{"id":1171,"date":"2026-03-29T14:14:31","date_gmt":"2026-03-29T06:14:31","guid":{"rendered":"https:\/\/www.ndnlab.com\/?p=1171"},"modified":"2026-03-29T14:15:52","modified_gmt":"2026-03-29T06:15:52","slug":"megascale-scaling-large-language-model-training-to-more-than-10000-gpus","status":"publish","type":"post","link":"https:\/\/www.ndnlab.com\/?p=1171","title":{"rendered":"MegaScale: Scaling Large Language Model Training to More Than 10,000 GPUs"},"content":{"rendered":"\n<p>NSDI &#8217;24: 21st USENIX Symposium on Networked Systems Design and Implementation, April 16\u201318, 2024, Santa Clara, CA, USA<\/p>\n\n\n\n<p><a href=\"https:\/\/www.usenix.org\/system\/files\/nsdi24-jiang-ziheng.pdf\">https:\/\/www.usenix.org\/system\/files\/nsdi24-jiang-ziheng.pdf<\/a><\/p>\n\n\n\n<p>\u4e00\u3001\u7814\u7a76\u80cc\u666f\u4e0e\u52a8\u673a\uff1a\u4e07\u5361\u8bad\u7ec3\u65f6\u4ee3\u7684\u7cfb\u7edf\u6311\u6218<\/p>\n\n\n\n<p>\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u7684\u8bad\u7ec3\u89c4\u6a21\u5df2\u4ece\u6570\u767e\u5757 GPU \u8dc3\u5347\u81f3\u6570\u4e07\u5757 GPU\u3002\u4ee5 GPT-3\uff081750 \u4ebf\u53c2\u6570\uff09\u548c PaLM\uff085400 \u4ebf\u53c2\u6570\uff09\u4e3a\u4ee3\u8868\u7684\u524d\u6cbf\u6a21\u578b\uff0c\u8981\u6c42\u5728\u6570\u5468\u751a\u81f3\u6570\u6708\u5185\u6301\u7eed\u5360\u7528\u6574\u4e2a\u96c6\u7fa4\u8fdb\u884c\u8bad\u7ec3\u3002\u8fd9\u4e0e\u4f20\u7edf\u7684\u6df1\u5ea6\u5b66\u4e60\u8bad\u7ec3\u4efb\u52a1\u5f62\u6210\u4e86\u9c9c\u660e\u5bf9\u6bd4\u2014\u2014\u8fc7\u53bb\u4e00\u4e2a ResNet \u8bad\u7ec3\u4efb\u52a1\u53ea\u9700\u8981\u51e0\u5341\u5230\u51e0\u767e\u5757 GPU\uff0c\u800c\u5982\u4eca\u4e00\u4e2a LLM \u8bad\u7ec3\u4efb\u52a1\u5c31\u9700\u8981\u72ec\u5360\u4e0a\u4e07\u5757 GPU\u3002<\/p>\n\n\n\n<p>\u5728\u8fd9\u79cd\u524d\u6240\u672a\u6709\u7684\u89c4\u6a21\u4e0b\uff0c\u7cfb\u7edf\u9762\u4e34\u4e24\u5927\u6838\u5fc3\u6311\u6218\u3002\u7b2c\u4e00\u662f\u8bad\u7ec3\u6548\u7387\u95ee\u9898\u3002\u6a21\u578b FLOPs \u5229\u7528\u7387\uff08MFU\uff09\u662f\u8861\u91cf\u8bad\u7ec3\u6548\u7387\u7684\u6807\u51c6\u6307\u6807\uff0c\u5b83\u53cd\u6620\u4e86\u5b9e\u9645\u541e\u5410\u91cf\u4e0e\u786c\u4ef6\u7406\u8bba\u5cf0\u503c\u7684\u6bd4\u503c\u3002LLM \u8bad\u7ec3\u5e76\u975e\u7b80\u5355\u7684\u6570\u636e\u5e76\u884c\uff0c\u800c\u662f\u9700\u8981\u5728\u6570\u636e\u5e76\u884c\u3001\u6d41\u6c34\u7ebf\u5e76\u884c\u548c\u5f20\u91cf\u5e76\u884c\u4e09\u4e2a\u7ef4\u5ea6\u4e0a\u540c\u65f6\u5206\u914d\u8ba1\u7b97\u548c\u901a\u4fe1\uff0c\u4efb\u4f55\u4e00\u4e2a\u73af\u8282\u7684\u4f4e\u6548\u90fd\u4f1a\u62c9\u4f4e\u6574\u4f53 MFU\u3002\u7b2c\u4e8c\u662f\u8bad\u7ec3\u7a33\u5b9a\u6027\u95ee\u9898\u3002\u8bad\u7ec3\u4e00\u4e2a\u4e07\u4ebf token \u7684\u6a21\u578b\u53ef\u80fd\u6301\u7eed\u6570\u5468\uff0c\u5728\u8fd9\u4e2a\u65f6\u95f4\u8de8\u5ea6\u5185\uff0c\u786c\u4ef6\u6545\u969c\u548c\u6027\u80fd\u6389\u961f\u8282\u70b9\uff08straggler\uff09\u662f\u5e38\u6001\u800c\u975e\u4f8b\u5916\u3002\u4e00\u6b21\u6545\u969c\u53ef\u80fd\u5bfc\u81f4\u4e0a\u4e07\u5757 GPU \u5168\u90e8\u505c\u6ede\uff0c\u800c\u56de\u9000\u5230\u4e0a\u4e00\u4e2a\u68c0\u67e5\u70b9\u53ef\u80fd\u610f\u5473\u7740\u6570\u5c0f\u65f6\u7684\u8bad\u7ec3\u8fdb\u5ea6\u4e22\u5931\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"461\" height=\"217\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/image-33.png\"  class=\"wp-image-1177\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/image-33.png 461w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/image-33-300x141.png 300w\" sizes=\"auto, (max-width: 461px) 100vw, 461px\" title=\"MegaScale: Scaling Large Language Model Training to More Than 10,000 GPUs\u63d2\u56fe\" alt=\"MegaScale: Scaling Large Language Model Training to More Than 10,000 GPUs\u63d2\u56fe\" \/><\/figure>\n\n\n\n<p>\u4e8c\u3001\u6838\u5fc3\u65b9\u6cd5\uff1a\u7b97\u6cd5\u4e0e\u7cfb\u7edf\u7684\u5168\u6808\u534f\u540c\u8bbe\u8ba1<\/p>\n\n\n\n<p>MegaScale \u91c7\u7528\u4e86\u7b97\u6cd5-\u7cfb\u7edf\u534f\u540c\u8bbe\u8ba1\u7684\u539f\u5219\uff0c\u5728\u6a21\u578b\u67b6\u6784\u3001\u901a\u4fe1\u7b56\u7565\u3001\u7b97\u5b50\u4f18\u5316\u3001\u6570\u636e\u7ba1\u9053\u548c\u7f51\u7edc\u8c03\u4f18\u7b49\u5c42\u9762\u8fdb\u884c\u4e86\u5168\u65b9\u4f4d\u4f18\u5316\u3002<\/p>\n\n\n\n<p>\u5728\u7b97\u6cd5\u5c42\u9762\uff0cMegaScale \u5f15\u5165\u4e86\u4e09\u9879\u5173\u952e\u6539\u8fdb\u3002\u7b2c\u4e00\u662f\u5e76\u884c Transformer \u5757\uff0c\u5c06\u4f20\u7edf\u4e32\u884c\u6267\u884c\u7684\u6ce8\u610f\u529b\u6a21\u5757\u548c MLP \u6a21\u5757\u6539\u4e3a\u5e76\u884c\u6267\u884c\uff0c\u51cf\u5c11\u4e86\u8ba1\u7b97\u65f6\u95f4\u3002\u7b2c\u4e8c\u662f\u6ed1\u52a8\u7a97\u53e3\u6ce8\u610f\u529b\uff0c\u5c06\u6ce8\u610f\u529b\u8ba1\u7b97\u7684\u590d\u6742\u5ea6\u4ece O(s\u00d7s) \u964d\u4f4e\u5230 O(s\u00d7w)\uff0c\u5176\u4e2d w \u8fdc\u5c0f\u4e8e\u5e8f\u5217\u957f\u5ea6 s\uff0c\u5728\u4e0d\u635f\u5931\u6a21\u578b\u7cbe\u5ea6\u7684\u524d\u63d0\u4e0b\u663e\u8457\u63d0\u5347\u4e86\u8bad\u7ec3\u901f\u5ea6\u3002\u7b2c\u4e09\u662f LAMB \u4f18\u5316\u5668\uff0c\u4f7f\u5f97\u6279\u91cf\u5927\u5c0f\u53ef\u4ee5\u6269\u5927 4 \u500d\u800c\u4e0d\u5f71\u54cd\u6a21\u578b\u6536\u655b\uff0c\u4ece\u800c\u5c06\u6d41\u6c34\u7ebf\u5e76\u884c\u4e2d\u7684\u6c14\u6ce1\uff08pipeline bubble\uff09\u51cf\u5c11\u4e86 87.5%\u3002<\/p>\n\n\n\n<p>\u5728\u901a\u4fe1\u5c42\u9762\uff0cMegaScale \u9488\u5bf9 3D \u5e76\u884c\u4e2d\u7684\u6bcf\u79cd\u901a\u4fe1\u6a21\u5f0f\u8bbe\u8ba1\u4e86\u4e13\u95e8\u7684\u91cd\u53e0\u7b56\u7565\u3002\u5bf9\u4e8e\u6570\u636e\u5e76\u884c\uff0c\u5c06 AllGather \u64cd\u4f5c\u9884\u53d6\u5230\u8fed\u4ee3\u5f00\u59cb\u9636\u6bb5\uff0c\u4f7f\u5176\u4e0e\u6570\u636e\u52a0\u8f7d\u91cd\u53e0\u3002\u5bf9\u4e8e\u6d41\u6c34\u7ebf\u5e76\u884c\uff0c\u5c06 Send \u548c Receive \u64cd\u4f5c\u89e3\u8026\uff0c\u4f7f\u53d1\u9001\u64cd\u4f5c\u80fd\u591f\u4e0e\u8ba1\u7b97\u91cd\u53e0\u3002\u5bf9\u4e8e\u5f20\u91cf\u5e76\u884c\u548c\u5e8f\u5217\u5e76\u884c\uff0c\u5c06 AllGather \u548c ReduceScatter \u64cd\u4f5c\u878d\u5408\u5230\u7ebf\u6027\u5c42\u7684 GEMM \u8ba1\u7b97\u4e2d\uff0c\u901a\u8fc7\u5c06 GEMM \u6838\u5207\u5206\u4e3a\u5c0f\u5757\u5e76\u4e0e\u901a\u4fe1\u8fdb\u884c\u6d41\u6c34\u7ebf\u6267\u884c\u6765\u9690\u85cf\u901a\u4fe1\u5f00\u9500\u3002<\/p>\n\n\n\n<p>\u5728\u6570\u636e\u7ba1\u9053\u65b9\u9762\uff0cMegaScale \u91c7\u7528\u5f02\u6b65\u6570\u636e\u9884\u5904\u7406\uff0c\u5728 GPU \u540c\u6b65\u68af\u5ea6\u65f6\u9884\u5904\u7406\u4e0b\u4e00\u6b65\u7684\u6570\u636e\u3002\u540c\u65f6\uff0c\u5229\u7528\u540c\u4e00\u8282\u70b9\u5185\u7684\u5f20\u91cf\u5e76\u884c\u7ec4\u8f93\u5165\u76f8\u540c\u7684\u7279\u70b9\uff0c\u91c7\u7528\u6811\u72b6\u6570\u636e\u52a0\u8f7d\u65b9\u5f0f\uff0c\u7531\u5355\u4e00\u52a0\u8f7d\u5668\u8bfb\u53d6\u6570\u636e\u5230\u5171\u4eab\u5185\u5b58\uff0c\u5404 GPU \u518d\u81ea\u884c\u62f7\u8d1d\uff0c\u6d88\u9664\u4e86\u5197\u4f59\u78c1\u76d8\u8bfb\u53d6\u3002<\/p>\n\n\n\n<p>\u5728\u7f51\u7edc\u5c42\u9762\uff0cMegaScale \u91c7\u7528\u4e09\u5c42 CLOS \u62d3\u6251\u8fde\u63a5\u8d85\u8fc7\u4e00\u4e07\u5757 GPU\uff0c\u901a\u8fc7 multi-rail \u65b9\u5f0f\u5c06 8 \u5757\u7f51\u5361\u8fde\u63a5\u5230 8 \u4e2a\u4e0d\u540c\u7684 ToR \u4ea4\u6362\u673a\u4ee5\u51cf\u5c11 ECMP \u54c8\u5e0c\u51b2\u7a81\uff0c\u5e76\u5f00\u53d1\u4e86\u878d\u5408 Swift \u548c DCQCN \u7684\u62e5\u585e\u63a7\u5236\u7b97\u6cd5\u6765\u964d\u4f4e PFC \u5f15\u53d1\u7684\u961f\u5934\u963b\u585e\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"891\" height=\"465\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/image-34.png\"  class=\"wp-image-1178\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/image-34.png 891w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/image-34-300x157.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/image-34-768x401.png 768w\" sizes=\"auto, (max-width: 891px) 100vw, 891px\" title=\"MegaScale: Scaling Large Language Model Training to More Than 10,000 GPUs\u63d2\u56fe1\" alt=\"MegaScale: Scaling Large Language Model Training to More Than 10,000 GPUs\u63d2\u56fe1\" \/><\/figure>\n\n\n\n<p>\u4e09\u3001\u5bb9\u9519\u4e0e\u8bca\u65ad\uff1a\u6df1\u5ea6\u53ef\u89c2\u6d4b\u6027\u9a71\u52a8\u7684\u7a33\u5b9a\u8bad\u7ec3<\/p>\n\n\n\n<p>MegaScale \u6784\u5efa\u4e86\u4e00\u5957\u5b8c\u6574\u7684\u5bb9\u9519\u8bad\u7ec3\u6846\u67b6\u3002Driver \u8fdb\u7a0b\u901a\u8fc7 Kubernetes \u7ba1\u7406\u8ba1\u7b97\u8d44\u6e90\uff0c\u6bcf\u4e2a Executor \u7ba1\u7406\u4e00\u4e2a\u8282\u70b9\u5e76\u5b9a\u671f\u53d1\u9001\u5fc3\u8df3\u3002\u5fc3\u8df3\u4fe1\u606f\u4e0d\u4ec5\u5305\u542b\u57fa\u672c\u72b6\u6001\uff0c\u8fd8\u5c01\u88c5\u4e86\u8bad\u7ec3\u8fdb\u7a0b\u65e5\u5fd7\u548c RDMA \u6d41\u91cf\u6307\u6807\u3002\u5f53 Driver \u68c0\u6d4b\u5230\u5f02\u5e38\u6216\u5fc3\u8df3\u8d85\u65f6\u65f6\uff0c\u4f1a\u89e6\u53d1\u6545\u969c\u6062\u590d\u6d41\u7a0b\uff1a\u6682\u505c\u6240\u6709\u8bad\u7ec3\u4efb\u52a1\uff0c\u547d\u4ee4\u5404\u8282\u70b9\u8fd0\u884c\u8f7b\u91cf\u7ea7\u81ea\u68c0\u8bca\u65ad\uff08\u5305\u62ec\u73af\u56de\u5e26\u5bbd\u6d4b\u8bd5\u3001RNIC \u95f4\u8fde\u901a\u6027\u6d4b\u8bd5\u548c NCCL \u901a\u4fe1\u6d4b\u8bd5\uff09\uff0c\u8bc6\u522b\u5e76\u9694\u79bb\u6545\u969c\u8282\u70b9\uff0c\u7531 Kubernetes \u8865\u5145\u5065\u5eb7\u8282\u70b9\uff0c\u6700\u540e\u4ece\u6700\u8fd1\u7684\u68c0\u67e5\u70b9\u6062\u590d\u8bad\u7ec3\u3002<\/p>\n\n\n\n<p>\u5728\u68c0\u67e5\u70b9\u4f18\u5316\u65b9\u9762\uff0cMegaScale \u91c7\u7528\u4e24\u9636\u6bb5\u65b9\u6cd5\u3002\u7b2c\u4e00\u9636\u6bb5\uff0c\u6bcf\u5757 GPU \u5c06\u72b6\u6001\u5199\u5165\u4e3b\u673a\u5185\u5b58\uff08\u5229\u7528 PCIe \u9ad8\u5e26\u5bbd\u4ec5\u9700\u51e0\u79d2\uff09\uff0c\u7136\u540e\u7acb\u5373\u7ee7\u7eed\u8bad\u7ec3\u3002\u7b2c\u4e8c\u9636\u6bb5\uff0c\u540e\u53f0\u8fdb\u7a0b\u5f02\u6b65\u5c06\u72b6\u6001\u4f20\u8f93\u5230\u5206\u5e03\u5f0f\u6587\u4ef6\u7cfb\u7edf\u3002\u6062\u590d\u65f6\uff0c\u5229\u7528\u540c\u4e00\u6570\u636e\u5e76\u884c\u7ec4\u5171\u4eab\u72b6\u6001\u5206\u7247\u7684\u7279\u70b9\uff0c\u4ec5\u7531\u4e00\u4e2a worker \u4ece HDFS \u8bfb\u53d6\u518d\u5e7f\u64ad\u7ed9\u7ec4\u5185\u5176\u4ed6\u6210\u5458\uff0c\u7ebf\u6027\u964d\u4f4e\u4e86 HDFS \u5e26\u5bbd\u538b\u529b\u3002<\/p>\n\n\n\n<p>\u5bf9\u4e8e\u96be\u4ee5\u901a\u8fc7\u81ea\u68c0\u53d1\u73b0\u7684\u6027\u80fd\u95ee\u9898\uff0cMegaScale \u5f00\u53d1\u4e86\u57fa\u4e8e CUDA Event \u7684\u6027\u80fd\u5206\u6790\u5de5\u5177\u3002\u8be5\u5de5\u5177\u4ee5\u6781\u4f4e\u5f00\u9500\uff08\u65e0\u9700 CUDA \u540c\u6b65\uff09\u8bb0\u5f55\u6bcf\u4e2a\u8282\u70b9\u4e0a\u5173\u952e\u4ee3\u7801\u6bb5\u7684\u6267\u884c\u65f6\u95f4\uff0c\u5e76\u63d0\u4f9b\u4e24\u79cd\u53ef\u89c6\u5316\u6a21\u5f0f\uff1a\u70ed\u529b\u56fe\u6a21\u5f0f\u53ef\u4ee5\u4e00\u773c\u8bc6\u522b\u51fa\u8ba1\u7b97\u6389\u961f\u8282\u70b9\uff08\u7ea6\u5360\u96c6\u7fa4\u7684 0.5%\uff09\uff0c\u65f6\u95f4\u7ebf\u6a21\u5f0f\u53ef\u4ee5\u4ece\u5206\u5e03\u5f0f\u89c6\u89d2\u5c55\u793a\u6d41\u6c34\u7ebf\u6267\u884c\u7684\u5b9e\u9645\u65f6\u5e8f\u548c\u6570\u636e\u4f9d\u8d56\u5173\u7cfb\u3002\u6b64\u5916\uff0c3D \u5e76\u884c\u8bad\u7ec3\u53ef\u89c6\u5316\u5de5\u5177\u53ef\u4ee5\u5728\u901a\u4fe1\u8d85\u65f6\u65f6\u5c55\u793a\u5404 GPU \u7684\u903b\u8f91\u62d3\u6251\u4f4d\u7f6e\u548c\u6570\u636e\u6d41\u65b9\u5411\uff0c\u5e2e\u52a9\u5feb\u901f\u5b9a\u4f4d\u963b\u585e\u6e90\u5934\u3002<\/p>\n\n\n\n<p>\u56db\u3001\u5b9e\u9a8c\u8bc4\u4f30\uff1a\u5728 12288 \u5757 GPU \u4e0a\u8fbe\u5230 55.2% MFU<\/p>\n\n\n\n<p>MegaScale \u5728 ByteDance \u5185\u90e8\u90e8\u7f72\u7684\u8d85\u8fc7\u4e00\u4e07\u5757 NVIDIA Ampere GPU \u7684 AI \u96c6\u7fa4\u4e0a\u8fdb\u884c\u4e86\u5168\u9762\u8bc4\u4f30\u3002<\/p>\n\n\n\n<p>\u5728\u8bad\u7ec3\u6548\u7387\u65b9\u9762\uff0cMegaScale \u5728 12288 \u5757 GPU \u4e0a\u8bad\u7ec3 1750 \u4ebf\u53c2\u6570\u6a21\u578b\u65f6\u8fbe\u5230\u4e86 55.2% \u7684 MFU\uff0c\u76f8\u6bd4 Megatron-LM \u7684 41.2% \u63d0\u5347\u4e86 1.34 \u500d\u3002\u6d88\u878d\u5b9e\u9a8c\u8868\u660e\uff0c\u5404\u9879\u4f18\u5316\u7684\u8d21\u732e\u5206\u522b\u4e3a\uff1a\u5e76\u884c Transformer \u5757\u548c\u6ed1\u52a8\u7a97\u53e3\u6ce8\u610f\u529b\u8d21\u732e 5.6%\uff0c3D \u5e76\u884c\u901a\u4fe1\u91cd\u53e0\u8d21\u732e 6.2%\uff0c\u9ad8\u6548\u7b97\u5b50\u8d21\u732e 1.7%\uff0c\u6570\u636e\u7ba1\u9053\u7b49\u6742\u9879\u4f18\u5316\u8d21\u732e 1.1%\uff0cLAMB \u4f18\u5316\u5668\u6269\u5927\u6279\u91cf\u8d21\u732e 3.0%\u3002\u5728\u5f31\u6269\u5c55\u6027\u5b9e\u9a8c\u4e2d\uff0cMegaScale \u5728\u4ece 2240 \u5230 11200 \u5757 GPU \u7684\u8303\u56f4\u5185\u4fdd\u6301\u4e86\u8fd1\u7ebf\u6027\u7684\u6269\u5c55\u6027\u80fd\uff0cMFU \u7a33\u5b9a\u5728 54% \u5de6\u53f3\u3002<\/p>\n\n\n\n<p>\u5728\u7a33\u5b9a\u6027\u65b9\u9762\uff0c\u8bba\u6587\u5c55\u793a\u4e86\u4e00\u6b21\u771f\u5b9e\u7684\u751f\u4ea7\u8bad\u7ec3\u8fc7\u7a0b\uff1a\u5728\u8d85\u8fc7\u4e00\u4e07\u5757 GPU \u4e0a\u8bad\u7ec3\u4e00\u4e2a\u6570\u5343\u4ebf\u53c2\u6570\u7684\u6a21\u578b\uff0c\u5386\u65f6\u6570\u5468\uff0closs \u6301\u7eed\u6536\u655b\u3002\u671f\u95f4\u8bad\u7ec3\u91cd\u542f\u8d85\u8fc7 100 \u6b21\uff0c\u5176\u4e2d 90% \u4ee5\u4e0a\u7684\u6545\u969c\u7531\u5bb9\u9519\u6846\u67b6\u81ea\u52a8\u8bc6\u522b\u548c\u4fee\u590d\uff0c\u4ece\u6545\u969c\u68c0\u6d4b\u5230\u6062\u590d\u7684\u5e73\u5747\u65f6\u95f4\u4e0d\u8d85\u8fc7 10 \u5206\u949f\uff0c\u6709\u6548\u8bad\u7ec3\u65f6\u95f4\u7387\u4fdd\u6301\u5728 90% \u4ee5\u4e0a\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"929\" height=\"463\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/image-35.png\"  class=\"wp-image-1179\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/image-35.png 929w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/image-35-300x150.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/03\/image-35-768x383.png 768w\" sizes=\"auto, (max-width: 929px) 100vw, 929px\" title=\"MegaScale: Scaling Large Language Model Training to More Than 10,000 GPUs\u63d2\u56fe2\" alt=\"MegaScale: Scaling Large Language Model Training to More Than 10,000 GPUs\u63d2\u56fe2\" \/><\/figure>\n\n\n\n<p>\u4e94\u3001\u7ecf\u9a8c\u6559\u8bad\uff1a\u5927\u89c4\u6a21\u8bad\u7ec3\u4e2d\u7684\u5178\u578b\u95ee\u9898<\/p>\n\n\n\n<p>\u8bba\u6587\u5206\u4eab\u4e86\u82e5\u5e72\u6709\u4ef7\u503c\u7684\u5b9e\u6218\u7ecf\u9a8c\u3002\u5173\u4e8e\u8ba1\u7b97\u6389\u961f\u8282\u70b9\uff0c\u901a\u8fc7 CUDA Event \u70ed\u529b\u56fe\u53d1\u73b0\u7ea6 0.5% \u7684\u673a\u5668\u5728\u6267\u884c\u76f8\u540c\u524d\u5411\u8ba1\u7b97\u65f6\u8017\u65f6\u591a\u51fa\u7ea6 10%\uff0c\u6392\u9664\u8fd9\u4e9b\u8282\u70b9\u540e MFU \u63d0\u5347\u4e86 0.7%\u3002\u5173\u4e8e MFU \u9010\u6b65\u4e0b\u964d\u7684\u95ee\u9898\uff0c\u7ecf\u6392\u67e5\u53d1\u73b0\u6839\u56e0\u662f\u4e0d\u89c4\u5219\u7684\u5783\u573e\u56de\u6536\u548c\u67d0\u4e9b PyTorch \u64cd\u4f5c\u5f15\u5165\u7684\u6027\u80fd\u6ce2\u52a8\uff0c\u5bfc\u81f4\u5404 rank \u53d1\u8d77 ReduceScatter \u7684\u65f6\u95f4\u9010\u6e10\u504f\u79fb\u3001\u8d8a\u6765\u8d8a\u4e0d\u540c\u6b65\uff0c\u4fee\u590d\u540e MFU \u4e0d\u518d\u968f\u6b65\u6570\u4e0b\u964d\u3002\u5173\u4e8e\u7f51\u7edc\u63a5\u53e3\u6296\u52a8\u95ee\u9898\uff0c\u9700\u8981\u5c06 NCCL \u8d85\u65f6\u9608\u503c\u663e\u5f0f\u8bbe\u7f6e\u4e3a\u8f83\u5927\u503c\u4ee5\u907f\u514d\u7f51\u5361\u77ed\u6682\u65ad\u5f00\u65f6\u8bad\u7ec3\u76f4\u63a5\u5d29\u6e83\uff0c\u540c\u65f6\u4ece\u4fe1\u53f7\u8d28\u91cf\u5c42\u9762\u5bf9\u7f51\u5361\u3001AOC \u7ebf\u7f06\u548c\u4ea4\u6362\u673a\u8fdb\u884c\u5e95\u5c42\u8d28\u91cf\u7ba1\u63a7\u3002<\/p>\n\n\n\n<p>\u516d\u3001\u603b\u7ed3\u4e0e\u610f\u4e49<\/p>\n\n\n\n<p>MegaScale \u662f\u76ee\u524d\u516c\u5f00\u53d1\u8868\u7684\u6700\u5927\u89c4\u6a21 LLM \u8bad\u7ec3\u7cfb\u7edf\u4e4b\u4e00\uff0c\u5b83\u4ece\u7cfb\u7edf\u89d2\u5ea6\u5c55\u793a\u4e86\u4e07\u5361\u7ea7\u8bad\u7ec3\u7684\u5b8c\u6574\u8bbe\u8ba1\u7a7a\u95f4\u3002\u5176\u6838\u5fc3\u8d21\u732e\u4e0d\u4ec5\u5728\u4e8e 55.2% MFU \u8fd9\u4e00\u6570\u5b57\u672c\u8eab\uff0c\u66f4\u5728\u4e8e\u5b83\u63ed\u793a\u4e86\u5927\u89c4\u6a21\u8bad\u7ec3\u4e2d\u901a\u4fe1\u91cd\u53e0\u3001\u5bb9\u9519\u6062\u590d\u548c\u6027\u80fd\u8bca\u65ad\u8fd9\u4e09\u4e2a\u7ef4\u5ea6\u7684\u7cfb\u7edf\u6027\u65b9\u6cd5\u8bba\u3002\u5bf9\u4e8e\u6784\u5efa\u4e0b\u4e00\u4ee3 AI \u57fa\u7840\u8bbe\u65bd\u2014\u2014\u65e0\u8bba\u662f\u66f4\u5927\u89c4\u6a21\u7684\u5bc6\u96c6\u6a21\u578b\u8bad\u7ec3\u8fd8\u662f MoE \u6a21\u578b\u8bad\u7ec3\u2014\u2014MegaScale \u7684\u5168\u6808\u534f\u540c\u8bbe\u8ba1\u601d\u8def\u548c\u6df1\u5ea6\u53ef\u89c2\u6d4b\u6027\u539f\u5219\u90fd\u63d0\u4f9b\u4e86\u6781\u5177\u4ef7\u503c\u7684\u53c2\u8003\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>NSDI &#8217;24: 21st USENIX Symposium on Networked Systems Design and Implementation, April 16\u201318, 2024, Santa Clara, CA, USA https:\/\/www.usenix.org\/system\/files\/nsdi24-jiang-ziheng.pdf \u4e00\u3001\u7814\u7a76\u80cc\u666f\u4e0e\u52a8\u673a\uff1a\u4e07\u5361\u8bad\u7ec3\u65f6\u4ee3\u7684\u7cfb\u7edf\u6311\u6218 \u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u7684\u8bad\u7ec3\u89c4\u6a21\u5df2\u4ece\u6570\u767e\u5757 GPU \u8dc3\u5347\u81f3\u6570\u4e07\u5757  &hellip; <a href=\"https:\/\/www.ndnlab.com\/?p=1171\">\u7ee7\u7eed\u9605\u8bfb <span class=\"meta-nav\">&rarr;<\/span><\/a><\/p>\n","protected":false},"author":6,"featured_media":1178,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[23,6],"tags":[],"class_list":["post-1171","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-23","category-weilaiwangluo"],"_links":{"self":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/1171","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/users\/6"}],"replies":[{"embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=1171"}],"version-history":[{"count":2,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/1171\/revisions"}],"predecessor-version":[{"id":1180,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/1171\/revisions\/1180"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/media\/1178"}],"wp:attachment":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=1171"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=1171"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=1171"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}