u51fa\u4e86\u4e00\u4e2a\u9762\u5411\u5f02\u6784\u00a0AI\u00a0\u82af\u7247\u7684\u667a\u80fd\u6d41\u6c34\u7ebf\u7f16\u8bd1\u4e0e\u8c03\u5ea6\u7cfb\u7edf\u00a0PipeThreader<\/strong>\u3002\u5b83\u7684\u76ee\u6807\u975e\u5e38\u660e\u786e\uff1a\u8ba9\u00a0AI\u00a0\u7f16\u8bd1\u5668\u80fd\u591f\u81ea\u52a8\u5b66\u4f1a\u5728\u5f02\u6784\u786c\u4ef6\u4e0a\u201c\u6392\u5175\u5e03\u9635\u201d\uff0c\u8ba9\u6bcf\u4e2a\u7b97\u529b\u5355\u5143\u5404\u53f8\u5176\u804c\u3001\u534f\u540c\u9ad8\u6548<\/strong>\u3002\u76f8\u5173\u8bba\u6587\u5df2\u88ab\u00a0OSDI 2025\u00a0\u6536\u5f55\u3002<\/p>\n\n\n\n

PipeThreader: Software-Defined Pipelining for Efficient DNN Execution<\/p>\n\n\n\n

\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/www.usenix.org\/conference\/osdi25\/presentation\/cheng (opens in new tab)<\/span><\/a><\/p>\n\n\n\n

\u5f02\u6784\u6267\u884c\u7684\u201c\u4e71\u9ebb\u201d\u56f0\u5883<\/h2>\n\n\n\n

\u5f53\u4eca\u5e7f\u6cdb\u4f7f\u7528\u7684\u5f02\u6784\u67b6\u6784 AI \u82af\u7247\uff0c\u53ef\u80fd\u540c\u65f6\u62e5\u6709\u591a\u4e2a\u4e0d\u540c\u7684\u4e13\u6709\u8ba1\u7b97\u5355\u5143\uff1a\u4e13\u7528\u4e8e\u77e9\u9635\u4e58\u6cd5\u7684 Tensor Engine\u3001\u5904\u7406\u901a\u7528\u8ba1\u7b97\u4efb\u52a1\u7684 General Compute Core \u4ee5\u53ca\u8fdb\u884c\u7247\u4e0a\/\u7247\u5916\u6570\u636e\u642c\u8fd0\u7684 DMA \u5355\u5143\u3002<\/p>\n\n\n\n

\u8fd9\u4e9b\u5355\u5143\u7684\u6027\u80fd\u5dee\u5f02\u6781\u5927\uff0c\u5982\u679c\u7f16\u8bd1\u5668\u76f4\u63a5\u7167\u642c\u201c\u9759\u6001\u7b97\u5b50\u56fe\u201d\u7684\u6267\u884c\u987a\u5e8f\uff0c\u5f80\u5f80\u4f1a\u51fa\u73b0\u90e8\u5206\u6a21\u5757\u7a7a\u8f6c\u3001\u6570\u636e\u963b\u585e\u7684\u60c5\u51b5\uff0c\u6700\u7ec8\u5bfc\u81f4\u6574\u4f53\u6027\u80fd\u4e25\u91cd\u4e0b\u6ed1\u3002<\/p>\n\n\n\n

\"diagram\"
\u56fe1\uff1a\u201c\u9759\u6001\u7b97\u5b50\u56fe\u201d\u6267\u884c\u65b9\u6848\uff08\u5de6\uff09\u548c\u4f18\u5316\u540e\u7684\u6d41\u6c34\u7ebf\u5e76\u884c\u6267\u884c\u65b9\u6848\uff08\u53f3\uff09<\/figcaption><\/figure>\n\n\n\n

\u4f20\u7edf\u201c\u4ee5\u7b97\u5b50\u4e3a\u4e2d\u5fc3\u201d\u7684\u7f16\u8bd1\u601d\u8def\u5df2\u7ecf\u65e0\u6cd5\u9002\u5e94\u8fd9\u79cd\u788e\u7247\u5316\u7684\u7b97\u529b\u683c\u5c40\u3002\u9488\u5bf9\u8fd9\u4e00\u60c5\u51b5\uff0cPipeThreader \u91cd\u65b0\u5b9a\u4e49\u4e86 AI \u7f16\u8bd1\u7684\u76ee\u6807\u2014\u2014\u4ece\u7b97\u5b50\u7f16\u8bd1\u8f6c\u5411\u6d41\u6c34\u7ebf\u7f16\u8bd1\uff08pipeline-oriented compilation\uff09\u3002<\/p>\n\n\n\n

PipeThreader \u8ba9\u7f16\u8bd1\u5668\u6210\u4e3a\u6d41\u6c34\u7ebf\u8bbe\u8ba1\u5e08<\/h2>\n\n\n\n

PipeThreader\u00a0\u7684\u8bbe\u8ba1\u7075\u611f\u6765\u81ea\u4e8e\u786c\u4ef6\u6d41\u6c34\u7ebf\u4e0e\u64cd\u4f5c\u7cfb\u7edf\u7ebf\u7a0b\u8c03\u5ea6\u7684\u7ed3\u5408\u3002\u7814\u7a76\u5458\u4eec\u5c06\u6a21\u578b\u6267\u884c\u8fc7\u7a0b\u62bd\u8c61\u4e3a\u4e00\u7ec4\u201c\u4e13\u6709\u6267\u884c\u4efb\u52a1\u201d\uff08specialized task\uff09\uff0c\u5e76\u4e3a\u6bcf\u4e2a\u4e13\u6709\u6267\u884c\u4efb\u52a1\u5206\u914d\u4e00\u4e2a\u201c\u5f02\u6784\u6267\u884c\u5355\u5143\u201d\uff08specialized execution unit\uff09\u3002\u8fd9\u4e00\u770b\u4f3c\u7b80\u5355\u7684\u91cd\u65b0\u62bd\u8c61\uff0c\u5374\u5e26\u6765\u4e86\u8d28\u7684\u98de\u8dc3\u3002<\/p>\n\n\n\n

PipeThreader\u00a0\u4e0d\u518d\u5355\u72ec\u5173\u6ce8\u6bcf\u4e2a\u7b97\u5b50\u7684\u6267\u884c\u65b9\u6848\uff0c\u800c\u662f\u4ece\u66f4\u9ad8\u7684\u5c42\u9762\u8fdb\u884c\u6574\u4f53\u89c4\u5212\uff1a\u5728\u8fd9\u4e9b\u5f02\u6784\u6267\u884c\u5355\u5143\u4e0a\uff0c\u54ea\u4e9b\u4efb\u52a1\u53ef\u4ee5\u5e76\u884c\u6267\u884c\u3001\u54ea\u4e9b\u6570\u636e\u901a\u8def\u4f1a\u5f62\u6210\u74f6\u9888\u3001\u5982\u4f55\u5728\u5f02\u6784\u5355\u5143\u95f4\u5e73\u8861\u8d1f\u8f7d\u3002<\/p>\n\n\n\n

\u4e3a\u4e86\u5728\u590d\u6742\u7684\u5f02\u6784\u786c\u4ef6\u4e0a\u5b9e\u73b0\u9ad8\u6548\u7684\u6d41\u6c34\u7ebf\u6267\u884c\uff0cPipeThreader\u00a0\u63d0\u51fa\u4e86\u4e09\u4e2a\u5173\u952e\u62bd\u8c61\uff1asTask\u3001sGraph\u00a0\u548c\u00a0sEU\u3002\u8fd9\u4e09\u8005\u6784\u6210\u4e86\u7cfb\u7edf\u7684\u6838\u5fc3\u8bed\u4e49\u5c42\uff0c\u4f7f\u7f16\u8bd1\u5668\u80fd\u591f\u4ece\u4efb\u52a1\u3001\u4f9d\u8d56\u5173\u7cfb\u548c\u786c\u4ef6\u8d44\u6e90\u4e09\u4e2a\u7ef4\u5ea6\u7edf\u4e00\u7406\u89e3\u6574\u4e2a\u8ba1\u7b97\u8fc7\u7a0b\u3002<\/p>\n\n\n\n