{"id":7820,"date":"2026-06-17T09:36:31","date_gmt":"2026-06-17T01:36:31","guid":{"rendered":"https:\/\/www.starverse-ai.com\/guide\/?p=7820"},"modified":"2026-06-17T09:36:32","modified_gmt":"2026-06-17T01:36:32","slug":"rdma-%e9%ab%98%e9%80%9f%e7%bd%91%e7%bb%9c%e9%83%a8%e7%bd%b2%e8%90%bd%e5%9c%b0%ef%bc%8c%e5%a4%a7%e6%a8%a1%e5%9e%8b%e5%88%86%e5%b8%83%e5%bc%8f%e7%a7%9f%e8%b5%81%e8%ae%ad%e7%bb%83%e4%bc%a0%e8%be%93","status":"publish","type":"post","link":"https:\/\/www.starverse-ai.com\/guide\/archives\/7820","title":{"rendered":"RDMA \u9ad8\u901f\u7f51\u7edc\u90e8\u7f72\u843d\u5730\uff0c\u5927\u6a21\u578b\u5206\u5e03\u5f0f\u79df\u8d41\u8bad\u7ec3\u4f20\u8f93\u5ef6\u8fdf\u4f18\u5316"},"content":{"rendered":"\n

\u4e00\u3001\u884c\u4e1a\u80cc\u666f\u4e0e\u6838\u5fc3\u75db\u70b9<\/h2>\n\n\n\n

\u5343\u4ebf\u3001\u4e07\u4ebf\u53c2\u6570\u5927\u6a21\u578b\u8bad\u7ec3\u4f9d\u8d56\u591a GPU \u5206\u5e03\u5f0f\u534f\u540c\uff0c\u6570\u636e\u5e76\u884c\u3001\u5f20\u91cf\u5e76\u884c\u3001\u4e13\u5bb6\u5e76\u884c\u6d41\u7a0b\u4e2d\uff0cAllReduce\u3001AllToAll \u96c6\u5408\u901a\u4fe1\u8017\u65f6\u5360\u5355\u8f6e\u8fed\u4ee3\u603b\u65f6\u957f 65%~78%\uff0c\u4f20\u7edf TCP\/IP \u4ee5\u592a\u7f51\u5185\u6838\u62f7\u8d1d\u3001\u534f\u8bae\u6808\u5f00\u9500\u3001CPU \u62a2\u5360\u7b97\u529b\u4e09\u5927\u95ee\u9898\u76f4\u63a5\u62c9\u957f\u8bad\u7ec3\u5468\u671f\uff0c\u62ac\u9ad8\u7b97\u529b\u79df\u8d41\u6210\u672c\u3002<\/p>\n\n\n\n

\u5b9e\u6d4b\u57fa\u51c6\u6570\u636e\uff1a128 \u5361 H100 \u96c6\u7fa4\uff0c\u4f20\u7edf 100G \u4ee5\u592a\u7f51 TCP \u7ec4\u7f51\u5355\u8f6e AllReduce \u5ef6\u8fdf\u5747\u503c 52\u03bcs\uff0cGPU \u7b97\u529b\u5229\u7528\u7387\u4ec5 61%\uff1b\u540c\u89c4\u683c RDMA RoCEv2 \u7ec4\u7f51\u5ef6\u8fdf\u964d\u81f3 4.3\u03bcs\uff0c\u7b97\u529b\u5229\u7528\u7387\u63d0\u5347\u81f3 89%\u3002\u4e2d\u5c0f AI \u4f01\u4e1a\u81ea\u5efa RDMA \u96c6\u7fa4\u5b58\u5728\u786c\u4ef6\u9009\u578b\u590d\u6742\u3001\u534f\u8bae\u8c03\u4f18\u95e8\u69db\u9ad8\u3001\u591a\u79df\u6237\u9694\u79bb\u96be\u5ea6\u5927\u7b49\u95ee\u9898\uff0c\u661f\u5b87\u667a\u7b97<\/strong>\u6807\u51c6\u5316 RDMA \u7b97\u529b\u79df\u8d41\u96c6\u7fa4\uff0c\u5b8c\u6210\u4ece\u786c\u4ef6\u7ec4\u7f51\u3001\u534f\u8bae\u6808\u90e8\u7f72\u3001NCCL \u9002\u914d\u5230\u4efb\u52a1\u8c03\u5ea6\u5168\u94fe\u8def\u5c01\u88c5\uff0c\u964d\u4f4e\u5206\u5e03\u5f0f\u8bad\u7ec3\u843d\u5730\u95e8\u69db\u3002<\/p>\n\n\n\n

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

\u4e8c\u3001\u6838\u5fc3\u6280\u672f\u5206\u4eab\uff1aRDMA \u964d\u5ef6\u8fdf\u5e95\u5c42\u903b\u8f91\u4e0e\u5206\u5c42\u8c03\u4f18\u65b9\u6848<\/h2>\n\n\n\n

2.1 RDMA \u4e0e\u4f20\u7edf TCP \u901a\u4fe1\u91cf\u5316\u6307\u6807\u5bf9\u6bd4<\/h3>\n\n\n\n

\u4e0b\u8868\u57fa\u4e8e\u661f\u5b87\u667a\u7b97 100G RoCEv2 \u751f\u4ea7\u96c6\u7fa4\u5b9e\u6d4b\uff0c\u786c\u4ef6\u7edf\u4e00\u4e3a ConnectX-6 \u7f51\u5361\u3001H100 80G GPU\uff1a<\/p>\n\n\n\n

\u6027\u80fd\u6307\u6807<\/th>\u4f20\u7edf TCP\/IP \u4ee5\u592a\u7f51<\/th>RDMA RoCEv2<\/th>\u91cf\u5316\u4f18\u5316\u5e45\u5ea6<\/th><\/tr><\/thead>
\u5355\u5305\u7aef\u5230\u7aef\u5ef6\u8fdf<\/td>50~100\u03bcs<\/td>1.5~5\u03bcs<\/td>\u5ef6\u8fdf\u964d\u4f4e 90%<\/td><\/tr>
100Gbps \u5e26\u5bbd CPU \u5360\u7528<\/td>32 \u6838\u6ee1\u8f7d<\/td>0.6 \u6838\u4ee5\u5185<\/td>CPU \u5f00\u9500\u4e0b\u964d 97%<\/td><\/tr>
\u6709\u6548\u94fe\u8def\u541e\u5410\u4e0a\u9650<\/td>68~79Gbps<\/td>94~97Gbps<\/td>\u6709\u6548\u5e26\u5bbd\u63d0\u5347 32%<\/td><\/tr>
128 \u5361 AllReduce \u8017\u65f6<\/td>52ms<\/td>7.8ms<\/td>\u96c6\u5408\u901a\u4fe1\u8017\u65f6\u51cf\u5c11 85%<\/td><\/tr>
34B \u5927\u6a21\u578b\u8bad\u7ec3\u5355\u8f6e\u8fed\u4ee3\u8017\u65f6<\/td>3210ms<\/td>1160ms<\/td>\u8bad\u7ec3\u901f\u5ea6\u63d0\u5347 176%<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

RDMA \u6838\u5fc3\u4f18\u5316\u8def\u5f84\uff1a\u7ed5\u8fc7\u64cd\u4f5c\u7cfb\u7edf\u5185\u6838\u534f\u8bae\u6808\uff0cRNIC \u7f51\u5361\u786c\u4ef6\u5b8c\u6210\u6570\u636e\u5305\u5c01\u88c5\u3001\u62e5\u585e\u63a7\u5236\u3001\u91cd\u4f20\uff1b\u91c7\u7528\u5355\u8fb9\u8bfb\u5199\u64cd\u4f5c\uff0c\u53d6\u6d88 CPU \u4e2d\u8f6c\u5185\u5b58\u62f7\u8d1d\uff1b\u914d\u5408 DCQCN \u62e5\u585e\u63a7\u5236\u7b97\u6cd5\u6291\u5236\u7f51\u7edc\u6296\u52a8\uff0c\u5c06\u541e\u5410\u6ce2\u52a8\u4ece \u00b130% \u6536\u655b\u81f3 \u00b15% \u4ee5\u5185\u3002<\/p>\n\n\n\n

2.2 \u5206\u5c42\u5ef6\u8fdf\u4f18\u5316\u843d\u5730\u65b9\u6848<\/h3>\n\n\n\n
    \n
  1. \u786c\u4ef6\u5c42<\/strong>\uff1a\u7edf\u4e00 100G RoCE \u7f51\u5361 + \u53f6\u810a\u62d3\u6251\u4ea4\u6362\u673a\uff0c\u5173\u95ed PCIe ACS \u5b89\u5168\u9694\u79bb\u901a\u9053\uff0c\u5b9e\u73b0 GPU \u76f4\u8fde\u7f51\u5361\uff0c\u6d88\u9664 PCIe \u8f6c\u53d1\u5f00\u9500\uff1b\u661f\u5b87\u667a\u7b97\u79df\u8d41\u96c6\u7fa4\u6807\u914d\u53cc RDMA \u7f51\u5361\u5197\u4f59\u94fe\u8def\uff0c\u94fe\u8def\u4e22\u5305\u7387\u7a33\u5b9a\u4f4e\u4e8e 0.01%\u3002<\/li>\n\n\n\n
  2. \u534f\u8bae\u5c42<\/strong>\uff1a\u90e8\u7f72 MLNX_OFED 5.8 \u9a71\u52a8\uff0c\u542f\u7528 RoCEv2 \u534f\u8bae\uff0c\u8c03\u4f18 DCQCN \u03b1=200\u3001\u03b2=10 \u53c2\u6570\uff1b\u7981\u7528\u9632\u706b\u5899\u5bf9 RDMA \u7aef\u53e3\u62e6\u622a\uff0c\u5168\u5c40\u5f00\u542f nvidia-peermem \u7ec4\u4ef6\u6253\u901a GPU \u663e\u5b58\u4e0e RDMA \u7f13\u51b2\u533a\u3002<\/li>\n\n\n\n
  3. \u6846\u67b6\u5c42<\/strong>\uff1aNCCL \u9002\u914d RDMA \u540e\u7aef\uff0c\u542f\u7528 NCCL_IB_DISABLE=0\u3001NCCL_P2P_DISABLE=0 \u73af\u5883\u53d8\u91cf\uff1b\u7ed3\u5408 ZeRO++ \u68af\u5ea6\u91cf\u5316\uff0cBF16 \u68af\u5ea6\u538b\u7f29\u81f3 FP8 \u4f20\u8f93\uff0c\u901a\u4fe1\u6570\u636e\u91cf\u7f29\u51cf 50%\uff0c\u7cbe\u5ea6\u635f\u5931\u63a7\u5236\u5728 0.3% \u4ee5\u5185\u3002<\/li>\n\n\n\n
  4. \u5e76\u884c\u7b56\u7565\u5c42<\/strong>\uff1aMoE \u6df7\u5408\u4e13\u5bb6\u6a21\u578b\u91c7\u7528 Hybrid-EP \u901a\u4fe1\u903b\u8f91\uff0c intra-node NVLink\u3001inter-node RDMA \u5206\u5c42\u901a\u4fe1\uff0c128 \u5361\u96c6\u7fa4\u541e\u5410\u91cf\u63d0\u5347 45%\u3002<\/li>\n<\/ol>\n\n\n\n

    \u4e09\u3001\u843d\u5730\u5b9e\u64cd\u7ecf\u9a8c\u5206\u4eab\uff1aRDMA \u96c6\u7fa4\u6807\u51c6\u5316\u90e8\u7f72\u6d41\u7a0b\uff08\u661f\u5b87\u667a\u7b97\u4ea4\u4ed8\u6807\u51c6\uff09<\/h2>\n\n\n\n

    3.1 \u56db\u9636\u6bb5\u6807\u51c6\u5316\u90e8\u7f72\u5de5\u65f6\u6570\u636e<\/h3>\n\n\n\n

    \u661f\u5b87\u667a\u7b97\u6c89\u6dc0\u53ef\u590d\u7528 RDMA \u96c6\u7fa4\u4ea4\u4ed8\u6d41\u7a0b\uff0c\u652f\u6301\u6309\u9700\u5f39\u6027\u79df\u8d41 8~1024 \u5361\u5206\u5e03\u5f0f\u7b97\u529b\uff0c\u5404\u73af\u8282\u5b9e\u6d4b\u8017\u65f6\u5982\u4e0b\uff1a<\/p>\n\n\n\n

      \n
    1. \u786c\u4ef6\u94fe\u8def\u6821\u9a8c\uff1a\u5185\u7f6e\u81ea\u7814 RDMA \u8fde\u901a\u6027\u68c0\u6d4b\u5de5\u5177\uff0c\u6279\u91cf\u8282\u70b9\u4e92\u901a\u6d4b\u8bd5\u226430 \u5206\u949f\uff0c\u81ea\u52a8\u8bc6\u522b\u5149\u6a21\u5757\u8870\u51cf\u3001\u7f51\u7ebf\u9519\u5e8f\u6545\u969c\u3002<\/li>\n\n\n\n
    2. RDMA \u534f\u8bae\u6808\u6279\u91cf\u90e8\u7f72\uff1a\u4e3b\u8282\u70b9\u955c\u50cf\u5206\u53d1\u9a71\u52a8\u3001rdma-core\u3001perftest \u5957\u4ef6\uff0c8 \u8282\u70b9\u96c6\u7fa4\u5168\u91cf\u914d\u7f6e\u8017\u65f6\u226490 \u5206\u949f\uff1b\u591a\u79df\u6237\u73af\u5883\u9694\u79bb QP \u961f\u5217\uff0c\u907f\u514d\u4efb\u52a1\u62a2\u5360\u7f51\u7edc\u8d44\u6e90\u3002<\/li>\n\n\n\n
    3. \u8c03\u5ea6\u4e0e\u901a\u4fe1\u7ec4\u4ef6\u9002\u914d\uff1a\u9884\u88c5 Slurm\/K8s \u8c03\u5ea6\u3001NCCL\u3001UCCL-EP \u901a\u4fe1\u5e93\uff0c\u53ef\u89c6\u5316\u63a7\u5236\u53f0\u4e00\u952e\u6302\u8f7d\u9ad8\u901f\u5171\u4eab\u5b58\u50a8\uff0c\u5b58\u50a8\u8bfb\u5199\u5e26\u5bbd\u7a33\u5b9a 650MB\/s\u3002<\/li>\n\n\n\n
    4. \u57fa\u51c6\u6027\u80fd\u538b\u6d4b\uff1a\u6267\u884c ib_send_lat\u3001all_reduce_perf \u81ea\u52a8\u5316\u8dd1\u5206\uff0c\u8f93\u51fa\u5ef6\u8fdf\u3001\u5e26\u5bbd\u3001\u4e22\u5305\u62a5\u544a\uff0c\u4ea4\u4ed8\u6807\u51c6\u4e3a\u5355\u8282\u70b9 AllReduce \u6548\u7387\u226594%\u3002<\/li>\n<\/ol>\n\n\n\n

      3.2 \u9ad8\u9891\u6545\u969c\u4e0e\u6807\u51c6\u5316\u89e3\u51b3\u65b9\u6848<\/h3>\n\n\n\n
        \n
      1. \u6545\u969c 1\uff1a\u591a\u5361\u8bad\u7ec3\u51fa\u73b0 NCCL \u8d85\u65f6\uff0cRDMA \u94fe\u8def\u95f4\u65ad\u4e22\u5305 \u6839\u56e0\uff1aDCQCN \u53c2\u6570\u4e0d\u5339\u914d\u3001\u4ea4\u6362\u673a PFC \u6d41\u63a7\u672a\u5f00\u542f\uff1b\u89e3\u51b3\u65b9\u6848\uff1a\u661f\u5b87\u667a\u7b97\u9884\u7f6e\u8c03\u4f18\u6a21\u677f\uff0c\u81ea\u52a8\u4e0b\u53d1\u4ea4\u6362\u673a\u6d41\u63a7\u914d\u7f6e\uff0c2 \u5206\u949f\u5b8c\u6210\u4fee\u590d\u3002<\/li>\n\n\n\n
      2. \u6545\u969c 2\uff1a\u5c0f\u6279\u91cf\u8bad\u7ec3\u5ef6\u8fdf\u6ce2\u52a8\u5927 \u6839\u56e0\uff1aQP \u961f\u5217\u8d44\u6e90\u5206\u914d\u4e0d\u8db3\uff1b\u89e3\u51b3\u65b9\u6848\uff1a\u811a\u672c\u6279\u91cf\u8c03\u5927 ibv_qp_max_wr \u53c2\u6570\uff0c\u9884\u6ce8\u518c\u5168\u5c40\u901a\u4fe1\u7f13\u51b2\u533a\u3002<\/li>\n\n\n\n
      3. \u6545\u969c 3\uff1a\u8de8\u673a\u67b6\u8282\u70b9\u5e26\u5bbd\u8870\u51cf 20% \u6839\u56e0\uff1a\u53f6\u810a\u62d3\u6251 ECMP \u8d1f\u8f7d\u4e0d\u5747\uff1b\u89e3\u51b3\u65b9\u6848\uff1a\u542f\u7528\u96c6\u4e2d\u6d41\u91cf\u8c03\u5ea6 TE \u7b97\u6cd5\uff0c\u94fe\u8def\u5229\u7528\u7387\u5747\u8861\u7ef4\u6301 80%\uff0c\u96c6\u5408\u901a\u4fe1\u6027\u80fd\u63d0\u5347 7%~11%\u3002<\/li>\n<\/ol>\n\n\n\n

        \u56db\u3001RDMA \u8fd0\u7ef4\u4e0e\u6027\u80fd\u8c03\u4f18\u5de5\u5177\u6808\u5b8c\u6574\u6e05\u5355<\/h2>\n\n\n\n

        \u5206\u4e3a\u786c\u4ef6\u8bca\u65ad\u3001\u901a\u4fe1\u538b\u6d4b\u3001\u8bad\u7ec3\u76d1\u63a7\u4e09\u7c7b\uff0c\u5747\u5728\u661f\u5b87\u667a\u7b97\u96c6\u7fa4\u9884\u88c5\uff0c\u65e0\u9700\u7528\u6237\u4e8c\u6b21\u7f16\u8bd1\uff1a<\/p>\n\n\n\n

        4.1 \u786c\u4ef6\u4e0e RDMA \u57fa\u7840\u8bca\u65ad\u5de5\u5177<\/h3>\n\n\n\n
          \n
        1. perftest \u5957\u4ef6\uff1aib_send_bw\uff08\u5e26\u5bbd\u6d4b\u8bd5\uff09\u3001ib_send_lat\uff08\u5ef6\u8fdf\u6d4b\u8bd5\uff09\uff0c\u9a8c\u8bc1\u5355\u8282\u70b9\u3001\u8de8\u8282\u70b9\u88f8\u7f51\u6027\u80fd\uff1b<\/li>\n\n\n\n
        2. infiniband-diags\uff1aibstat\u3001ibping\uff0c\u5b9a\u4f4d\u7f51\u5361\u3001\u4ea4\u6362\u673a\u94fe\u8def\u786c\u4ef6\u6545\u969c\uff1b<\/li>\n\n\n\n
        3. ofed_info\uff1a\u6821\u9a8c OFED \u9a71\u52a8\u7248\u672c\u3001rdma-core \u5e93\u5b8c\u6574\u6027\u3002<\/li>\n<\/ol>\n\n\n\n

          4.2 \u5206\u5e03\u5f0f\u8bad\u7ec3\u901a\u4fe1\u538b\u6d4b\u5de5\u5177<\/h3>\n\n\n\n
            \n
          1. NCCL test\uff1aall_reduce_perf\uff0c\u6a21\u62df\u771f\u5b9e\u5927\u6a21\u578b\u68af\u5ea6\u540c\u6b65\uff0c\u8f93\u51fa\u4e0d\u540c\u6d88\u606f\u5c3a\u5bf8\u4e0b\u901a\u4fe1\u8017\u65f6\uff1b<\/li>\n\n\n\n
          2. UCCL-EP\uff1a\u4e13\u5bb6\u5e76\u884c\u901a\u4fe1\u57fa\u51c6\u6d4b\u8bd5\uff0c\u9002\u914d MoE \u7a00\u758f\u8bad\u7ec3\u573a\u666f\uff1b<\/li>\n\n\n\n
          3. \u661f\u5b87\u667a\u7b97\u81ea\u7814 RDMA \u5de1\u68c0\u5e73\u53f0\uff1a\u5b9e\u65f6\u91c7\u96c6\u7f51\u5361\u541e\u5410\u3001QP \u9519\u8bef\u8ba1\u6570\u3001PFC \u6682\u505c\u4e8b\u4ef6\uff0c\u5f02\u5e38\u6307\u6807\u89e6\u53d1\u544a\u8b66\u3002<\/li>\n<\/ol>\n\n\n\n

            4.3 \u96c6\u7fa4\u5168\u94fe\u8def\u76d1\u63a7\u5de5\u5177<\/h3>\n\n\n\n

            Prometheus+Grafana \u5b9a\u5236\u76d1\u63a7\u770b\u677f\uff0c\u6307\u6807\u5305\u542b\uff1a\u5355\u5361\u8fed\u4ee3\u8017\u65f6\u3001RDMA \u5e73\u5747 p99 \u5ef6\u8fdf\u3001\u94fe\u8def\u4e22\u5305\u7387\u3001GPU \u7b97\u529b\u5229\u7528\u7387\u3001CPU \u534f\u8bae\u6808\u5360\u7528\uff0c\u652f\u6301\u6309\u8bad\u7ec3\u4efb\u52a1\u3001\u79df\u6237\u5206\u7ec4\u7b5b\u9009\u6570\u636e\u3002<\/p>\n\n\n\n

            \u4e94\u3001\u56e2\u961f\u534f\u4f5c\u3001\u7ba1\u7406\u4f53\u7cfb\u4e0e\u5de5\u7a0b\u804c\u4e1a\u5fc3\u5f97<\/h2>\n\n\n\n

            5.1 RDMA \u7b97\u529b\u96c6\u7fa4\u8fd0\u7ef4\u56e2\u961f\u5206\u5de5\u6807\u51c6<\/h3>\n\n\n\n

            \u9762\u5411 AI \u8bad\u7ec3\u3001\u7f51\u7edc\u3001\u7b97\u529b\u8c03\u5ea6\u4e09\u7c7b\u5c97\u4f4d\uff0c\u661f\u5b87\u667a\u7b97\u843d\u5730\u6807\u51c6\u5316\u534f\u4f5c\u673a\u5236\uff1a<\/p>\n\n\n\n

              \n
            1. \u7f51\u7edc\u5de5\u7a0b\u5e08\uff1a\u8d1f\u8d23 RoCE \u7ec4\u7f51\u3001\u4ea4\u6362\u673a DCQCN \u53c2\u6570\u3001\u786c\u4ef6\u94fe\u8def\u6545\u969c\u5904\u7406\uff1b\u8f93\u51fa\u6708\u5ea6\u7f51\u7edc\u6027\u80fd\u57fa\u7ebf\u62a5\u544a\u3002<\/li>\n\n\n\n
            2. AI \u7b97\u6cd5\u8fd0\u7ef4\u5de5\u7a0b\u5e08\uff1aNCCL \u9002\u914d\u3001ZeRO \u5e76\u884c\u7b56\u7565\u3001\u8bad\u7ec3\u6846\u67b6\u73af\u5883\u8c03\u4f18\uff0c\u627f\u63a5\u5ba2\u6237\u5206\u5e03\u5f0f\u8bad\u7ec3\u4efb\u52a1\u6392\u969c\u3002<\/li>\n\n\n\n
            3. \u7b97\u529b\u8c03\u5ea6\u5de5\u7a0b\u5e08\uff1aSlurm\/K8s \u591a\u79df\u6237\u9694\u79bb\u3001\u5f39\u6027\u6269\u5bb9\u3001\u79df\u8d41\u8d44\u6e90\u6210\u672c\u6838\u7b97\uff0c\u4f18\u5316\u4efb\u52a1\u6392\u961f\u7b49\u5f85\u65f6\u957f\u3002 \u534f\u4f5c\u6d41\u7a0b\uff1a\u5ba2\u6237\u8bad\u7ec3\u5ef6\u8fdf\u5f02\u5e38\u5de5\u5355\uff0c10 \u5206\u949f\u5185\u7f51\u7edc + \u7b97\u6cd5\u8fd0\u7ef4\u53cc\u5c97\u534f\u540c\u5b9a\u4f4d\uff0c\u533a\u5206\u786c\u4ef6\u7f51\u7edc\u95ee\u9898\u4e0e\u6846\u67b6\u5e76\u884c\u7b56\u7565\u95ee\u9898\uff0c\u95ed\u73af\u65f6\u6548\u22642 \u5c0f\u65f6\u3002<\/li>\n<\/ol>\n\n\n\n

              5.2 \u5de5\u7a0b\u843d\u5730\u804c\u4e1a\u5fc3\u5f97<\/h3>\n\n\n\n
                \n
              1. \u4f18\u5148\u6807\u51c6\u5316\uff0c\u964d\u4f4e\u8bd5\u9519\u6210\u672c\uff1aRDMA \u8c03\u4f18\u53c2\u6570\u3001\u90e8\u7f72\u811a\u672c\u3001\u6545\u969c\u5904\u7406\u6a21\u677f\u7edf\u4e00\u6c89\u6dc0\uff0c\u907f\u514d\u5355\u4eba\u7ecf\u9a8c\u65ad\u5c42\uff1b\u661f\u5b87\u667a\u7b97\u5c06 300 + \u96c6\u7fa4\u843d\u5730\u7ecf\u9a8c\u5c01\u88c5\u4e3a\u955c\u50cf\uff0c\u65b0\u96c6\u7fa4\u4ea4\u4ed8\u6548\u7387\u63d0\u5347 60%\u3002<\/li>\n\n\n\n
              2. \u4ee5\u91cf\u5316\u6570\u636e\u66ff\u4ee3\u4e3b\u89c2\u5224\u65ad\uff1a\u6240\u6709\u7f51\u7edc\u4f18\u5316\u52a8\u4f5c\u5fc5\u987b\u7559\u5b58\u8c03\u4f18\u524d\u540e\u5ef6\u8fdf\u3001\u541e\u5410\u3001\u7b97\u529b\u5229\u7528\u7387\u5bf9\u6bd4\u6570\u636e\uff0c\u4e0d\u4f9d\u8d56\u7ecf\u9a8c\u731c\u6d4b\u3002<\/li>\n\n\n\n
              3. \u79df\u8d41\u6a21\u5f0f\u5e73\u8861\u6210\u672c\u4e0e\u6027\u80fd\uff1a\u4e2d\u5c0f\u56e2\u961f\u81ea\u5efa RDMA \u786c\u4ef6\u95f2\u7f6e\u7387\u8d85 40%\uff0c\u6309\u9700\u79df\u8d41\u53ef\u5c06\u7b97\u529b\u56fa\u5b9a\u6295\u5165\u964d\u4f4e 75%\uff0c\u540c\u65f6\u4eab\u53d7\u6301\u7eed\u8fed\u4ee3\u7684\u7f51\u7edc\u8c03\u4f18\u8fd0\u7ef4\u670d\u52a1\u3002<\/li>\n\n\n\n
              4. \u957f\u671f\u4f18\u5316\u65b9\u5411\uff1a\u901a\u4fe1\u4e0e\u8ba1\u7b97\u91cd\u53e0\u8c03\u5ea6\uff0c\u901a\u8fc7 RDMA \u5f02\u6b65\u4f20\u8f93\u63a9\u76d6\u68af\u5ea6\u540c\u6b65\u8017\u65f6\uff0c\u8fdb\u4e00\u6b65\u91ca\u653e GPU \u7b97\u529b\u4e0a\u9650\u3002<\/li>\n<\/ol>\n\n\n\n

                \u516d\u3001\u843d\u5730\u4ef7\u503c\u603b\u7ed3<\/h2>\n\n\n\n

                \u57fa\u4e8e\u661f\u5b87\u667a\u7b97 RDMA \u79df\u8d41\u96c6\u7fa4\u5b9e\u6d4b\u6570\u636e\uff1a\u540c\u7b49\u53c2\u6570\u89c4\u6a21\u5927\u6a21\u578b\u8bad\u7ec3\uff0cRDMA \u7ec4\u7f51\u76f8\u6bd4\u4f20\u7edf\u4ee5\u592a\u7f51\uff0c\u8bad\u7ec3\u65f6\u957f\u7f29\u77ed 58%\uff0c\u5355\u4f4d token \u7b97\u529b\u79df\u8d41\u6210\u672c\u964d\u4f4e 42%\uff1b\u96c6\u7fa4\u89c4\u6a21\u6269\u5bb9\u81f3 512 \u5361\u65f6\uff0c\u5206\u5e03\u5f0f\u901a\u4fe1\u6548\u7387\u4ecd\u7ef4\u6301 82% \u4ee5\u4e0a\u7ebf\u6027\u6269\u5c55\u80fd\u529b\u3002<\/p>\n\n\n\n

                RDMA \u9ad8\u901f\u7f51\u7edc\u662f\u5927\u6a21\u578b\u5206\u5e03\u5f0f\u8bad\u7ec3\u4e0d\u53ef\u66ff\u4ee3\u7684\u5e95\u5c42\u57fa\u7840\u8bbe\u65bd\uff0c\u6807\u51c6\u5316\u79df\u8d41\u6a21\u5f0f\u53ef\u5927\u5e45\u964d\u4f4e AI \u4f01\u4e1a\u786c\u4ef6\u6295\u5165\u3001\u7f51\u7edc\u8fd0\u7ef4\u3001\u6280\u672f\u8c03\u4f18\u4e09\u91cd\u95e8\u69db\uff0c\u4f9d\u6258\u5206\u5c42\u8c03\u4f18\u65b9\u6848\u3001\u5b8c\u6574\u5de5\u5177\u6808\u4e0e\u6807\u51c6\u5316\u8fd0\u7ef4\u4f53\u7cfb\uff0c\u5b9e\u73b0\u4f4e\u5ef6\u8fdf\u3001\u9ad8\u541e\u5410\u3001\u9ad8\u7a33\u5b9a\u7684\u5343\u4ebf\u53c2\u6570\u6a21\u578b\u89c4\u6a21\u5316\u8bad\u7ec3\u843d\u5730\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

                \u4e00\u3001\u884c\u4e1a\u80cc\u666f\u4e0e\u6838\u5fc3\u75db\u70b9 \u5343\u4ebf\u3001\u4e07\u4ebf\u53c2\u6570\u5927\u6a21\u578b\u8bad\u7ec3\u4f9d\u8d56\u591a GP…<\/p>\n","protected":false},"author":2,"featured_media":7821,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-7820","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-zixun"],"views":29,"_links":{"self":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts\/7820","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/comments?post=7820"}],"version-history":[{"count":1,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts\/7820\/revisions"}],"predecessor-version":[{"id":7822,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts\/7820\/revisions\/7822"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/media\/7821"}],"wp:attachment":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/media?parent=7820"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/categories?post=7820"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/tags?post=7820"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}