\u983b\u5bec\u5206\u6790\u548cGPU\u4e92\u9023<\/strong><\/h2>\n

\u5728\u9999\u6e2f\u8cc7\u6599\u4e2d\u5fc3\u5efa\u69cb\u591aGPU\u7cfb\u7d71\u6642\uff0c\u7406\u89e3\u983b\u5bec\u5206\u914d\u81f3\u95dc\u91cd\u8981\u3002\u4ee5\u4e0b\u662f\u4f7f\u7528\u96d9CPU\u4f3a\u670d\u5668\u914d\u7f6e\u7684\u8a73\u7d30\u5206\u6790\uff1a<\/p>\n

\r\n# \u983b\u5bec\u5206\u914d\u793a\u4f8b\uff08\u96d9Intel Xeon\u5e73\u53f0\uff09\r\nCPU1 \u2192 GPU1: PCIe Gen4 x16 (31.5 GB\/s)\r\nCPU1 \u2192 GPU2: PCIe Gen4 x16 (31.5 GB\/s)\r\nCPU2 \u2192 GPU3: PCIe Gen4 x16 (31.5 GB\/s)\r\nCPU2 \u2192 GPU4: PCIe Gen4 x16 (31.5 GB\/s)\r\n\r\nCPU\u9593\u901a\u8a0a\uff1aUPI\u9023\u7d50\r\n3\u689dUPI\u9023\u7d50 \u00d7 23.3 GB\/s = \u7e3d\u8a0869.9 GB\/s\r\n<\/code><\/pre>\n
\u9999\u6e2f\u7279\u5b9a\u914d\u7f6e\u8003\u91cf\u56e0\u7d20<\/strong><\/h2>\n
\u9999\u6e2f\u7684\u6c23\u5019\u70baGPU\u4f3a\u670d\u5668\u90e8\u7f72\u5e36\u4f86\u7368\u7279\u6311\u6230\u3002\u9ad8\u6fd5\u5ea6\u548c\u9ad8\u6eab\u9700\u8981\u7279\u5b9a\u7684PCIe\u62d3\u64b2\u7d50\u69cb\u8003\u91cf\uff1a<\/p>\n
\n
PCIe\u63d2\u69fd\u7684\u71b1\u8a2d\u8a08\u529f\u8017\uff08TDP\uff09\u5206\u5e03<\/li>\n
\u900f\u904e\u7b56\u7565\u6027GPU\u4f48\u7f6e\u512a\u5316\u6c23\u6d41<\/li>\n
\u9ad8\u5bc6\u5ea6\u914d\u7f6e\u7684\u5197\u9918\u6563\u71b1\u7cfb\u7d71<\/li>\n<\/ul>\n
\u70ba\u5728\u9999\u6e2f\u74b0\u5883\u4e2d\u7372\u5f97\u6700\u4f73\u6548\u80fd\uff0c\u5efa\u8b70\u63a1\u7528\u4ee5\u4e0bPCIe\u63d2\u69fd\u914d\u7f6e\uff1a<\/p>\n
\r\n# \u63a8\u85a6\u7684PCIe\u63d2\u69fd\u914d\u7f6e\r\n\u63d2\u69fd1\uff1aGPU1\uff08\u4e3b\u8981\uff09 - PCIe Gen4 x16\r\n\u63d2\u69fd3\uff1aGPU2 - PCIe Gen4 x16\r\n\u63d2\u69fd5\uff1aGPU3 - PCIe Gen4 x16\r\n\u63d2\u69fd7\uff1aGPU4 - PCIe Gen4 x16\r\n\r\n# \u6ce8\u610f\uff1a\u4fdd\u6301\u6700\u5c112\u500b\u63d2\u69fd\u9593\u8ddd\u4ee5\u9032\u884c\u6563\u71b1\u7ba1\u7406\r\n<\/code><\/pre>\n
\u6548\u80fd\u512a\u5316\u6280\u8853<\/strong><\/h2>\n
\u8981\u5728\u9999\u6e2f\u4f3a\u670d\u5668\u79df\u7528\u74b0\u5883\u4e2d\u6700\u5927\u5316GPU\u4f3a\u670d\u5668\u6548\u80fd\uff0c\u8acb\u5be6\u65bd\u9019\u4e9bPCIe\u62d3\u64b2\u512a\u5316\uff1a<\/p>\n
\n
NUMA\u7bc0\u9ede\u512a\u5316\n
\n
\u5c07GPU\u7d81\u5b9a\u5230\u672c\u5730NUMA\u7bc0\u9ede<\/li>\n
\u6700\u5c0f\u5316\u8de8NUMA\u901a\u8a0a<\/li>\n
\u512a\u5316\u8a18\u61b6\u9ad4\u5206\u914d\u6a21\u5f0f<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n
\u4ee5\u4e0b\u662fLinux\u4e2dNUMA\u7d81\u5b9a\u7684\u5be6\u969b\u793a\u4f8b\uff1a<\/p>\n
\r\n# NUMA\u7d81\u5b9a\u793a\u4f8b\r\nnumactl --cpunodebind=0 --membind=0 .\/gpu_application  # \u7528\u65bcGPU0\/1\r\nnumactl --cpunodebind=1 --membind=1 .\/gpu_application  # \u7528\u65bcGPU2\/3\r\n\r\n# \u6aa2\u67e5NUMA\u62d3\u64b2\r\nnvidia-smi topo -m\r\n<\/code><\/pre>\n
PCIe\u62d3\u64b2\u7d50\u69cb\u57fa\u6e96\u6e2c\u8a66<\/strong><\/h2>\n
\u6548\u80fd\u9a57\u8b49\u5c0d\u9999\u6e2fGPU\u4f3a\u670d\u5668\u79df\u7528\u74b0\u5883\u81f3\u95dc\u91cd\u8981\u3002\u4ee5\u4e0b\u662f\u62d3\u64b2\u6e2c\u8a66\u7684\u7cfb\u7d71\u65b9\u6cd5\uff1a<\/p>\n
\r\n# \u983b\u5bec\u6e2c\u8a66\u8173\u672c\r\n#!\/bin\/bash\r\nfor i in {0..3}; do\r\n    for j in {0..3}; do\r\n        if [ $i -ne $j ]; then\r\n            nvidia-smi topo -p2p r -i $i -j $j\r\n            .\/bandwidth_test --src $i --dst $j\r\n        fi\r\n    done\r\ndone\r\n<\/code><\/pre>\n
\u5e38\u898b\u554f\u984c\u6545\u969c\u6392\u9664<\/strong><\/h2>\n
\u5728\u9999\u6e2f\u4f3a\u670d\u5668\u8a17\u7ba1\u8a2d\u65bd\u90e8\u7f72GPU\u4f3a\u670d\u5668\u6642\uff0c\u9700\u8981\u6ce8\u610f\u9019\u4e9bPCIe\u62d3\u64b2\u76f8\u95dc\u554f\u984c\uff1a<\/p>\n
\n
PCIe\u9023\u7d50\u8a13\u7df4\u5931\u6557\n
\n
\u6aa2\u67e5\u5be6\u9ad4\u9023\u63a5\u5b8c\u6574\u6027<\/li>\n
\u9a57\u8b49BIOS PCIe\u4ee3\u6578\u8a2d\u5b9a<\/li>\n
\u76e3\u63a7\u7cfb\u7d71\u4e8b\u4ef6\u65e5\u8a8c<\/li>\n<\/ul>\n<\/li>\n
\u983b\u5bec\u964d\u7d1a\n
\n
\u76e3\u63a7PCIe\u9023\u7d50\u5bec\u5ea6\u5354\u5546<\/li>\n
\u9a57\u8b49\u6563\u71b1\u6548\u80fd<\/li>\n
\u6aa2\u67e5\u4f9b\u96fb\u7a69\u5b9a\u6027<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n
GPU\u57fa\u790e\u8a2d\u65bd\u7684\u672a\u4f86\u5c55\u671b<\/strong><\/h2>\n
\u5c55\u671b\u9999\u6e2fGPU\u4f3a\u670d\u5668\u79df\u7528\u9818\u57df\u7684\u672a\u4f86\uff0c\u8acb\u8003\u616e\u9019\u4e9b\u65b0\u8208\u6280\u8853\uff1a<\/p>\n
\n
PCIe Gen5\u5be6\u73fe\n
\n
\u6bcf\u500bx16\u63d2\u69fd63 GB\/s\u7406\u8ad6\u983b\u5bec<\/li>\n
\u589e\u5f37\u7684\u932f\u8aa4\u6aa2\u6e2c\u548c\u4fee\u6b63<\/li>\n
\u6539\u9032\u7684\u96fb\u6e90\u7ba1\u7406\u529f\u80fd<\/li>\n<\/ul>\n<\/li>\n
\u904b\u7b97\u5feb\u901f\u9023\u7d50\uff08CXL\uff09\u6574\u5408\n
\n
\u5feb\u53d6\u4e00\u81f4\u6027\u652f\u63f4<\/li>\n
\u8a18\u61b6\u9ad4\u6c60\u5316\u529f\u80fd<\/li>\n
\u964d\u4f4eGPU-CPU\u901a\u8a0a\u5ef6\u9072<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n
\u7d50\u8ad6<\/strong><\/h2>\n
\u5728\u9999\u6e2f\u7684\u4f3a\u670d\u5668\u79df\u7528\u74b0\u5883\u4e2d\uff0c\u512a\u5316GPU\u4f3a\u670d\u5668\u4e2d\u7684PCIe\u62d3\u64b2\u7d50\u69cb\u4ecd\u7136\u662f\u9ad8\u6548\u80fd\u904b\u7b97\u7684\u57fa\u790e\u3002\u7406\u89e3PCIe\u901a\u9053\u3001\u983b\u5bec\u5206\u914d\u548c\u6563\u71b1\u8003\u91cf\u56e0\u7d20\u4e4b\u9593\u7684\u8907\u96dc\u95dc\u4fc2\uff0c\u80fd\u5920\u5be6\u73fe\u6700\u4f73GPU\u4f3a\u670d\u5668\u914d\u7f6e\u3002\u96a8\u8457\u9999\u6e2f\u7e7c\u7e8c\u767c\u5c55\u6210\u70ba\u4e3b\u8981\u8cc7\u6599\u4e2d\u5fc3\u6a1e\u7d10\uff0c\u5be6\u65bd\u9019\u4e9bPCIe\u62d3\u64b2\u6700\u4f73\u5be6\u52d9\u5c07\u78ba\u4fddGPU\u4f3a\u670d\u5668\u79df\u7528\u90e8\u7f72\u7684\u6700\u5927\u6548\u80fd\u548c\u53ef\u9760\u6027\u3002<\/p>\n
\u5982\u9700\u9032\u4e00\u6b65\u5354\u52a9\u512a\u5316\u9999\u6e2f\u8cc7\u6599\u4e2d\u5fc3\u7684GPU\u4f3a\u670d\u5668PCIe\u62d3\u64b2\u7d50\u69cb\uff0c\u8acb\u806f\u7d61\u6211\u5011\u7684\u6280\u8853\u5718\u968a\u8a0e\u8ad6\u60a8\u7684\u5177\u9ad4\u4f3a\u670d\u5668\u79df\u7528\u9700\u6c42\u3002<\/p>\n
[\/vc_column_text][\/vc_column][\/vc_row]<\/p>\n<\/div>","protected":false},"excerpt":{"rendered":"
[vc_row][vc_column][vc_column_text css=””]  […]<\/p>\n
Read More…<\/a><\/p>\n","protected":false},"author":2,"featured_media":21896,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[61],"tags":[6178,6179,6180,6181,4325],"class_list":["post-21902","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-hong-kong-dedicated-server-tc","tag-gpu-server-pcie-topology-tc","tag-pcie-bandwidth-optimization-tc","tag-gpu-interconnect-architecture-tc","tag-data-center-gpu-configuration-tc","tag-hong-kong-gpu-hosting-tc"],"acf":[],"yoast_head":"\n