{"id":8460,"date":"2026-09-25T09:00:21","date_gmt":"2026-09-25T01:00:21","guid":{"rendered":"https:\/\/1ow.top\/index.php\/2026\/09\/25\/pcie%e6%98%be%e5%8d%a1%e8%a2%ab%e4%bd%8e%e4%bc%b0%e4%ba%86%ef%bc%81%e5%86%85%e6%a0%b8%e8%a1%a5%e9%bd%90%e9%80%9a%e4%bf%a1%e9%87%8d%e6%9e%84%ef%bc%8cdeepseek%e6%8e%a8%e7%90%86%e5%90%9e%e5%90%90\/"},"modified":"2026-09-25T09:00:21","modified_gmt":"2026-09-25T01:00:21","slug":"pcie%e6%98%be%e5%8d%a1%e8%a2%ab%e4%bd%8e%e4%bc%b0%e4%ba%86%ef%bc%81%e5%86%85%e6%a0%b8%e8%a1%a5%e9%bd%90%e9%80%9a%e4%bf%a1%e9%87%8d%e6%9e%84%ef%bc%8cdeepseek%e6%8e%a8%e7%90%86%e5%90%9e%e5%90%90","status":"publish","type":"post","link":"https:\/\/1ow.top\/index.php\/2026\/09\/25\/pcie%e6%98%be%e5%8d%a1%e8%a2%ab%e4%bd%8e%e4%bc%b0%e4%ba%86%ef%bc%81%e5%86%85%e6%a0%b8%e8%a1%a5%e9%bd%90%e9%80%9a%e4%bf%a1%e9%87%8d%e6%9e%84%ef%bc%8cdeepseek%e6%8e%a8%e7%90%86%e5%90%9e%e5%90%90\/","title":{"rendered":"PCIe\u663e\u5361\u88ab\u4f4e\u4f30\u4e86\uff01\u5185\u6838\u8865\u9f50+\u901a\u4fe1\u91cd\u6784\uff0cDeepSeek\u63a8\u7406\u541e\u5410\u7ffb\u8fd17\u500d"},"content":{"rendered":"
\"PCIe\u663e\u5361\u88ab\u4f4e\u4f30\u4e86\uff01\u5185\u6838\u8865\u9f50+\u901a\u4fe1\u91cd\u6784\uff0cDeepSeek\u63a8\u7406\u541e\u5410\u7ffb\u8fd17\u500d\"\/<\/div><\/figure>\n
\"PCIe\u663e\u5361\u88ab\u4f4e\u4f30\u4e86\uff01\u5185\u6838\u8865\u9f50+\u901a\u4fe1\u91cd\u6784\uff0cDeepSeek\u63a8\u7406\u541e\u5410\u7ffb\u8fd17\u500d\"\/<\/div><\/figure>\n
\"PCIe\u663e\u5361\u88ab\u4f4e\u4f30\u4e86\uff01\u5185\u6838\u8865\u9f50+\u901a\u4fe1\u91cd\u6784\uff0cDeepSeek\u63a8\u7406\u541e\u5410\u7ffb\u8fd17\u500d\"\/<\/div><\/figure>\n
\"PCIe\u663e\u5361\u88ab\u4f4e\u4f30\u4e86\uff01\u5185\u6838\u8865\u9f50+\u901a\u4fe1\u91cd\u6784\uff0cDeepSeek\u63a8\u7406\u541e\u5410\u7ffb\u8fd17\u500d\"\/<\/div><\/figure>\n

\u5141\u4e2d \u53d1\u81ea \u51f9\u975e\u5bfa<\/p>\n

\u968f\u7740\u5927\u6a21\u578b\u5bf9\u7b97\u529b\u9700\u6c42\u6301\u7eed\u6500\u5347\uff0cGPU\u5361\u7684\u91c7\u8d2d\u6210\u672c\u3001\u4f9b\u7ed9\u7ea6\u675f\u6301\u7eed\u52a0\u5267\u3002<\/p>\n

AI\u63a8\u7406\u7684\u7ade\u4e89\u6b63\u5728\u6084\u7136\u53d1\u751f\u8f6c\u5411\uff1a\u4e0d\u518d\u5355\u7eaf\u6bd4\u62fc\u201c\u8c01\u80fd\u591f\u4e70\u5230\u6027\u80fd\u6700\u9876\u7ea7\u7684\u786c\u4ef6\u201d\uff0c\u800c\u662f\u8f6c\u5411\u201c\u8c01\u53ef\u4ee5\u628a\u624b\u4e2d\u5df2\u6709\u7684\u7b97\u529b\u8d44\u6e90\u7528\u5f97\u66f4\u503c\u201d\u3002<\/p>\n

\u5f88\u591aGPU\u5361\uff0c\u5f00\u6e90\u6846\u67b6\u80fd\u591f\u5b8c\u6210\u6a21\u578b\u52a0\u8f7d\u3001\u6743\u91cd\u4e0b\u8f7d\u3001\u670d\u52a1\u62c9\u8d77\uff0c\u5b9e\u73b0\u201c\u80fd\u8dd1\u8d77\u6765\u201d\uff0c\u4f46\u5b9e\u9645\u538b\u6d4b\u6027\u80fd\u5f80\u5f80\u8fdc\u4f4e\u4e8e\u5b98\u65b9\u5ba3\u4f20\u6c34\u5e73\u3002\u6a21\u578b\u672c\u8eab\u6ca1\u6709\u7f3a\u9677\uff0c\u786c\u4ef6\u4e5f\u53ef\u6b63\u5e38\u5de5\u4f5c\uff0c\u6027\u80fd\u635f\u5931\u7684\u6838\u5fc3\uff0c\u6765\u81ea\u6a21\u578b\u6846\u67b6\u4e0e\u786c\u4ef6\u4e4b\u95f4\u5b58\u5728\u7684\u6027\u80fd\u9e3f\u6c9f\u3002<\/p>\n

\u8fd9\u7c7bGPU\u5361\u6ca1\u6709\u9ad8\u901f\u5361\u95f4\u4e92\u8054\uff0c\u540c\u65f6\u4e0d\u5728\u4e3b\u6d41\u5f00\u6e90\u6846\u67b6\u5b98\u65b9\u9a8c\u8bc1\u77e9\u9635\u4e2d\u3002\u76f4\u63a5\u4f7f\u7528\u793e\u533a\u9ed8\u8ba4\u90e8\u7f72\u65b9\u6848\uff0c\u4f1a\u51fa\u73b0\u7b97\u5b50\u81ea\u52a8\u56de\u9000\u81f3\u4f4e\u6548\u901a\u7528\u5b9e\u73b0\u3001\u96c6\u5408\u901a\u4fe1\u6cbf\u7528NVLink\u786c\u4ef6\u7684\u8c03\u4f18\u53c2\u6570\u3001\u663e\u5b58\u4e0e\u5e76\u884c\u914d\u7f6e\u6cbf\u7528\u5176\u4ed6\u786c\u4ef6\u9ed8\u8ba4\u503c\u7b49\u4e00\u7cfb\u5217\u95ee\u9898\u3002\u786c\u4ef6\u672c\u8eab\u5177\u5907\u7684\u7b97\u529b\u6f5c\u529b\uff0c\u88ab\u8f6f\u4ef6\u5c42\u7684\u9002\u914d\u77ed\u677f\u6240\u7981\u9522\u3002<\/p>\n

\u662f\u77f3\u79d1\u6280\uff08METASTONE\uff09\u662f\u4e00\u5bb6\u201c\u72ec\u7acb\u7b2c\u4e09\u65b9\u5168\u6808\u7b97\u529b\u8fd0\u8425\u5546\u201d\u7684\u79d1\u6280\u4f01\u4e1a\u3002\u4e0d\u7ed1\u5b9a\u4efb\u4f55\u7279\u5b9a\u5927\u6a21\u578b\u5382\u5546\uff0c\u81f4\u529b\u4e8e\u4e3a\u5ba2\u6237\u63d0\u4f9b\u6db5\u76d6\u786c\u4ef6\u3001\u7ec4\u7f51\u3001\u8c03\u5ea6\u3001\u4f18\u5316\u4e0e\u8fd0\u8425\u7684\u4e00\u7ad9\u5f0f\u5168\u6808\u7b97\u529b\u4ea4\u4ed8\u670d\u52a1\u3002<\/p>\n

\u51ed\u501f\u6e90\u81ea\u56fd\u5bb6\u7ea7\u8ba1\u7b97\u4e2d\u5fc3\u7684\u4e30\u5bcc\u7ecf\u9a8c\uff0c\u516c\u53f8\u786e\u4fdd\u4e86\u7b97\u529b\u96c6\u7fa4\u9ad8\u8fbe99.95%\u4ee5\u4e0a\u7684\u5b89\u5168\u7a33\u5b9a\u8fd0\u884c\uff08SLA\uff09\u3002\u76ee\u524d\uff0c\u662f\u77f3\u79d1\u6280\u5df2\u7eb3\u7ba1\u8d85\u8fc720000P\u7684\u7b97\u529b\u8d44\u6e90\uff0c\u8fd0\u842510\u591a\u4e2a\u667a\u7b97\u4e2d\u5fc3\uff0c\u62e5\u67092\u4e2a\u56fd\u5bb6\u7ea7\u8d85\u7b97\u4e2d\u5fc3\u3002\u56e2\u961f\u6210\u5458\u66fe\u8363\u83b73\u9879\u6208\u767b\u00b7\u8d1d\u5c14\u5956\uff0c\u5177\u5907\u6df1\u539a\u7684\u8d85\u5927\u89c4\u6a21\u96c6\u7fa4\u5168\u6808\u6280\u672f\u670d\u52a1\u5b9e\u529b\u7b49\u3002<\/p>\n

\u662f\u77f3\u79d1\u6280\u81ea\u7814Meta-Infer\u9ad8\u6548\u90e8\u7f72\u5f15\u64ce\uff0c\u662f\u9762\u5411\u5f02\u6784\u52a0\u901f\u82af\u7247\u7684\u4f01\u4e1a\u7ea7\u9ad8\u6027\u80fd\u5927\u6a21\u578b\u63a8\u7406\u5f15\u64ce\uff0c\u5c1d\u8bd5\u901a\u8fc7\u7eaf\u8f6f\u4ef6\u4f18\u5316\u624b\u6bb5\u5f25\u5408\u8fd9\u9053\u786c\u4ef6-\u6846\u67b6\u4e4b\u95f4\u7684\u7f1d\u9699\uff0c\u5728\u4e0d\u6539\u52a8\u6a21\u578b\u7ed3\u6784\u3001\u4e0d\u4fee\u6539\u4efb\u52a1\u8bed\u4e49\u7684\u524d\u63d0\u4e0b\uff0c\u5145\u5206\u6316\u6398GPU\u5361\u67b6\u6784\u786c\u4ef6\u7684\u63a8\u7406\u80fd\u529b\uff0c\u4e3a\u884c\u4e1a\u63d0\u4f9b\u4e00\u79cd\u53ef\u80fd\u6027\uff1a<\/p>\n

\u4f9d\u9760\u8f6f\u4ef6\u4f18\u5316\uff0c\u6210\u672c\u66f4\u4f4e\u7684GPU\u5361\uff0c\u6709\u673a\u4f1a\u5728\u90e8\u5206\u63a8\u7406\u4e1a\u52a1\u6307\u6807\u4e0a\uff0c\u903c\u8fd1\u539f\u672c\u9700\u8981\u66f4\u9ad8\u7aefGPU\u624d\u80fd\u5b9e\u73b0\u7684\u670d\u52a1\u80fd\u529b\u3002<\/p>\n

Meta-Infer\u9ad8\u6548\u63a8\u7406\u5f15\u64ce\u7684\u6574\u5957\u4f18\u5316\u903b\u8f91\uff0c\u5206\u4e3a\u4e24\u5927\u5173\u952e\u9636\u6bb5\uff1a<\/p>\n

\u6700\u7ec8\u5c06\u6c89\u6dc0\u4e3aMeta-Infer\u9ad8\u6548\u63a8\u7406\u5f15\u64ce\u7684\u56db\u9879\u6838\u5fc3\u80fd\u529b\uff1a\u5185\u6838\u8865\u9f50\u3001\u7b97\u5b50\u4f18\u5316\u4e0e\u901a\u4fe1\u91cd\u6784\u3001\u5e76\u884c\u4e0e\u5bb9\u91cf\u8c03\u4f18\u3001\u7f13\u5b58\u590d\u7528\u3002<\/p>\n

\u4e3b\u6d41\u63a8\u7406\u6846\u67b6\u5185\u7f6e\u4e86\u5927\u91cf\u9ad8\u6027\u80fd\u7b97\u5b50\uff0c\u4f46\u5bf9\u4e8e\u4e0d\u5728\u5b98\u65b9\u9a8c\u8bc1\u6e05\u5355\u7684\u957f\u5c3e\u786c\u4ef6\uff0c\u6846\u67b6\u4e0d\u4f1a\u62a5\u9519\uff0c\u53ea\u4f1a\u9759\u9ed8\u7ed5\u5f00\u52a0\u901f\u8def\u5f84\uff0c\u964d\u7ea7\u8fd0\u884c\u4f4e\u6548\u901a\u7528\u903b\u8f91\u3002<\/p>\n

\u5f88\u591a\u9ad8\u6027\u80fd\u5185\u6838\u5e76\u975e\u7f3a\u5931\uff0c\u53ea\u662f\u5143\u6570\u636e\u3001\u9875\u5c3a\u5bf8\u3001\u786c\u4ef6\u5185\u6838\u9002\u914d\u4e0d\u5339\u914d\uff0c\u5bfc\u81f4\u65e0\u6cd5\u88ab\u89e6\u53d1\u6267\u884c\u3002<\/p>\n

Meta-Infer\u9ad8\u6548\u63a8\u7406\u5f15\u64ce\uff0c\u901a\u8fc7\u5185\u6838\u8865\u9f50\u5b8c\u6210\u9002\u914d\u4fee\u590d\uff1a<\/p>\n

\u4ee5DeepSeek-V4.1-Flash\u4e3a\u4f8b\uff0c\u57288\u5f20PCIe-Only\u663e\u5361\u73af\u5883\u7684\u793e\u533aDay0\u57fa\u7ebf\u7248\u672c\u4e2d\uff0c\u8f93\u5165\u541e\u5410\u4ec51932 tok\/s\u3002<\/p>\n

\u7ecf\u8fc7\u7b97\u6a21\u534f\u540c\u9636\u6bb5\u7684\u4fee\u590d\u9002\u914d\uff0c\u8865\u9f50sparse-MLA Prefill\u5feb\u8def\u5f84\uff0c\u66ff\u6362FP8\u7a20\u5bc6GEMM\u6162\u5185\u6838\uff0c\u6269\u5927PCIe-IPC\u901a\u4fe1\u5feb\u8def\u5f84\u8986\u76d6\u8303\u56f4\uff0c\u541e\u5410\u76f4\u63a5\u63d0\u5347\u81f35850 tok\/s\u3002<\/p>\n

\u8fd9\u4e00\u9636\u6bb5\u7684\u63d0\u5347\uff0c\u5e76\u4e0d\u662f\u521b\u9020\u5168\u65b0\u7b97\u6cd5\uff0c\u800c\u662f\u628a\u786c\u4ef6\u4e0e\u6846\u67b6\u672c\u5c31\u5177\u5907\u3001\u5374\u88ab\u9002\u914d\u95ee\u9898\u9501\u4f4f\u7684\u6027\u80fd\u91ca\u653e\u51fa\u6765\u3002<\/p>\n

\u540c\u6837\u7684\u601d\u8def\u4e5f\u590d\u73b0\u5728DeepSeek-V4-Flash\u3001GLM5.3\u7b49\u6a21\u578b\u4e0a\uff1a\u4fee\u6b63\u6ce8\u610f\u529b\u5934\u8865\u9f50\u7684\u5197\u4f59\u8ba1\u7b97\uff0c\u5bf9KV\u7f13\u5b58\u6807\u8bb0\u903b\u8f91\u505a\u5411\u91cf\u5316\u52a0\u901f\uff0c\u9002\u914d\u786c\u4ef6\u7279\u6027\u91cd\u6784\u7b97\u5b50\u62c6\u5206\u903b\u8f91\uff0c\u62d3\u5c55CUDA\u8ba1\u7b97\u56fe\u8986\u76d6\u8303\u56f4\uff0c\u8ba9\u66f4\u591a\u8bf7\u6c42\u6279\u6b21\u53ef\u4ee5\u8fdb\u5165\u52a0\u901f\u56fe\u6267\u884c\u3002\u4ec5\u4ec5\u5b8c\u6210\u7b97\u6a21\u534f\u540c\uff0c\u591a\u6b3e\u6a21\u578b\u5c31\u62ff\u523020%-33%\u4e0d\u7b49\u7684\u541e\u5410\u589e\u76ca\u3002<\/p>\n

\u5b8c\u6210\u786c\u4ef6\u9002\u914d\u3001\u89e3\u9501\u52a0\u901f\u8def\u5f84\u4e4b\u540e\uff0c\u6027\u80fd\u74f6\u9888\u4f1a\u8fdb\u4e00\u6b65\u8f6c\u79fb\u5230\u901a\u4fe1\u5f00\u9500\u3001\u5e76\u884c\u8c03\u5ea6\u3001\u663e\u5b58\u5206\u914d\u3001\u91cd\u590d\u8ba1\u7b97\u7b49\u73af\u8282\u3002PCIe-Only\u67b6\u6784\u5361\u95f4\u901a\u4fe1\u5e26\u5bbd\u8fdc\u4f4e\u4e8eNVLink\uff0c\u5982\u679c\u76f4\u63a5\u5957\u7528\u9762\u5411\u9ad8\u901f\u4e92\u8054\u786c\u4ef6\u7684\u9ed8\u8ba4\u7b56\u7565\uff0cGPU\u4f1a\u82b1\u5927\u91cf\u65f6\u95f4\u7b49\u5f85\u901a\u4fe1\u5b8c\u6210\uff0c\u7b97\u529b\u8d44\u6e90\u88ab\u95f2\u7f6e\u6d6a\u8d39\u3002<\/p>\n

\u4f9d\u6258\u7b97\u5b50\u4f18\u5316\u4e0e\u901a\u4fe1\u91cd\u6784\uff0cMeta-Infer\u9ad8\u6548\u63a8\u7406\u5f15\u64ce\u5bf9\u6ce8\u610f\u529b\u3001\u6295\u5f71\u7b49\u5173\u952e\u7b97\u5b50\u505a\u7b97\u5b50\u878d\u5408\uff0c\u538b\u7f29\u5355\u6b65\u8ba1\u7b97\u8017\u65f6\uff1b\u5c06\u8ba1\u7b97\u4efb\u52a1\u548c\u96c6\u5408\u901a\u4fe1\u505a\u65f6\u95f4\u91cd\u53e0\u63a9\u76d6\uff0c\u628a\u7edf\u8ba1\u7c7b\u8ba1\u7b97\u5d4c\u5165\u901a\u4fe1\u95f4\u9699\u5e76\u884c\u6267\u884c\uff1b\u6539\u5199\u96c6\u5408\u901a\u4fe1\u903b\u8f91\u9002\u914dPCIe\u5e26\u5bbd\u7279\u6027\uff0c\u964d\u4f4e\u6d88\u606f\u4f20\u8f93\u5e26\u6765\u7684\u7b49\u5f85\u5f00\u9500\u3002<\/p>\n

\u5728\u6b64\u57fa\u7840\u4e0a\u5f00\u5c55\u5e76\u884c\u4e0e\u5bb9\u91cf\u8c03\u4f18\uff0c\u653e\u5f03\u6846\u67b6\u5168\u5c40\u56fa\u5316\u7684\u5e76\u884c\u53c2\u6570\uff0c\u9488\u5bf9Prefill\u9884\u586b\u5145\u3001Decode\u751f\u6210\u4e0d\u540c\u8fd0\u884c\u9636\u6bb5\u5dee\u5f02\u5316\u914d\u7f6e\u5f20\u91cf\u5e76\u884c\u3001\u4e0a\u4e0b\u6587\u5e76\u884c\u7b56\u7565\uff1b\u7ed3\u5408\u786c\u4ef6\u663e\u5b58\u7279\u6027\uff0c\u52a8\u6001\u8c03\u6574\u9759\u6001\u663e\u5b58\u5360\u6bd4\u3001KV\u7f13\u5b58\u5bb9\u91cf\u53c2\u6570\uff0c\u89c4\u907f\u663e\u5b58\u6ea2\u51fa\u4e0e\u7b97\u5b50\u56de\u9000\uff1b\u9762\u5411\u4e0d\u540c\u4e1a\u52a1\u8d1f\u8f7d\uff0c\u7075\u6d3b\u5339\u914d\u6d41\u6c34\u7ebf\u3001\u5f20\u91cf\u5e76\u884c\u7684\u7ec4\u5408\u5f62\u6001\uff0c\u8ba9\u9ad8\u5e76\u53d1\u77ed\u8bf7\u6c42\u3001\u8d85\u957f\u4e0a\u4e0b\u6587\u8bf7\u6c42\u90fd\u53ef\u4ee5\u62ff\u5230\u6700\u4f18\u6267\u884c\u914d\u7f6e\u3002<\/p>\n

\u9762\u5411\u957f\u6587\u672c\u3001\u89c6\u9891\u751f\u6210\u573a\u666f\uff0c\u7f13\u5b58\u590d\u7528\u80fd\u529b\u63d0\u4f9b\u98ce\u9669\u611f\u77e5\u7684\u7f13\u5b58\u590d\u7528\u673a\u5236\uff0c\u6839\u636e\u5b9e\u65f6\u751f\u6210\u72b6\u6001\u52a8\u6001\u5224\u65ad\u7f13\u5b58\u590d\u7528\u4e0e\u5237\u65b0\u65f6\u673a\uff0c\u5728\u4e0d\u727a\u7272\u8f93\u51fa\u8d28\u91cf\u7684\u524d\u63d0\u4e0b\u524a\u51cf\u91cd\u590d\u8ba1\u7b97\u4e0e\u663e\u5b58\u8bfb\u5199\u3002<\/p>\n

\u5728DeepSeek-V4.1-Flash\u7684\u5b9e\u8df5\u4e2d\uff0c\u5b8c\u6210\u7b97\u6a21\u534f\u540c\u4e4b\u540e\uff0c\u7ecf\u8fc7\u901a\u7b97\u91cd\u6784\u7684\u5168\u5957\u8c03\u4f18\uff1a\u6ce8\u610f\u529b\u7b97\u5b50\u878d\u5408\u3001\u5408\u7406\u88c1\u526a\u6295\u673a\u89e3\u7801\u8349\u7a3f\u957f\u5ea6\u3001\u8ba1\u7b97\u901a\u4fe1\u91cd\u53e0\u3001\u533a\u5206Prefill\/Decode\u5e76\u884c\u7b56\u7565\u3001\u91cd\u65b0\u8c03\u6821\u663e\u5b58\u5bb9\u91cf\u53c2\u6570\uff0c\u8f93\u5165\u541e\u5410\u4ece5850 tok\/s\u8fdb\u4e00\u6b65\u63d0\u5347\u81f313274 tok\/s\uff0c\u6574\u4f53\u5b9e\u73b06.87\u500d\u541e\u5410\u63d0\u5347\uff0c\u5e76\u4e14\u652f\u63011M\u8d85\u957f\u4e0a\u4e0b\u6587\u3002<\/p>\n

\u8fd9\u5957\u65b9\u6cd5\u8bba\u5177\u5907\u901a\u7528\u6027\uff0c\u8986\u76d6\u6587\u672c\u5927\u6a21\u578b\u3001\u957f\u4e0a\u4e0b\u6587\u3001\u89c6\u9891\u751f\u6210\u591a\u7c7b\u4efb\u52a1\uff1a<\/p>\n

DeepSeek-V4-Flash\uff1a\u7ecf\u8fc7\u5168\u5957\u4f18\u5316\u8f93\u5165\u541e\u5410\u4ece14546 tok\/s\u63d0\u5347\u81f322584 tok\/s\uff0c\u63d0\u53471.55\u500d\uff1b<\/p>\n

GLM5.3\uff1a\u8f93\u5165\u541e\u5410\u81ea3236.78 tok\/s\u63d0\u5347\u81f36222.72 tok\/s\uff0c\u63d0\u53471.92\u500d\uff1bP95\u9996Token\u65f6\u5ef6\u4ece141.6\u79d2\u4e0b\u964d\u81f346.6\u79d2\uff0c\u4e0a\u4e0b\u6587\u652f\u6301\u4e0a\u9650\u4ece27\u4e07Token\u62d3\u5c55\u5230105\u4e07Token\uff1b<\/p>\n

\u5728\u76f8\u540c\u5e76\u53d1\u70b9\u914d\u5bf9\u6bd4\u8f83\u4e0b\uff0cB300\u7684\u8f93\u5165\u541e\u5410\u7ea6\u4e3a\u8fd9\u53f08\u5361\u6574\u673a\u76844.9\u20135.6\u500d\uff08DeepSeek-V4-Flash\uff0cB300\u4fa7\u6309SGLang cookbook\u63a8\u8350\u53c2\u6570\u914d\u7f6e\uff0cC8\u8f93\u516533,937 \/ C256 60,486 tok\/s\uff09\uff0cGLM-5.3 \u00b9\u4e0a\u8fd9\u4e00\u6bd4\u503c\u4e3a3.5\u20134.7\u500d\uff08B300\u4fa7C8\u8f93\u516516,384 \/ C64 23,503 tok\/s\uff09\u3002<\/p>\n

MiniMax H3\u89c6\u9891\u751f\u6210\u6a21\u578b\uff1a\u4f9d\u6258\u7a00\u758f\u6ce8\u610f\u529b\u3001\u81ea\u7814\u98ce\u9669\u611f\u77e5\u7f13\u5b58\u590d\u7528\u3001Turbo LoRA\u591a\u624b\u6bb5\u7ec4\u5408\uff0c15\u79d2\u53c2\u8003\u56fe\u751f\u89c6\u9891\u7aef\u5230\u7aef\u751f\u6210\u901f\u5ea6\u63d0\u53472.48\u500d\uff0c\u5cf0\u503c\u663e\u5b58\u4e0e\u539f\u59cb\u7a20\u5bc6\u57fa\u7ebf\u4fdd\u6301\u6301\u5e73\u3002<\/p>\n

\u5355\u673a8\u5361\u6574\u673a\u914d\u7f6e\uff0c\u6587\u751f\u89c6\u9891\u6700\u9ad8\u541e\u5410\u8fbe\u5230\u57fa\u7ebf\u76845.33\u500d\uff0c\u53c2\u8003\u56fe\u751f\u89c6\u9891\u8fbe\u5230\u57fa\u7ebf\u76844.98\u500d\uff0c\u5728\u4e0d\u660e\u663e\u635f\u5931\u753b\u8d28\u7684\u524d\u63d0\u4e0b\u5927\u5e45\u7f29\u77ed\u89c6\u9891\u751f\u6210\u8017\u65f6\u3002<\/p>\n

\u628a\u8fd9\u7ec4\u6574\u673a\u541e\u5410\u653e\u5230\u9876\u7ea7GPU\u7684\u540c\u4e00\u628a\u6807\u5c3a\u4e0a\uff0c\u5dee\u8ddd\u6bd4\u76f4\u89c9\u66f4\u5c0f\uff1a\u5728\u7edf\u4e00\u6574\u673a\u53e3\u5f84\u4e0b\uff08\u5355\u673a8\u5361\u30015\u79d2 \/ 768P \/ 16\uff1a9\uff09\uff0c\u6587\u751f\u89c6\u9891296\u6761\/\u65f6\u3001\u53c2\u8003\u56fe\u751f\u89c6\u9891131\u6761\/\u65f6\uff0c\u5bf9\u5e94B300\u7684439\u6761\/\u65f6\u3001225\u6761\/\u65f6\uff0c\u5206\u522b\u7ea6\u4e3a\u517667%\u548c58%\uff08B300\u4fa7\u6309SGLang cookbook\u63a8\u8350\u53c2\u6570\u914d\u7f6e\uff0c\u53d616\u5b9e\u4f8b\u3001\u6bcf\u5b9e\u4f8bInflight=1\u7684\u6700\u4f18\u6863\uff1b\u5176\u4e2d\u6587\u751f\u89c6\u9891\u5bf9\u7167\u7684\u662fB300\u7684FL2VA\u7ed3\u679c\uff09\u3002<\/p>\n

\u5982\u679c\u8fdb\u4e00\u6b65\u6309\u6574\u673a\u541e\u5410\u6298\u7b97\uff0c\u5728\u51e0\u4e4e\u65e0\u635f\u7684Ours Cache\u65b9\u6848\u4e0b\uff0c\u7ea62.6\u53f06000D\u53ef\u5bf9\u5e941\u53f0B300\u7684\u6587\u751f\u89c6\u9891\u541e\u5410\uff0c\u53c2\u8003\u56fe\u751f\u89c6\u9891\u7ea6\u4e3a2.9\u53f0\uff1b\u5728Ours Cache + LoRA\u65b9\u6848\u4e0b\uff0c\u8fd9\u4e00\u6bd4\u4f8b\u8fdb\u4e00\u6b65\u7f29\u5c0f\u5230\u7ea61.5\u53f0\u548c1.7\u53f0\u3002<\/p>\n

\u786c\u4ef6\u672c\u8eab\u6ca1\u6709\u53d8\u5316\uff0c\u6027\u80fd\u5dee\u8ddd\u7684\u7f29\u5c0f\u4e3b\u8981\u6765\u81ea\u8f6f\u4ef6\u6808\u3001\u7f13\u5b58\u7b56\u7565\u4e0e\u6a21\u578b\u4f18\u5316\u65b9\u5f0f\u7684\u7ec4\u5408\u3002<\/p>\n

\u4e0b\u9762\u8fd9\u6bb5\u662f\u53c2\u8003\u56fe\u751f\u89c6\u9891\uff08Ref2VA\uff09\u4efb\u52a1\u7684\u540c\u9898\u5bf9\u7167\uff0cDense\u57fa\u7ebf\u3001Cache-DiT\u3001Turbo LoRA\u4e0e\u672c\u6587\u65b9\u6848\u5e76\u6392\u5448\u73b0\u3002<\/p>\n

Meta-Infer\u9ad8\u6548\u63a8\u7406\u5f15\u64ce\u4e5f\u5728\u56fd\u4ea7GPU\u4e0a\u5b8c\u6210\u4e86\u9a8c\u8bc1\u3002<\/p>\n

\u56fd\u4ea7GPU\u5f80\u5f80\u4e5f\u4e0d\u5728\u4e3b\u6d41\u6846\u67b6\u7684\u5b98\u65b9\u9a8c\u8bc1\u77e9\u9635\u5185\uff0c\u56e0\u6b64\u9700\u8981\u9488\u5bf9\u5177\u4f53\u786c\u4ef6\u91cd\u65b0\u68b3\u7406\u6267\u884c\u8def\u5f84\u3002<\/p>\n

\u4f8b\u5982\uff0c\u6ce8\u610f\u529b\u6a21\u5757\u9700\u8981\u663e\u5f0f\u542f\u7528\u9762\u5411\u8be5\u5e73\u53f0\u4f18\u5316\u7684NSA\u7a00\u758f\u6ce8\u610f\u529b\u5b9e\u73b0\uff0c\u907f\u514d\u56de\u9000\u5230\u4f4e\u6548\u8def\u5f84\uff1bShared Expert\u878d\u5408\u7b49\u9ed8\u8ba4\u9762\u5411NVIDIA\/AMD\u5e73\u53f0\u7684\u7279\u6027\u5219\u9700\u8981\u5173\u95ed\uff0c\u4ee5\u7ed5\u5f00\u4e0d\u9002\u914d\u5b9e\u73b0\u3002<\/p>\n

\u901a\u4fe1\u7b56\u7565\u4e5f\u9700\u8981\u6839\u636e\u63a8\u7406\u9636\u6bb5\u5206\u522b\u914d\u7f6e\uff0cPrefill\u91c7\u7528\u504f\u541e\u5410\u4f18\u5316\u7684DeepEP normal\u6a21\u5f0f\uff0cDecode\u91c7\u7528\u504f\u65f6\u5ef6\u4f18\u5316\u7684low_latency\u6a21\u5f0f\uff0c\u5e76\u5c06Shared Expert\u4e0eRouted Expert\u62c6\u5206\u5230\u4e24\u6761stream\u4e2d\u5e76\u884c\u63d0\u4ea4\uff0c\u4ee5\u8fdb\u4e00\u6b65\u4e0e\u96c6\u5408\u901a\u4fe1\u91cd\u53e0\u3002<\/p>\n

\u5176\u4e2d\uff0c\u4ec5\u5c06Shared Expert\u4e0eRouted Expert\u6539\u4e3a\u5e76\u884c\u63d0\u4ea4\uff0c\u5c31\u572835\u7ec4\u540c\u914d\u7f6e\u914d\u5bf9\u6d4b\u8bd5\u4e2d\u5e26\u6765\u4e8611.26%\u7684\u541e\u5410\u63d0\u5347\u3002<\/p>\n

\u8fd9\u8bf4\u660e\u5bf9\u4e8e\u56fd\u4ea7\u7b97\u529b\u800c\u8a00\uff0c\u201c\u80fd\u8dd1\u8d77\u6765\u201d\u540c\u6837\u53ea\u662f\u7b2c\u4e00\u6b65\u3002<\/p>\n

\u901a\u8fc7\u9488\u5bf9\u6267\u884c\u8def\u5f84\u3001\u901a\u4fe1\u6a21\u5f0f\u548c\u5e76\u884c\u7b56\u7565\u8fdb\u884c\u7cfb\u7edf\u9002\u914d\uff0c\u4ecd\u53ef\u4ee5\u8fdb\u4e00\u6b65\u91ca\u653e\u786c\u4ef6\u7684\u5b9e\u9645\u63a8\u7406\u6027\u80fd\uff0c\u4f7f\u5176\u66f4\u63a5\u8fd1\u89c4\u6a21\u5316\u751f\u4ea7\u670d\u52a1\u6240\u9700\u7684\u6027\u80fd\u6c34\u5e73\u3002<\/p>\n

\u5168\u90e8\u4f18\u5316\u5de5\u4f5c\u5747\u6536\u655b\u5728Meta-Infer\u9ad8\u6548\u63a8\u7406\u5f15\u64ce\u5185\u90e8\uff0c\u5168\u7a0b\u65e0\u9700\u6539\u52a8\u6a21\u578b\u6743\u91cd\u4e0e\u6a21\u578b\u7ed3\u6784\uff0c\u4e0d\u6539\u53d8\u539f\u6709\u4e1a\u52a1\u4efb\u52a1\u8bed\u4e49\u3002<\/p>\n

\u5bf9\u4e8e\u5f00\u6e90\u793e\u533a\u5927\u91cf\u4e0d\u5728\u5b98\u65b9\u9a8c\u8bc1\u77e9\u9635\u5185\u7684\u957f\u5c3e\u786c\u4ef6\uff0c\u884c\u4e1a\u8fc7\u5f80\u73b0\u72b6\u5f80\u5f80\u662f\uff1a\u65b0\u6a21\u578b\u53d1\u5e03\uff0c\u6743\u91cd\u53ef\u4ee5\u4e0b\u8f7d\u3001\u670d\u52a1\u53ef\u4ee5\u62c9\u8d77\uff0c\u505a\u5230\u201c\u80fd\u8dd1\u8d77\u6765\u201d\uff0c\u5de5\u4f5c\u5c31\u5ba3\u544a\u7ed3\u675f\u3002<\/p>\n

\u4f46\u8fd9\u5957\u5b9e\u8df5\u8bc1\u660e\uff0c\u201c\u80fd\u8dd1\u201d\u548c\u201c\u53ef\u7528\u3001\u53ef\u89c4\u6a21\u5316\u5bf9\u5916\u63d0\u4f9b\u751f\u4ea7\u670d\u52a1\u201d\u4e4b\u95f4\uff0c\u8fd8\u5b58\u5728\u5de8\u5927\u7684\u8f6f\u4ef6\u4f18\u5316\u7a7a\u95f4\u3002<\/p>\n

\u5728\u786c\u4ef6\u4f9b\u7ed9\u5b58\u5728\u7ea6\u675f\u3001\u9ad8\u7aef\u7b97\u529b\u6210\u672c\u5c45\u9ad8\u4e0d\u4e0b\u7684\u5927\u80cc\u666f\u4e0b\uff0cMeta-Infer\u9ad8\u6548\u63a8\u7406\u5f15\u64ce\u4ee3\u8868\u4e86\u4e00\u6761\u65b0\u8def\u7ebf\u2014\u2014\u4e0d\u5fc5\u4e00\u5473\u8ffd\u9010\u6700\u9876\u7ea7\u786c\u4ef6\uff0c\u901a\u8fc7\u5185\u6838\u8865\u9f50\u3001\u7b97\u5b50\u901a\u4fe1\u91cd\u6784\u3001\u5e76\u884c\u5bb9\u91cf\u8c03\u4f18\u3001\u7f13\u5b58\u590d\u7528\u8fd9\u5957\u8f6f\u4ef6\u4f53\u7cfb\uff0c\u5145\u5206\u6316\u6398\u73b0\u6709PCIe\u67b6\u6784\u786c\u4ef6\u7684\u6f5c\u529b\u3002<\/p>\n

\u8fd9\u4e3a\u5927\u6a21\u578b\u63a8\u7406\u4e1a\u52a1\u5e26\u6765\u4e00\u79cd\u503c\u5f97\u5173\u6ce8\u7684\u53ef\u80fd\u6027\uff1a\u501f\u52a9\u6df1\u5ea6\u8f6f\u4ef6\u4f18\u5316\uff0c\u6210\u672c\u66f4\u53cb\u597d\u7684PCIe GPU\uff0c\u6709\u673a\u4f1a\u5728\u90e8\u5206\u63a8\u7406\u4e1a\u52a1\u6307\u6807\u4e0a\u5411\u66f4\u9ad8\u7aef\u786c\u4ef6\u7684\u670d\u52a1\u80fd\u529b\u9760\u62e2\u3002<\/p>\n

\u786c\u4ef6\u7684\u4e0a\u9650\u7531\u82af\u7247\u51b3\u5b9a\uff0c\u4f46\u786c\u4ef6\u80fd\u5b9e\u9645\u53d1\u6325\u51fa\u591a\u5c11\u80fd\u529b\uff0c\u5219\u7531\u8f6f\u4ef6\u51b3\u5b9a\u3002<\/p>\n

\u00b9 \u65e0\u635fNVFP4\u91cf\u5316\u7248\u672c\uff1a\u6a21\u578b\u6743\u91cd\u4ee5NVFP4\u90e8\u7f72\uff0c\u4fdd\u6301\u65e2\u6709\u6a21\u578b\u8ba1\u7b97\u5b9a\u4e49\uff0c\u672a\u5f15\u5165\u989d\u5916\u7cbe\u5ea6\u635f\u5931\u3002<\/p>\n

\n

\u7248\u6743\u58f0\u660e\uff1a<\/strong><\/p>\n

\u672c\u535a\u5ba2\u90e8\u5206\u5185\u5bb9\u4e3a\u8f6c\u8f7d\u6587\u7ae0\uff0c\u65e8\u5728\u5206\u4eab\u6709\u4ef7\u503c\u7684\u4fe1\u606f\uff0c\u7248\u6743\u5f52\u539f\u4f5c\u8005\u6240\u6709\u3002<\/p>\n

\u8f6c\u8f7d\u4ec5\u4e3a\u4e2a\u4eba\u5b66\u4e60\u4e0e\u4ea4\u6d41\u76ee\u7684\uff0c\u4e0d\u5bf9\u6587\u7ae0\u89c2\u70b9\u8d1f\u8d23\uff0c\u4ea6\u4e0d\u7528\u4e8e\u4efb\u4f55\u5546\u4e1a\u7528\u9014\u3002<\/p>\n

\u5982\u6d89\u53ca\u7248\u6743\u95ee\u9898\uff0c\u8bf7\u8054\u7cfb\u672c\u4eba\u5220\u9664\u3002<\/p>\n<\/blockquote>\n","protected":false},"excerpt":{"rendered":"

\u5141\u4e2d \u53d1\u81ea \u51f9\u975e\u5bfa<\/p>\n

\u968f\u7740\u5927\u6a21\u578b\u5bf9\u7b97\u529b\u9700\u6c42\u6301\u7eed\u6500\u5347\uff0cGPU\u5361\u7684\u91c7\u8d2d\u6210\u672c\u3001\u4f9b\u7ed9\u7ea6\u675f\u6301\u7eed\u52a0\u5267\u3002<\/p>\n

AI\u63a8\u7406\u7684\u7ade\u4e89\u6b63\u5728\u6084\u7136\u53d1\u751f\u8f6c\u5411\uff1a\u4e0d\u518d\u5355\u7eaf\u6bd4\u62fc\u201c\u8c01\u80fd\u591f\u4e70\u5230\u6027\u80fd\u6700\u9876\u7ea7\u7684\u786c\u4ef6\u201d\uff0c\u800c\u662f\u8f6c\u5411\u201c\u8c01\u53ef\u4ee5\u628a\u624b\u4e2d\u5df2\u6709\u7684\u7b97\u529b\u8d44\u6e90\u7528\u5f97\u66f4\u503c\u201d\u3002<\/p>\n

\u5f88\u591aGPU\u5361\uff0c\u5f00\u6e90\u6846\u67b6\u80fd\u591f\u5b8c\u6210\u6a21\u578b\u52a0\u8f7d\u3001\u6743\u91cd\u4e0b\u8f7d\u3001\u670d\u52a1\u62c9\u8d77\uff0c\u5b9e\u73b0\u201c\u80fd\u8dd1\u8d77\u6765\u201d\uff0c\u4f46\u5b9e\u9645\u538b\u6d4b\u6027\u80fd\u5f80\u5f80\u8fdc\u4f4e\u4e8e\u5b98\u65b9\u5ba3\u4f20\u6c34\u5e73\u3002\u6a21\u578b\u672c\u8eab\u6ca1\u6709\u7f3a\u9677\uff0c\u786c\u4ef6\u4e5f\u53ef\u6b63\u5e38\u5de5\u4f5c\uff0c\u6027\u80fd\u635f\u5931\u7684\u6838\u5fc3\uff0c\u6765\u81ea\u6a21\u578b\u6846<\/p>\n","protected":false},"author":1,"featured_media":8458,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[246,1261,1262,24],"class_list":["post-8460","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-xwzx","tag-deepseek","tag-1261","tag-1262","tag-24"],"_links":{"self":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/posts\/8460","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/comments?post=8460"}],"version-history":[{"count":0,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/posts\/8460\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/media\/8458"}],"wp:attachment":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/media?parent=8460"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/categories?post=8460"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/tags?post=8460"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}