{"id":1155617,"date":"2025-11-13T02:16:13","date_gmt":"2025-11-13T10:16:13","guid":{"rendered":"https:\/\/www.microsoft.com\/en-us\/research\/?post_type=msr-blog-post&p=1155617"},"modified":"2025-11-13T02:16:36","modified_gmt":"2025-11-13T10:16:36","slug":"waferllm","status":"publish","type":"msr-blog-post","link":"https:\/\/www.microsoft.com\/en-us\/research\/articles\/waferllm\/","title":{"rendered":"WaferLLM\uff1a\u4ece\u7cfb\u7edf\u8f6f\u4ef6\u89c6\u89d2\u91cd\u6784\u6676\u5706\u7ea7AI\u8ba1\u7b97"},"content":{"rendered":"\n

\u7f16\u8005\u6309\uff1a\u5f53\u201c\u66f4\u5927\u3001\u66f4\u5feb\u3001\u66f4\u9ad8\u6548\u201d\u7684 AI \u8ba1\u7b97\u6210\u4e3a\u4e1a\u754c\u8ffd\u6c42\u7684\u65b9\u5411\uff0c\u5728\u786c\u4ef6\u9769\u65b0\u4e4b\u5916\uff0c\u7cfb\u7edf\u8f6f\u4ef6\u5c42\u7684\u7a81\u7834\u540c\u6837\u5173\u952e\u3002\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u4e0e\u7231\u4e01\u5821\u5927\u5b66\u8054\u5408\u63d0\u51fa\u7684 WaferLLM \u6b63\u662f\u8fd9\u4e00\u63a2\u7d22\u7684\u91cd\u8981\u5c1d\u8bd5\u3002\u8be5\u7814\u7a76\u805a\u7126\u4e8e\u6676\u5706\u7ea7 AI \u8ba1\u7b97\u5e73\u53f0\u7684\u7cfb\u7edf\u8f6f\u4ef6\u4f18\u5316\uff0c\u4ece\u67b6\u6784\u539f\u7406\u5230\u63a8\u7406\u6027\u80fd\u5168\u9762\u91cd\u6784\uff0c\u4e3a AI \u8ba1\u7b97\u7684\u672a\u6765\u63d0\u4f9b\u65b0\u7684\u89c6\u89d2\u3002\u76f8\u5173\u8bba\u6587\u5df2\u88ab OSDI 2025 \u63a5\u6536\u3002<\/p>\n\n\n\n

\n\n\n\n

\u968f\u7740 AI \u6a21\u578b\u89c4\u6a21\u6301\u7eed\u81a8\u80c0\u3001\u8ba1\u7b97\u590d\u6742\u5ea6\u6025\u5267\u4e0a\u5347\uff0c\u4f20\u7edf\u82af\u7247\u67b6\u6784\u7684\u6027\u80fd\u4e0e\u80fd\u6548\u6b63\u5728\u903c\u8fd1\u6781\u9650\u3002\u8981\u7a81\u7834\u8fd9\u4e00\u74f6\u9888\uff0c\u4e1a\u754c\u5f00\u59cb\u63a2\u7d22\u66f4\u5927\u5c3a\u5ea6\u3001\u66f4\u7d27\u5bc6\u96c6\u6210\u7684\u8ba1\u7b97\u5f62\u6001\uff1a\u6676\u5706\u7ea7 AI \u8ba1\u7b97\uff08Wafer-Scale AI Compute\uff09\u3002\u5b83\u53ef\u5c06\u4e0a\u767e\u4e07\u4e2a\u8ba1\u7b97\u6838\u5fc3\u4e0e\u6d77\u91cf\u7247\u4e0a\u5b58\u50a8\u6574\u5408\u5728\u4e00\u6574\u7247\u7845\u6676\u5706\u4e0a\uff0c\u8ba9\u6570\u636e\u5728\u6676\u5706\u5185\u90e8\u6d41\u52a8\uff0c\u5927\u5e45\u51cf\u5c11\u8de8\u82af\u7247\u901a\u4fe1\u7684\u6210\u672c\u3002<\/p>\n\n\n\n

\u4f46\u786c\u4ef6\u521b\u65b0\u53ea\u662f\u5f00\u59cb\u3002\u8981\u60f3\u771f\u6b63\u91ca\u653e\u6676\u5706\u7ea7 AI \u82af\u7247\u7684\u6f5c\u529b\uff0c\u7cfb\u7edf\u8f6f\u4ef6\u5fc5\u987b\u540c\u6b65\u91cd\u6784\u3002\u4e3a\u6b64\uff0c\u6765\u81ea\u7231\u4e01\u5821\u5927\u5b66\u548c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86\u9762\u5411\u6676\u5706\u7ea7\u5e73\u53f0\u7684\u5927\u8bed\u8a00\u6a21\u578b\uff08LLMs\uff09\u63a8\u7406\u7cfb\u7edf WaferLLM<\/strong>\u3002\u5b83\u5b9e\u73b0\u4e86\u6bcf token \u4e9a\u6beb\u79d2\u7ea7\uff08<1ms\uff09\u5ef6\u8fdf<\/strong>\uff0c\u5c55\u793a\u4e86\u6676\u5706\u7ea7\u8ba1\u7b97\u5728 AI scaling \u4e0a\u7684\u5de8\u5927\u6f5c\u529b\u3002\u76f8\u5173\u8bba\u6587\u5df2\u88ab OSDI 2025 \u63a5\u6536\u3002<\/p>\n\n\n\n

WaferLLM: Large Language Model Inference at Wafer Scale<\/strong><\/p>\n\n\n\n

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

GitHub \u94fe\u63a5\uff1ahttps:\/\/github.com\/MeshInfra\/WaferLLM (opens in new tab)<\/span><\/a><\/p>\n\n\n\n

\u4eceAI\u89c4\u6a21\u6cd5\u5219\u5230\u6676\u5706\u7ea7\u8ba1\u7b97<\/h2>\n\n\n\n

\u5728\u8fc7\u53bb\u51e0\u5e74\u4e2d\uff0cAI \u89c4\u6a21\u6cd5\u5219\uff08Scaling Laws\uff09\u88ab\u591a\u6b21\u9a8c\u8bc1\uff1a\u66f4\u5927\u7684\u6a21\u578b\u548c\u66f4\u5f3a\u7684\u7b97\u529b\u5f80\u5f80\u80fd\u591f\u5e26\u6765\u66f4\u9ad8\u7684\u7cbe\u5ea6\u4e0e\u201c\u6d8c\u73b0\u80fd\u529b\u201d\u3002\u8fd9\u8ba9\u4e1a\u754c\u4e0d\u65ad\u8ffd\u6c42\u66f4\u5927\u89c4\u6a21\u7684\u8ba1\u7b97\u5e73\u53f0\u3002<\/p>\n\n\n\n

\u7136\u800c\uff0c\u591a GPU \u96c6\u7fa4\u5728\u6269\u5c55\u65f6\u4f1a\u906d\u9047\u901a\u4fe1\u74f6\u9888\uff0c\u8de8\u82af\u7247\u6570\u636e\u4f20\u8f93\u7684\u5ef6\u8fdf\u548c\u80fd\u8017\u6781\u9ad8\uff0c\u62b5\u6d88\u4e86\u7b97\u529b\u6269\u5c55\u6240\u5e26\u6765\u7684\u6536\u76ca\u3002\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e9b\u95ee\u9898\uff0c\u786c\u4ef6\u5382\u5546\u5f00\u59cb\u63a2\u7d22\u201c\u6676\u5706\u7ea7\u96c6\u6210\uff08System-on-Wafer, SoW\uff09\u201d\u6280\u672f\uff0c\u4e5f\u5c31\u662f\u5728\u540c\u4e00\u7247\u6676\u5706\u4e0a\u96c6\u6210\u66f4\u591a\u6838\u5fc3\u4e0e\u5185\u5b58\uff0c\u4f7f\u6570\u636e\u80fd\u591f\u5728\u7247\u4e0a\u5c40\u90e8\u4f20\u8f93\u3002<\/p>\n\n\n\n

\u8fd1\u5e74\u7684\u534a\u5bfc\u4f53\u5de5\u827a\u8fdb\u5c55\u8ba9\u8fd9\u4e00\u613f\u666f\u53d8\u5f97\u53ef\u884c\u3002\u5f97\u76ca\u4e8e\u5148\u8fdb\u7684\u5c01\u88c5\u3001\u66f4\u5c0f\u7684\u5236\u7a0b\u8282\u70b9\u4ee5\u53ca\u9ad8\u6548\u7684\u6563\u70ed\u6280\u672f\uff0c\u6676\u5706\u7ea7\u67b6\u6784\u80fd\u591f\u5b89\u5168\u7a33\u5b9a\u5730\u5de5\u4f5c\u3002\u4ee5\u53f0\u79ef\u7535\uff08TSMC\uff09\u7684 SoW \u6280\u672f\u4e3a\u4f8b\uff0c\u5176\u53ef\u5c06\u82af\u7247\u9762\u79ef\u6269\u5c55\u81f3\u6574\u7247\u6676\u5706\uff0c\u662f\u4f20\u7edf\u82af\u7247\u7684\u6570\u5341\u500d\u3002\u7247\u4e0a\u901a\u4fe1\u7684\u80fd\u6548\u6bd4\u677f\u7ea7\u4e92\u8fde\u9ad8\u51fa10\u5230100\u500d\uff0c\u66f4\u9002\u5408\u9ad8\u5e26\u5bbd\u3001\u4f4e\u5ef6\u8fdf\u7684 AI \u63a8\u7406\u4efb\u52a1\u3002<\/p>\n\n\n\n

\u76ee\u524d\uff0cNVIDIA\u3001AMD \u7b49\u516c\u53f8\u90fd\u5728\u63a8\u52a8\u66f4\u5927\u5c3a\u5bf8\u7684\u82af\u7247\u548c\u6676\u5706\u7ea7\u65b9\u6848\u3002\u4e1a\u754c\u9884\u6d4b\uff0c\u6676\u5706\u7ea7\u8ba1\u7b97\u5c06\u6210\u4e3a\u4e0b\u4e00\u4ee3 AI \u57fa\u7840\u8bbe\u65bd\u7684\u91cd\u8981\u65b9\u5411\u3002\u4e0e\u6b64\u540c\u65f6\uff0c\u5f88\u591a\u4f01\u4e1a\u4e5f\u5f00\u59cb\u5c1d\u8bd5\u5728\u6676\u5706\u7ea7\u7cfb\u7edf\u4e0a\u8fd0\u884c\u5927\u8bed\u8a00\u6a21\u578b\u3002<\/p>\n\n\n\n

\"table\"
\u88681\uff1aSystem-on-Die\uff08\u82af\u7247\u7ea7\u96c6\u6210\uff09 vs. System-on-Wafer\uff08\u6676\u5706\u7ea7\u96c6\u6210\uff09<\/figcaption><\/figure>\n\n\n\n

\u6676\u5706\u7ea7\u5e26\u6765\u7684\u8f6f\u4ef6\u6311\u6218\u4e0ePLMR\u6a21\u578b<\/h2>\n\n\n\n

\u6676\u5706\u7ea7\u7cfb\u7edf\u5728\u89c4\u6a21\u4e0e\u7ed3\u6784\u4e0a\u4e0e\u4f20\u7edf GPU \u5b8c\u5168\u4e0d\u540c\u3002\u6676\u5706\u7ea7\u5e73\u53f0\u867d\u7136\u5728\u7269\u7406\u7ed3\u6784\u4e0a\u7a81\u7834\u4e86\u9650\u5236\uff0c\u4f46\u5728\u7cfb\u7edf\u8f6f\u4ef6\u5c42\u9762\u5e26\u6765\u4e86\u5168\u65b0\u7684\u590d\u6742\u6027\u3002<\/p>\n\n\n\n

\"diagram\"
\u56fe1\uff1a\u6676\u5706\u7ea7 AI \u82af\u7247\u793a\u4f8b<\/figcaption><\/figure>\n\n\n\n

\u4e3a\u4e86\u7cfb\u7edf\u6027\u523b\u753b\u8fd9\u4e9b\u7279\u5f81\uff0c\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86 PLMR \u6a21\u578b\uff0c\u4ece\u56db\u4e2a\u7ef4\u5ea6\u603b\u7ed3\u4e86\u6676\u5706\u7ea7 AI \u82af\u7247\u7684\u8f6f\u4ef6\u6311\u6218\uff1a<\/p>\n\n\n\n

\"table\"<\/figure>\n\n\n\n

\u82e5\u8981\u5728\u6676\u5706\u7ea7\u5e73\u53f0\u4e0a\u7f16\u5199\u9ad8\u6027\u80fd AI \u8f6f\u4ef6\uff0c\u5fc5\u987b\u9075\u5b88\u201cPLMR \u539f\u5219\u201d\uff1a<\/p>\n\n\n\n

    \n
  • \u8ba1\u7b97\u4e0e\u6570\u636e\u5212\u5206\u5e94\u517c\u987e P\u3001M\uff1a\u5408\u7406\u5206\u5e03\u8d1f\u8f7d\u3001\u907f\u514d\u8d8a\u754c\u3002<\/li>\n\n\n\n
  • \u901a\u4fe1\u8bbe\u8ba1\u8981\u5c0a\u91cd L\uff1a\u51cf\u5c11\u8fdc\u8ddd\u79bb\u4f20\u8f93\u3001\u4f18\u5148\u5c40\u90e8\u8bbf\u95ee\u3002<\/li>\n\n\n\n
  • \u6d88\u606f\u8def\u7531\u4e0e\u90bb\u57df\u901a\u4fe1\u8981\u7b26\u5408 R \u7684\u9650\u5236\uff1a\u964d\u4f4e\u591a\u8df3\u901a\u4fe1\u5e26\u6765\u7684\u5f00\u9500\u3002<\/li>\n<\/ul>\n\n\n\n

    \u76f4\u63a5\u5957\u7528\u4f20\u7edf\u7684 GPU \u96c6\u7fa4\u6846\u67b6\uff08\u5982 MPI\uff09\uff0c\u5219\u4f1a\u51fa\u73b0\u901a\u4fe1\u5f00\u9500\u8fc7\u5927\u3001\u5185\u5b58\u5931\u8861\u6216\u540c\u6b65\u5ef6\u8fdf\u7b49\u95ee\u9898\u3002<\/p>\n\n\n\n

    WaferLLM\uff1a\u6676\u5706\u7ea7\u5e73\u53f0\u4e0a\u7684\u9ad8\u901fLLMs\u63a8\u7406\u7cfb\u7edf<\/h2>\n\n\n\n

    \u57fa\u4e8e PLMR \u6a21\u578b\uff0c\u7814\u7a76\u5458\u4eec\u5f00\u53d1\u4e86 WaferLLM\u2014\u2014\u4e00\u4e2a\u4e3a\u6676\u5706\u7ea7\u5e73\u53f0\uff08\u5982 Cerebras WSE-2\uff09\u91cf\u8eab\u5b9a\u5236\u7684 LLMs \u63a8\u7406\u7cfb\u7edf\u3002\u5b83\u5728\u5e76\u884c\u7b56\u7565\u3001\u901a\u4fe1\u7b97\u5b50\u548c\u7f13\u5b58\u7ba1\u7406\u4e0a\u8fdb\u884c\u4e86\u5168\u9762\u91cd\u6784\uff0c\u5b9e\u73b0\u4e86\u6bcf token \u5c0f\u4e8e1\u6beb\u79d2\u7684\u5ef6\u8fdf\u3002<\/p>\n\n\n\n

    \u5728\u5b9e\u73b0\u6676\u5706\u7ea7 LLMs \u63a8\u7406\u7684\u8fc7\u7a0b\u4e2d\uff0cWaferLLM \u9488\u5bf9\u67b6\u6784\u7279\u70b9\u8fdb\u884c\u4e86\u7cfb\u7edf\u6027\u7684\u4f18\u5316\u8bbe\u8ba1\u3002\u9996\u5148\uff0c\u5728\u5e76\u884c\u7b56\u7565\u4e0a\uff0c\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86\u66f4\u9002\u914d\u6676\u5706\u7ea7\u67b6\u6784\u7684\u5206\u5e03\u65b9\u6848\u3002\u901a\u5e38\uff0c\u5927\u8bed\u8a00\u6a21\u578b\u7684\u63a8\u7406\u53ef\u5206\u4e3a\u4e24\u4e2a\u9636\u6bb5\uff1aPrefill \u9636\u6bb5\u8d1f\u8d23\u52a0\u8f7d\u6a21\u578b\u6743\u91cd\u5e76\u6267\u884c\u5927\u89c4\u6a21\u77e9\u9635\u4e58\u6cd5\uff08GEMM\uff09\uff0c\u800c Decode \u9636\u6bb5\u5219\u5728\u751f\u6210\u6bcf\u4e2a token \u65f6\u6267\u884c\u77e9\u9635\u5411\u91cf\u4e58\u6cd5\uff08GEMV\uff09\u5e76\u8bbf\u95ee KV \u7f13\u5b58\u3002<\/p>\n\n\n\n

    WaferLLM \u9488\u5bf9\u8fd9\u4e24\u4e2a\u9636\u6bb5\u91c7\u7528\u4e86\u4e0d\u540c\u7684\u4f18\u5316\u7b56\u7565\u3002\u5728 Prefill \u9636\u6bb5\uff0c\u5bf9\u6fc0\u6d3b\u4e0e\u6743\u91cd\u77e9\u9635\u8fdb\u884c\u4e8c\u7ef4\u5212\u5206\uff1b\u5728 Decode \u9636\u6bb5\uff0c\u4f7f\u7528\u590d\u5236\u4e0e\u5207\u5206\u76f8\u7ed3\u5408\u7684\u65b9\u5f0f\uff0c\u4f7f\u8ba1\u7b97\u4e0e\u6570\u636e\u66f4\u5747\u5300\u5730\u5206\u5e03\u5728\u5404\u4e2a\u6838\u5fc3\u4e0a\u3002\u6b64\u5916\uff0c\u7cfb\u7edf\u8fd8\u5728\u9884\u5904\u7406\u9636\u6bb5\u5b8c\u6210\u4e86\u77e9\u9635\u8f6c\u7f6e\uff0c\u4ee5\u907f\u514d\u8fd0\u884c\u65f6\u7684\u9ad8\u5ef6\u8fdf\u901a\u4fe1\u3002<\/p>\n\n\n\n

    \u9488\u5bf9\u6676\u5706\u7ea7\u82af\u7247\u72ec\u7279\u7684\u7f51\u683c\u4e92\u8fde\u7ed3\u6784\uff0cWaferLLM \u8bbe\u8ba1\u4e86\u4e24\u79cd\u9ad8\u6548\u7b97\u5b50\uff1a<\/p>\n\n\n\n

      \n
    • MeshGEMM<\/strong>\uff1a\u901a\u8fc7\u5faa\u73af\u79fb\u4f4d (cyclic-shifting) \u4e0e\u4ea4\u9519 (interleaving) \u7b56\u7565\u7f29\u77ed\u901a\u4fe1\u8def\u5f84\uff0c\u4f7f\u901a\u4fe1\u5f00\u9500\u964d\u4f4e2\u81f35\u500d\uff0c\u6574\u4f53\u6027\u80fd\u63d0\u53471.3\u81f32\u500d\u3002<\/li>\n\n\n\n
    • MeshGEMV<\/strong>\uff1a\u7528 K-tree Allreduce \u673a\u5236\u7f29\u77ed\u5173\u952e\u901a\u4fe1\u8def\u5f84\uff0c\u4f7f\u901a\u4fe1\u5f00\u9500\u51cf\u5c111.5\u81f36\u500d\uff0c\u6574\u4f53\u52a0\u901f\u53ef\u8fbe2.5\u81f33\u500d\u3002<\/li>\n<\/ul>\n\n\n\n

      \u5728\u7f13\u5b58\u7ba1\u7406\u65b9\u9762\uff0cWaferLLM \u63d0\u51fa\u4e86\u57fa\u4e8e\u79fb\u4f4d\uff08shift\uff09\u7684 KV \u7f13\u5b58\u91cd\u5206\u5e03\u65b9\u6cd5\u3002\u8be5\u7b56\u7565\u5141\u8bb8\u76f8\u90bb\u6838\u5fc3\u5e76\u884c\u8c03\u6574\u7f13\u5b58\u5e03\u5c40\uff0c\u4ece\u800c\u5e73\u8861\u5185\u5b58\u4f7f\u7528\uff0c\u907f\u514d\u663e\u5b58\u6ea2\u51fa\uff08OOM\uff09\uff0c\u5e76\u652f\u6301\u66f4\u957f\u7684\u4e0a\u4e0b\u6587\u63a8\u7406\u3002<\/p>\n\n\n\n

      \u5728\u771f\u5b9e\u7684 Cerebras WSE-2 \u6676\u5706\u7ea7\u82af\u7247\u6d4b\u8bd5\u4e2d\uff0cWaferLLM \u5c55\u73b0\u51fa\u4e86\u663e\u8457\u7684\u4f18\u52bf\u3002\u5f53 SGLang \u6216 vLLM \u4ece\u5355\u5f20 A100 \u6269\u5c55\u52308\u5f20 A100 GPU \u65f6\uff0c\u89e3\u7801\u901f\u7387\u4ece78\u63d0\u5347\u81f3\u7ea6260 tokens\/s\uff0c\u867d\u7136\u63d0\u5347\u660e\u663e\uff0c\u4f46\u4ecd\u672a\u8fbe\u5230\u7406\u60f3\u76848\u500d\u6269\u5c55\u3002\u968f\u7740\u96c6\u7fa4\u89c4\u6a21\u8fdb\u4e00\u6b65\u6269\u5927\uff0c\u8de8\u8282\u70b9\u901a\u4fe1\u7684\u6210\u672c\u751a\u81f3\u8d85\u8fc7\u4e86\u8ba1\u7b97\u6536\u76ca\uff0c\u6574\u4f53\u6027\u80fd\u5f00\u59cb\u4e0b\u964d\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0cWaferLLM \u5728\u6676\u5706\u7ea7\u82af\u7247\u4e0a\u907f\u514d\u4e86\u8fd9\u79cd\u901a\u4fe1\u74f6\u9888\uff0c\u63a8\u7406\u901f\u5ea6\u8fbe\u52302700 tokens\/s\uff0c\u5373\u6bcf\u4e2a token \u7684\u5ef6\u8fdf\u4f4e\u4e8e1\u6beb\u79d2\u3002<\/p>\n\n\n\n

      \u5728\u5355\u5361\u89c4\u6a21\u4e0b\uff0cWaferLLM \u7684\u80fd\u91cf\u6548\u7387\u4e0e A100 \u76f8\u5f53\uff08\u80fd\u91cf\u6bd4\u7ea6\u4e3a0.92\uff09\u3002\u5f53\u7cfb\u7edf\u6269\u5c55\u81f3\u66f4\u5927\u89c4\u6a21\u65f6\uff0cWaferLLM \u7684\u80fd\u91cf\u6bd4\u63d0\u5347\u52302.22\uff0c\u8bf4\u660e\u6676\u5706\u7ea7\u67b6\u6784\u5728\u9ad8\u5e76\u884c\u5ea6\u573a\u666f\u4e0b\u80fd\u5b9e\u73b0\u66f4\u4f18\u7684\u6027\u80fd\u4e0e\u80fd\u8017\u5e73\u8861\u3002<\/p>\n\n\n\n

      \"table\"
      \u56fe2\uff1a\u6676\u5706\u7ea7\u7cfb\u7edf\u4e0e\u591a GPU \u7cfb\u7edf\u7684\u6269\u5c55\u6548\u7387\u6bd4\u8f83<\/figcaption><\/figure>\n\n\n\n

      \u603b\u4f53\u6765\u770b\uff0c\u6676\u5706\u7ea7\u8ba1\u7b97\u5728\u8d85\u5927\u89c4\u6a21\u3001\u4f4e\u5ef6\u8fdf\u63a8\u7406\u4efb\u52a1\u4e2d\u5c55\u73b0\u51fa\u5f3a\u52b2\u7684\u6f5c\u529b\uff0c\u7279\u522b\u9002\u5408\u9700\u8981\u5feb\u901f\u751f\u6210\u548c\u9ad8\u5b9e\u65f6\u6027\u7684 test-time scaling \u573a\u666f\u3002<\/p>\n\n\n\n

      AI\u7cfb\u7edf\u5806\u6808\u7684\u201c\u6676\u5706\u7ea7\u91cd\u6784\u201d<\/h2>\n\n\n\n

      WaferLLM \u53ea\u662f\u4e00\u4e2a\u8d77\u70b9\u3002\u8981\u8ba9\u6676\u5706\u7ea7 AI \u8ba1\u7b97\u771f\u6b63\u8d70\u5411\u6210\u719f\uff0c\u8fd8\u9700\u8981\u591a\u4e2a\u5c42\u9762\u534f\u540c\u521b\u65b0\uff1a<\/p>\n\n\n\n

        \n
      • \u6a21\u578b\u67b6\u6784\u4e0e\u786c\u4ef6\u534f\u540c<\/strong><\/li>\n<\/ul>\n\n\n\n

        \u672a\u6765\u6a21\u578b\u7684\u8bbe\u8ba1\u5e94\u66f4\u7d27\u5bc6\u5730\u7ed3\u5408\u786c\u4ef6\u7279\u6027\uff0c\u5728\u8bad\u7ec3\u9636\u6bb5\u7eb3\u5165\u901a\u4fe1\u548c\u5206\u5e03\u7b56\u7565\u7684\u4f18\u5316\uff0c\u4ee5\u5145\u5206\u53d1\u6325\u6676\u5706\u7ea7\u67b6\u6784\u6240\u63d0\u4f9b\u7684\u9ad8\u5e76\u884c\u5ea6\u548c\u9ad8\u5e26\u5bbd\u4f18\u52bf\u3002<\/p>\n\n\n\n

          \n
        • \u7cfb\u7edf\u8f6f\u4ef6\u4e0e\u7f16\u7a0b\u6a21\u578b\u9769\u65b0<\/strong><\/li>\n<\/ul>\n\n\n\n

          \u4f20\u7edf MPI \u6216\u7ebf\u7a0b\u6a21\u578b\u96be\u4ee5\u9002\u5e94\u9762\u5411\u6570\u767e\u4e07\u6838\u7684\u5f02\u6b65\u5206\u5e03\u5f0f\u7cfb\u7edf\uff0c\u9700\u8981\u91cd\u65b0\u8bbe\u8ba1\u8fd0\u884c\u65f6\u3001\u8c03\u5ea6\u4e0e\u5185\u5b58\u7ba1\u7406\u673a\u5236\u3002\u4e00\u4e2a\u201c\u6676\u5706\u539f\u751f\uff08wafer-native\uff09\u201d\u7684\u7f16\u7a0b\u8303\u5f0f\u6216\u5c06\u6210\u4e3a\u7a81\u7834\u53e3\u3002<\/p>\n\n\n\n

            \n
          • \u786c\u4ef6\u62d3\u6251\u4e0e\u7f51\u7edc\u521b\u65b0<\/strong><\/li>\n<\/ul>\n\n\n\n

            \u5f53\u524d\u4e3b\u6d41\u7684\u4e8c\u7ef4\u7f51\u683c\u62d3\u6251\u5728\u90e8\u5206\u901a\u4fe1\u6a21\u5f0f\uff08\u5982 All-Reduce\uff09\u4e0b\u5e76\u4e0d\u7406\u60f3\u3002\u672a\u6765\u53ef\u80fd\u51fa\u73b0\u6df7\u5408\u6216\u53ef\u91cd\u6784\u7684 NoC \u7ed3\u6784\uff0c \u4ee5\u652f\u6301\u66f4\u7075\u6d3b\u7684\u6a21\u578b\u67b6\u6784\u548c\u901a\u8baf\u9700\u6c42\u3002<\/p>\n\n\n\n

            \u6676\u5706\u7ea7 AI \u8ba1\u7b97\u6b63\u5728\u91cd\u65b0\u5b9a\u4e49\u7b97\u529b\u7684\u8fb9\u754c\u3002\u5b83\u65e2\u662f\u4e00\u573a\u786c\u4ef6\u9769\u547d\uff0c\u4e5f\u662f\u7cfb\u7edf\u8f6f\u4ef6\u4e0e\u6a21\u578b\u534f\u540c\u8bbe\u8ba1\u7684\u8303\u5f0f\u8f6c\u53d8\u3002\u672a\u6765\uff0c\u8ba1\u7b97\u7684\u6781\u9650\u5c06\u4e0d\u518d\u7531\u82af\u7247\u9762\u79ef\u51b3\u5b9a\uff0c\u800c\u53d6\u51b3\u4e8e\u7cfb\u7edf\u4e0e\u667a\u80fd\u7684\u5171\u751f\u6f14\u5316\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

            \u7f16\u8005\u6309\uff1a\u5f53\u201c\u66f4\u5927\u3001\u66f4\u5feb\u3001\u66f4\u9ad8\u6548\u201d\u7684 AI \u8ba1\u7b97\u6210\u4e3a\u4e1a\u754c\u8ffd\u6c42\u7684\u65b9\u5411\uff0c\u5728\u786c\u4ef6\u9769\u65b0\u4e4b\u5916\uff0c\u7cfb\u7edf\u8f6f\u4ef6\u5c42\u7684\u7a81\u7834\u540c\u6837\u5173\u952e\u3002\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u4e0e\u7231\u4e01\u5821\u5927\u5b66\u8054\u5408\u63d0\u51fa\u7684 WaferLLM \u6b63\u662f\u8fd9\u4e00\u63a2\u7d22\u7684\u91cd\u8981\u5c1d\u8bd5\u3002\u8be5\u7814\u7a76\u805a\u7126\u4e8e\u6676\u5706\u7ea7 AI \u8ba1\u7b97\u5e73\u53f0\u7684\u7cfb\u7edf\u8f6f\u4ef6\u4f18\u5316\uff0c\u4ece\u67b6\u6784\u539f\u7406\u5230\u63a8\u7406\u6027\u80fd\u5168\u9762\u91cd\u6784\uff0c\u4e3a AI \u8ba1\u7b97\u7684\u672a\u6765\u63d0\u4f9b\u65b0\u7684\u89c6\u89d2\u3002\u76f8\u5173\u8bba\u6587\u5df2\u88ab OSDI 2025 \u63a5\u6536\u3002 \u968f\u7740 AI \u6a21\u578b\u89c4\u6a21\u6301\u7eed\u81a8\u80c0\u3001\u8ba1\u7b97\u590d\u6742\u5ea6\u6025\u5267\u4e0a\u5347\uff0c\u4f20\u7edf\u82af\u7247\u67b6\u6784\u7684\u6027\u80fd\u4e0e\u80fd\u6548\u6b63\u5728\u903c\u8fd1\u6781\u9650\u3002\u8981\u7a81\u7834\u8fd9\u4e00\u74f6\u9888\uff0c\u4e1a\u754c\u5f00\u59cb\u63a2\u7d22\u66f4\u5927\u5c3a\u5ea6\u3001\u66f4\u7d27\u5bc6\u96c6\u6210\u7684\u8ba1\u7b97\u5f62\u6001\uff1a\u6676\u5706\u7ea7 AI \u8ba1\u7b97\uff08Wafer-Scale AI Compute\uff09\u3002\u5b83\u53ef\u5c06\u4e0a\u767e\u4e07\u4e2a\u8ba1\u7b97\u6838\u5fc3\u4e0e\u6d77\u91cf\u7247\u4e0a\u5b58\u50a8\u6574\u5408\u5728\u4e00\u6574\u7247\u7845\u6676\u5706\u4e0a\uff0c\u8ba9\u6570\u636e\u5728\u6676\u5706\u5185\u90e8\u6d41\u52a8\uff0c\u5927\u5e45\u51cf\u5c11\u8de8\u82af\u7247\u901a\u4fe1\u7684\u6210\u672c\u3002 \u4f46\u786c\u4ef6\u521b\u65b0\u53ea\u662f\u5f00\u59cb\u3002\u8981\u60f3\u771f\u6b63\u91ca\u653e\u6676\u5706\u7ea7 AI \u82af\u7247\u7684\u6f5c\u529b\uff0c\u7cfb\u7edf\u8f6f\u4ef6\u5fc5\u987b\u540c\u6b65\u91cd\u6784\u3002\u4e3a\u6b64\uff0c\u6765\u81ea\u7231\u4e01\u5821\u5927\u5b66\u548c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86\u9762\u5411\u6676\u5706\u7ea7\u5e73\u53f0\u7684\u5927\u8bed\u8a00\u6a21\u578b\uff08LLMs\uff09\u63a8\u7406\u7cfb\u7edf WaferLLM\u3002\u5b83\u5b9e\u73b0\u4e86\u6bcf token \u4e9a\u6beb\u79d2\u7ea7\uff08<\/p>\n","protected":false},"author":34512,"featured_media":1155632,"template":"","meta":{"msr-url-field":"","msr-podcast-episode":"","msrModifiedDate":"","msrModifiedDateEnabled":false,"ep_exclude_from_search":false,"_classifai_error":"","msr-content-parent":1012650,"msr_hide_image_in_river":null,"footnotes":""},"research-area":[13556,13547],"msr-locale":[268881],"msr-post-option":[],"class_list":["post-1155617","msr-blog-post","type-msr-blog-post","status-publish","has-post-thumbnail","hentry","msr-research-area-artificial-intelligence","msr-research-area-systems-and-networking","msr-locale-zh_cn"],"msr_assoc_parent":{"id":1012650,"type":"lab"},"_links":{"self":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1155617","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post"}],"about":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/types\/msr-blog-post"}],"author":[{"embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/users\/34512"}],"version-history":[{"count":7,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1155617\/revisions"}],"predecessor-version":[{"id":1155634,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1155617\/revisions\/1155634"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/media\/1155632"}],"wp:attachment":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/media?parent=1155617"}],"wp:term":[{"taxonomy":"msr-research-area","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/research-area?post=1155617"},{"taxonomy":"msr-locale","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-locale?post=1155617"},{"taxonomy":"msr-post-option","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-post-option?post=1155617"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}