{"id":3132,"date":"2024-09-06T16:35:53","date_gmt":"2024-09-06T08:35:53","guid":{"rendered":"http:\/\/blog.cloud500.net\/?p=3132"},"modified":"2024-10-09T16:26:17","modified_gmt":"2024-10-09T08:26:17","slug":"rdma-%e7%bd%91%e7%bb%9c%e4%b8%8e-ai-%e7%a0%94%e7%a9%b6%e6%8a%a5%e5%91%8a","status":"publish","type":"post","link":"http:\/\/blog.cloud500.ai\/?p=3132","title":{"rendered":"RDMA \u7f51\u7edc\u4e0e AI \u7814\u7a76\u62a5\u544a"},"content":{"rendered":"\n<h2 class=\"wp-block-heading has-text-align-center\"><strong>\u4ec0\u4e48\u662f RDMA<\/strong><\/h2>\n\n\n\n<p>RDMA \u6280\u672f\u5141\u8bb8\u4e24\u53f0\u670d\u52a1\u5668\u5728\u4e0d\u7ecf\u8fc7\u5904\u7406\u5668\u3001\u7f13\u5b58\u6216\u64cd\u4f5c\u7cfb\u7edf\u7684\u5e72\u9884\u4e0b\uff0c\u76f4\u63a5\u8bfb\u53d6\u6216\u5199\u5165\u5bf9\u65b9\u7684\u5185\u5b58\u3002\u8fd9\u9879\u6280\u672f\u80fd\u591f\u91ca\u653e CPU \u8d44\u6e90\uff0c\u964d\u4f4e\u5ef6\u8fdf\uff0c\u52a0\u5feb\u6570\u636e\u4f20\u8f93\u901f\u5ea6\uff0c\u4ece\u800c\u8ba9\u7f51\u7edc\u3001\u5b58\u50a8\u548c\u8ba1\u7b97\u5e94\u7528\u53d7\u76ca\u532a\u6d45\u3002RDMA \u4e3b\u8981\u901a\u8fc7\u670d\u52a1\u5668\u7684\u7f51\u7edc\u63a5\u53e3\u5361\uff08NIC\uff09\u5b9e\u73b0\uff0c\u5b83\u7ed5\u8fc7\u64cd\u4f5c\u7cfb\u7edf\u548c\u7f51\u7edc\u5185\u6838\uff0c\u663e\u8457\u63d0\u5347\u4e86\u4e24\u53f0\u670d\u52a1\u5668\u95f4\u7684\u7f51\u7edc\u6027\u80fd\u548c\u6570\u636e\u4ea4\u6362\u901f\u5ea6\u3002RDMA \u6700\u521d\u662f\u4e3a\u9ad8\u6027\u80fd\u8ba1\u7b97\u96c6\u7fa4\u8bbe\u8ba1\u7684\uff0c\u7279\u522b\u9002\u7528\u4e8e\u9700\u8981\u5927\u89c4\u6a21\u5e76\u884c\u5904\u7406\u7684\u8ba1\u7b97\u4efb\u52a1\u3002\u8fd9\u79cd\u5927\u89c4\u6a21\u5e76\u884c\u8ba1\u7b97\u6b63\u662f AI\/ML \u6280\u672f\u7684\u57fa\u7840\uff0cRDMA \u5728\u8fd9\u4e9b\u9886\u57df\u7684\u5e94\u7528\u90e8\u7f72\u4e2d\u626e\u6f14\u7740\u81f3\u5173\u91cd\u8981\u7684\u89d2\u8272\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0cRDMA \u6280\u672f\u6700\u521d\u662f\u5728 InfiniBand \u7f51\u7edc\u4e0a\u5f00\u53d1\u7684\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading has-text-align-center\"><strong>\u4ec0\u4e48\u662f\u57fa\u4e8e\u878d\u5408\u4ee5\u592a\u7f51\u7684 RDMA (RoCE)<\/strong><\/h2>\n\n\n\n<p>\u57fa\u4e8e\u878d\u5408\u4ee5\u592a\u7f51\u6280\u672f\u7684 RDMA\uff08Remote Direct Memory Access over Converged Ethernet\uff0c\u7b80\u79f0RoCE\uff09\u662f\u4e00\u79cd\u7f51\u7edc\u534f\u8bae\uff0c\u5b83\u5141\u8bb8\u5728\u4ee5\u592a\u7f51\u7f51\u7edc\u4e0a\u5b9e\u73b0\u9ad8\u6548\u7684\u76f4\u63a5\u5185\u5b58\u8bbf\u95ee\u4ea4\u6362\u3002RoCE \u7684\u65e9\u671f\u7248\u672c\u4f9d\u8d56\u4e8e\u878d\u5408\u4ee5\u592a\u7f51\u67b6\u6784\uff0c\u4f46\u968f\u7740\u6280\u672f\u7684\u53d1\u5c55\uff0c\u6700\u65b0\u7684 RoCE \u7248\u672c\u5df2\u80fd\u591f\u5728\u6807\u51c6\u7684\u4ee5\u592a\u7f51\u7f51\u7edc\u4e0a\u8fd0\u884c\u3002RoCE \u534f\u8bae\u5728\u4ee5\u592a\u7f51\u4e4b\u4e0a\u8fd0\u884c\uff0c\u9700\u8981\u4ee5\u592a\u7f51\u7f51\u7edc\u63a5\u53e3\u5361\uff08NIC\uff09\u7684\u652f\u6301\u6765\u5b8c\u6210\u6570\u636e\u4f20\u8f93\u3002\u5728\u884c\u4e1a\u5c42\u9762\uff0c\u4e1a\u754c\u6b63\u5728\u5bf9\u4ee5\u592a\u7f51\u6280\u672f\u8fdb\u884c\u5927\u91cf\u6295\u8d44\uff0c\u4ee5\u6539\u5584\u62e5\u585e\u63a7\u5236\u673a\u5236\u5e76\u51cf\u5c11\u6570\u636e\u6d41\u91cf\u7684\u635f\u5931\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u6570\u636e\u4e2d\u5fc3\u5b89\u88c5\u7684\u4ee5\u592a\u7f51\u4ea4\u6362\u673a\u7aef\u53e3\u6570\u91cf\u5df2\u8d85\u8fc7 4 \u4ebf\u4e2a\u3002\u9274\u4e8e\u4ee5\u592a\u7f51\u7684\u666e\u53ca\u6027\uff0c\u5b83\u5c06\u5728 AI\/ML \u7f51\u7edc\u4e2d\u53d1\u6325\u8d8a\u6765\u8d8a\u91cd\u8981\u7684\u4f5c\u7528\u3002\u968f\u7740\u6280\u672f\u7684\u4e0d\u65ad\u8fdb\u6b65\uff0c\u9884\u8ba1\u672a\u6765\u5c06\u6709\u66f4\u591a\u7684 RDMA \u64cd\u4f5c\u901a\u8fc7\u4ee5\u592a\u7f51\u6765\u6267\u884c\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading has-text-align-center\"><strong>\u4ecb\u7ecd<\/strong><\/h2>\n\n\n\n<p>\u5728 2022 \u5e74\u4e4b\u524d\uff0cRDMA \u6280\u672f\u4e3b\u8981\u5e94\u7528\u4e8e\u9ad8\u6027\u80fd\u8ba1\u7b97\uff08HPC\uff09\u9886\u57df\uff0c\u7279\u522b\u662f\u5728\u8d85\u7b97\u9879\u76ee\u4e2d\u5f97\u5230\u5e7f\u6cdb\u5e94\u7528\uff0c\u800c\u5728\u4e91\u548c\u4f01\u4e1a\u6570\u636e\u4e2d\u5fc3\u4e2d\u7684\u5e94\u7528\u76f8\u5bf9\u8f83\u5c11\u3002\u7136\u800c\uff0c\u968f\u7740 AI \u548c ML \u5728 2022 \u5e74\u5e95\u6210\u4e3a\u6295\u8d44\u7684\u7126\u70b9\uff0c\u6570\u636e\u4e2d\u5fc3\u7684\u652f\u51fa\u8fc5\u901f\u4ece\u4f20\u7edf\u8ba1\u7b97\u8f6c\u5411\u4e86 AI\/ML \u9886\u57df\u3002\u8fd9\u79cd\u8f6c\u53d8\u5e26\u6765\u7684\u589e\u957f\u662f\u7a7a\u524d\u7684\uff0c\u5230 2023 \u5e74\u5e95\uff0cRDMA \u57fa\u7840\u7f51\u7edc\u7684\u589e\u957f\u901f\u5ea6\u751a\u81f3\u8d85\u8fc7\u4e86 2022 \u5e74\u548c 2021 \u5e74\u7684\u603b\u548c\u3002\u8fd9\u4e00\u8fc5\u731b\u7684\u589e\u957f\u4f7f RDMA \u6210\u4e3a\u4e86\u4e3b\u6d41\uff0c\u5e76\u5bf9 AI\/ML \u7684\u6269\u5c55\u81f3\u5173\u91cd\u8981\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading has-text-align-center\"><strong>\u670d\u52a1\u5668\u5e02\u573a\u7684\u53d8\u5316<\/strong><\/h2>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full\"><img decoding=\"async\" loading=\"lazy\" width=\"896\" height=\"844\" src=\"http:\/\/blog.cloud500.ai\/wp-content\/uploads\/2024\/10\/640-4.png\" alt=\"\" class=\"wp-image-3176\"\/><\/figure><\/div>\n\n\n<p>\u968f\u7740\u5ba2\u6237\u9010\u6e10\u4ece\u901a\u7528\u670d\u52a1\u5668\u8f6c\u5411 AI\/ML \u670d\u52a1\u5668\uff0c\u670d\u52a1\u5668\u5e02\u573a\u6b63\u5728\u7ecf\u5386\u663e\u8457\u7684\u53d8\u9769\u3002\u9884\u8ba1\u5728\u8fd9\u4e00\u8f6c\u578b\u8fc7\u7a0b\u4e2d\uff0cAI\/ML \u670d\u52a1\u5668\u7684\u6570\u91cf\u5c06\u4ece 2023 \u5e74\u7684 100&nbsp;\u4e07\u53f0\u589e\u957f\u5230 2028 \u5e74\u7684 600&nbsp;\u4e07\u53f0\uff08\u53c2\u89c1\u56fe 1\uff09\u3002\u4e0e\u6b64\u540c\u65f6\uff0cAI\/ML \u670d\u52a1\u5668\u7684\u652f\u51fa\u9884\u8ba1\u5c06\u5728 2028 \u5e74\u63a5\u8fd1 3,000 \u4ebf\u7f8e\u5143\uff08\u53c2\u89c1\u56fe 2\uff09\u3002\u6570\u636e\u4e2d\u5fc3\u5e02\u573a\u4ece\u672a\u7ecf\u5386\u8fc7\u5982\u6b64\u8fc5\u901f\u4e14\u663e\u8457\u7684\u589e\u957f\uff0c\u800c AI\/ML \u65e0\u7591\u5c06\u6210\u4e3a\u63a8\u52a8\u6570\u636e\u4e2d\u5fc3\u652f\u51fa\u7ffb\u500d\u7684\u4e3b\u8981\u9a71\u52a8\u529b\u3002\u9884\u8ba1\u5230 2028 \u5e74\uff0c\u8fd9 600 \u4e07\u53f0 AI\/ML \u670d\u52a1\u5668\u5c06\u914d\u5907\u540e\u7aef\u7f51\u7edc\u6216 AI Fabric\uff0c\u5b9e\u73b0\u8ba1\u7b97\u8d44\u6e90\u7684\u9ad8\u6548\u4e92\u8054\u4e92\u901a\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img decoding=\"async\" loading=\"lazy\" width=\"836\" height=\"810\" src=\"http:\/\/blog.cloud500.ai\/wp-content\/uploads\/2024\/09\/image-15.png\" alt=\"\" class=\"wp-image-3134\"\/><\/figure>\n\n\n\n<p>\u968f\u7740\u6280\u672f\u7684\u4e0d\u65ad\u8fdb\u6b65\uff0c\u670d\u52a1\u5668\u4e2d\u642d\u8f7d\u7684 GPU \u548c AI ASIC \u6570\u91cf\u9884\u8ba1\u5c06\u6301\u7eed\u589e\u957f\u3002\u76ee\u524d\uff0c\u914d\u5907 8 \u4e2a GPU \u7684\u670d\u52a1\u5668\u914d\u7f6e\u6700\u4e3a\u666e\u904d\uff0c\u4f46\u9884\u8ba1\u5230\u5341\u5e74\u540e\uff0c\u5e02\u573a\u5c06\u6269\u5c55\u81f3 16 \u4e2a\u751a\u81f3 32 \u4e2a GPU \u7684\u914d\u7f6e\u3002\u968f\u7740\u8bad\u7ec3\u6a21\u578b\u7684\u53c2\u6570\u91cf\u4ece\u6570\u5341\u4ebf\u6269\u5c55\u81f3\u6570\u4e07\u4ebf\uff0c\u6bcf\u4e2a GPU \u6240\u9700\u7684\u5185\u5b58\u5bb9\u91cf\u4e5f\u5c06\u76f8\u5e94\u589e\u52a0\u3002\u5728\u8fd9\u6837\u7684\u89c4\u6a21\u4e0b\uff0c\u670d\u52a1\u5668\u95f4\u8fdb\u884c\u9ad8\u6548\u6570\u636e\u4f20\u8f93\u7684\u80fd\u529b\u53d8\u5f97\u5c24\u4e3a\u5173\u952e\uff0c\u5b83\u662f\u5b9e\u73b0\u548c\u6269\u5c55\u8fd9\u4e9b\u5e9e\u5927\u8bad\u7ec3\u6a21\u578b\u7684\u57fa\u7840\u3002RDMA \u51ed\u501f\u5176\u9ad8\u6548\u7684\u8fdc\u7a0b\u76f4\u63a5\u5185\u5b58\u8bbf\u95ee\u80fd\u529b\uff0c\u5728\u8fd9\u4e9b\u9ad8\u6027\u80fd\u670d\u52a1\u5668\u4e2d\u626e\u6f14\u7740\u81f3\u5173\u91cd\u8981\u7684\u89d2\u8272\u3002\u7279\u522b\u662f\u5728\u9700\u8981\u8de8\u670d\u52a1\u5668\u8fdb\u884c\u5185\u5b58\u548c\u8d44\u6e90\u8bbf\u95ee\u7684\u573a\u666f\u4e2d\uff0cRDMA \u7684\u4f5c\u7528\u5c24\u4e3a\u7a81\u51fa\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading has-text-align-center\"><strong>\u4f5c\u4e1a\u5b8c\u6210\u65f6\u95f4\uff08JCT\uff09\u548c\u6027\u80fd\u6307\u6807<\/strong><\/h2>\n\n\n\n<p>\u76f4\u63a5\u8bbf\u95ee\u5176\u4ed6\u670d\u52a1\u5668\u7684\u5185\u5b58\u6709\u52a9\u4e8e\u63d0\u5347AI\u6a21\u578b\u7684\u6574\u4f53\u6027\u80fd\u3002RDMA \u901a\u8fc7\u5feb\u901f\u5c06\u6570\u636e\u4f20\u8f93\u5230 GPU\uff0c\u6709\u6548\u5730\u7f29\u77ed\u4e86\u4f5c\u4e1a\u5b8c\u6210\u65f6\u95f4\uff08Job Completion Time, JCT\uff09\u3002\u5728\u65e9\u671f\u7684 AI\/ML \u96c6\u7fa4\u4e2d\uff0cGPU \u6838\u5fc3\u7684\u95f2\u7f6e\u662f\u4e00\u4e2a\u4e3b\u8981\u95ee\u9898\uff0c\u8fd9\u53ef\u80fd\u5bfc\u81f4\u6574\u4e2a AI\/ML \u96c6\u7fa4\u56e0\u4e3a\u6570\u636e\u5305\u4e22\u5931\u6216\u6570\u636e\u4f20\u8f93\u5ef6\u8fdf\u800c\u6548\u7387\u4f4e\u4e0b\u3002RDMA \u901a\u8fc7\u89e3\u51b3\u8fd9\u4e9b\u7f51\u7edc\u74f6\u9888\uff0c\u663e\u8457\u63d0\u9ad8\u4e86 JCT \u548c AI\/ML \u96c6\u7fa4\u7684\u6574\u4f53\u6027\u80fd\u3002\u867d\u7136\u4ee5\u592a\u7f51\u548c InfiniBand \u5728\u6027\u80fd\u4e0a\u53ef\u80fd\u5b58\u5728\u5dee\u5f02\uff0c\u4f46 RDMA \u76f8\u8f83\u4e8e\u4f20\u7edf\u7684\u7f51\u7edc\u89e3\u51b3\u65b9\u6848\uff0c\u65e0\u7591\u662f\u4e00\u4e2a\u663e\u8457\u7684\u8fdb\u6b65\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading has-text-align-center\"><strong>\u7f51\u7edc\u63a5\u53e3\u5361\uff08NIC\uff09\u5e02\u573a<\/strong><\/h2>\n\n\n\n<p>\u6240\u6709 InfiniBand NIC \u90fd\u652f\u6301 RDMA\uff0c\u4f46\u5e76\u975e\u6240\u6709\u73b0\u6709\u7684\u4ee5\u592a\u7f51 NIC \u90fd\u5177\u5907 RDMA \u6216 RoCE \u529f\u80fd\u3002\u5bf9\u4e8e\u90a3\u4e9b\u5e0c\u671b\u62d3\u5c55\u5230 AI\/ML \u5e02\u573a\u7684\u4f20\u7edf\u4ee5\u592a\u7f51 NIC \u4f9b\u5e94\u5546\u6765\u8bf4\uff0c\u96c6\u6210 RoCE \u529f\u80fd\u81f3\u5176\u4ea7\u54c1\u4e2d\u662f\u5173\u952e\uff0c\u8fd9\u6837\u624d\u80fd\u4e0e\u5df2\u7ecf\u63d0\u4f9b RoCE \u652f\u6301\u7684\u7ade\u4e89\u5bf9\u624b\u76f8\u6297\u8861\u3002\u968f\u7740 NIC \u7684\u4f20\u8f93\u901f\u5ea6\u63d0\u5347\u81f3 400 Gbps \u751a\u81f3\u66f4\u9ad8\uff0c\u9884\u8ba1\u5927\u591a\u6570\u4ee5\u592a\u7f51 NIC \u4ea7\u54c1\u5c06\u5f00\u59cb\u652f\u6301 RoCE\u3002\u968f\u7740\u529f\u80fd\u7684\u589e\u5f3a\u548c\u7aef\u53e3\u901f\u5ea6\u7684\u63d0\u9ad8\uff0c\u9884\u8ba1\u4ee5\u592a\u7f51 NIC \u7684\u5e73\u5747\u552e\u4ef7\u4e5f\u5c06\u968f\u4e4b\u4e0a\u5347\u3002<\/p>\n\n\n\n<p>Ethernet NIC \u4e2d RoCE \u7684\u6027\u80fd\u8868\u73b0\u56e0\u4f9b\u5e94\u5546\u6240\u91c7\u7528\u7684\u5904\u7406\u5668\u7c7b\u578b\u3001\u9644\u52a0\u7684\u5378\u8f7d\u5f15\u64ce\u4ee5\u53ca\u5de5\u7a0b\u6280\u672f\u7684\u4e0d\u540c\u800c\u6709\u6240\u5dee\u5f02\uff0c\u4ece\u800c\u5f62\u6210\u4e0d\u540c\u7684\u6027\u80fd\u5c42\u7ea7\u3002\u968f\u7740\u65b0\u4e00\u4ee3\u4ea7\u54c1\u7684\u63a8\u51fa\uff0c\u4f9b\u5e94\u5546\u4eec\u6b63\u81f4\u529b\u4e8e\u4e0d\u65ad\u4f18\u5316\u5176\u4ea7\u54c1\u6027\u80fd\uff0c\u4ee5\u5b9e\u73b0 RoCE \u6027\u80fd\u6307\u6807\u5728\u4e0d\u540c\u4f9b\u5e94\u5546\u4e4b\u95f4\u7684\u66f4\u9ad8\u4e00\u81f4\u6027\u3002\u8fd9\u79cd\u6027\u80fd\u4e0a\u7684\u4e00\u81f4\u6027\u4e0d\u4ec5\u4fc3\u8fdb\u4e86\u4ea7\u54c1\u95f4\u7684\u4e92\u64cd\u4f5c\u6027\uff0c\u800c\u4e14\u4e3a\u7ec8\u7aef\u7528\u6237\u63d0\u4f9b\u4e86\u66f4\u5e7f\u6cdb\u7684\u9009\u62e9\u7a7a\u95f4\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading has-text-align-center\"><strong>AI\/ML \u540e\u7aef\u7f51\u7edc<\/strong><\/h2>\n\n\n\n<p>\u5927\u591a\u6570 AI\/ML \u670d\u52a1\u5668\u90fd\u914d\u5907\u4e86\u72ec\u7acb\u7684\u540e\u7aef\u7f51\u7edc\uff0c\u8fd9\u79cd\u7f51\u7edc\u914d\u7f6e\u72ec\u7acb\u4e8e\u6570\u636e\u4e2d\u5fc3\u7684\u5176\u4ed6\u7ec4\u6210\u90e8\u5206\uff0c\u53ef\u4ee5\u57fa\u4e8e InfiniBand \u6216\u4ee5\u592a\u7f51\u6280\u672f\u6784\u5efa\u3002\u540e\u7aef\u7f51\u7edc\u7684\u8bbe\u8ba1\u4e13\u6ce8\u4e8e\u5f3a\u5316 AI\/ML \u96c6\u7fa4\u5185\u90e8\u7684\u8fde\u63a5\uff0c\u5c24\u5176\u662f GPU \u4e4b\u95f4\u7684\u4e92\u8054\u4ee5\u53ca GPU \u4e0e\u5185\u5b58\u7684\u76f4\u63a5\u8fde\u63a5\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u540e\u7aef\u7f51\u7edc\u662f\u73b0\u6709\u7f51\u7edc\u7684\u6269\u5c55\uff0c\u5b83\u901a\u8fc7\u4e3a\u6bcf\u53f0\u670d\u52a1\u5668\u589e\u52a0\u66f4\u591a\u7684\u7aef\u53e3\uff0c\u663e\u8457\u63d0\u5347\u4e86\u7f51\u7edc\u6536\u5165\u7684\u673a\u4f1a\u3002<\/p>\n\n\n\n<p>AI\/ML \u7cfb\u7edf\u53ef\u4ee5\u914d\u7f6e\u591a\u4e2a\u540e\u7aef\u7f51\u7edc\uff0c\u4ee5\u4e13\u6ce8\u4e8e\u4e0d\u540c\u7684\u4efb\u52a1\u3002\u4f8b\u5982\uff0c\u5c3d\u7ba1 RDMA \u80fd\u591f\u5728\u4ee5\u592a\u7f51\u548c InfiniBand \u4e0a\u8fd0\u884c\uff0c\u4f46\u4e0d\u540c\u7684 GPU \u4f9b\u5e94\u5546\u6216 AI ASIC \u5236\u9020\u5546\u53ef\u80fd\u4f1a\u91c7\u7528\u989d\u5916\u7684\u7f51\u7edc\u8fde\u63a5\uff0c\u4ee5\u5b9e\u73b0\u66f4\u9ad8\u7684\u6027\u80fd\u89e3\u51b3\u65b9\u6848\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading has-text-align-center\"><strong>\u5e02\u573a\u89c4\u6a21<\/strong><\/h2>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full\"><img decoding=\"async\" loading=\"lazy\" width=\"909\" height=\"363\" src=\"http:\/\/blog.cloud500.ai\/wp-content\/uploads\/2024\/10\/640-5.png\" alt=\"\" class=\"wp-image-3179\"\/><\/figure><\/div>\n\n\n<p>\u5728 2021 \u5e74\u4e4b\u524d\uff0c\u5177\u5907 RDMA \u80fd\u529b\u7684\u5e02\u573a\u89c4\u6a21\u6bcf\u5e74\u5927\u7ea6\u5728 4 \u4ebf\u5230 7 \u4ebf\u7f8e\u5143\u4e4b\u95f4\uff0c\u4e3b\u8981\u7531\u9ad8\u6027\u80fd\u8ba1\u7b97\uff08HPC\uff09\u5e94\u7528\u6240\u63a8\u52a8\u3002\u7136\u800c\uff0c\u5230\u4e86 2023 \u5e74\uff0c\u968f\u7740 AI\/ML \u90e8\u7f72\u7684\u8fc5\u731b\u589e\u957f\uff0c\u5bf9 RDMA \u7684\u9700\u6c42\u5df2\u7ecf\u7a81\u7834 60\u00a0\u4ebf\u7f8e\u5143\u5927\u5173\u3002\u636e\u9884\u6d4b\uff0c\u8fd9\u4e00\u6570\u5b57\u5230 2028 \u5e74\u5c06\u8d85\u8fc7 220\u00a0\u4ebf\u7f8e\u5143\uff08\u89c1\u8868 1\u3001\u8868 2 \u548c\u56fe 3\u3001\u56fe 4\uff09\u3002\u968f\u7740\u8fd0\u8425\u5546\u589e\u52a0\u5bf9 AI\/ML \u7684\u8d44\u672c\u652f\u51fa\uff08CAPEX\uff09\u5e76\u8fdb\u884c\u672a\u6765\u89c4\u5212\uff0c\u9884\u8ba1\u672a\u6765\u51e0\u5e74 RDMA \u76f8\u5173\u7684\u9879\u76ee\u548c\u5e02\u573a\u9884\u671f\u5c06\u4f1a\u8fdb\u4e00\u6b65\u4e0a\u8c03\u3002<\/p>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full\"><img decoding=\"async\" loading=\"lazy\" width=\"830\" height=\"814\" src=\"http:\/\/blog.cloud500.ai\/wp-content\/uploads\/2024\/10\/640-6.png\" alt=\"\" class=\"wp-image-3180\"\/><\/figure><\/div>\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full\"><img decoding=\"async\" loading=\"lazy\" width=\"826\" height=\"800\" src=\"http:\/\/blog.cloud500.ai\/wp-content\/uploads\/2024\/10\/640-7.png\" alt=\"\" class=\"wp-image-3181\"\/><\/figure><\/div>\n\n\n<p>RDMA \u5e02\u573a\u4e3b\u8981\u5206\u4e3a\u4e24\u4e2a\u4e0d\u540c\u7684\u7ec6\u5206\u9886\u57df\u3002\u9996\u5148\u662f\u6280\u672f\u5c42\u9762\uff0c\u76ee\u524d RDMA \u6280\u672f\u4e3b\u8981\u90e8\u7f72\u5728InfiniBand \u4e0a\uff0c\u4f46\u9884\u8ba1\u672a\u6765\u5728\u4ee5\u592a\u7f51\u9886\u57df\u5c06\u51fa\u73b0\u66f4\u591a RDMA \u7f51\u7edc\u7684\u5e94\u7528\u3002\u7b2c\u4e8c\u4e2a\u7ec6\u5206\u9886\u57df\u662f\u7f51\u7edc\u63a5\u53e3\u5361\uff08NIC\uff09\u548c\u4ea4\u6362\u673a\u5e02\u573a\u3002\u5728 InfiniBand \u7f51\u7edc\u4e2d\uff0cNIC \u548c\u4ea4\u6362\u673a\u7684\u91c7\u8d2d\u901a\u5e38\u540c\u6b65\u8fdb\u884c\u3002\u7136\u800c\uff0c\u5728\u4ee5\u592a\u7f51\u7f51\u7edc\u4e2d\uff0c\u4ea4\u6362\u673a\u7684\u91c7\u8d2d\u5f80\u5f80\u7531\u4e0d\u540c\u7684\u56e2\u961f\u8d1f\u8d23\uff0c\u5176\u91c7\u8d2d\u8282\u594f\u4e0e NIC \u7684\u91c7\u8d2d\u4e5f\u4e0d\u5c3d\u76f8\u540c\u3002\u7531\u4e8e\u7ec4\u7ec7\u7ed3\u6784\u7684\u9650\u5236\uff0c\u5373\u4f7f\u5728 AI\/ML \u9886\u57df\uff0c\u670d\u52a1\u5668\u56e2\u961f\u4e0e\u7f51\u7edc\u56e2\u961f\u7684\u5206\u79bb\u73b0\u8c61\u4ecd\u5c06\u7ee7\u7eed\u5b58\u5728\u3002\u56e0\u6b64\uff0c\u5728\u672a\u6765\u6784\u5efa AI\/ML \u7f51\u7edc\u65f6\uff0c\u6bcf\u4e2a\u5ba2\u6237\u90fd\u5c06\u5bf9 NIC \u548c\u4ea4\u6362\u673a\u4f9b\u5e94\u5546\u6709\u81ea\u5df1\u72ec\u7279\u7684\u504f\u597d\uff0c\u8fd9\u4e9b\u504f\u597d\u53ef\u80fd\u4e0e\u4f20\u7edf\u8ba1\u7b97\u90e8\u7f72\u4e2d\u7684\u9009\u62e9\u6709\u6240\u533a\u522b\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading has-text-align-center\"><strong>\u7ed3\u8bba<\/strong><\/h2>\n\n\n\n<p>RDMA \u548c RoCE \u5728 AI\/ML \u7f51\u7edc\u4e2d\u81f3\u5173\u91cd\u8981\uff0c\u7f3a\u5c11\u8fd9\u4e24\u9879\u6280\u672f\uff0cAI\/ML \u7684\u53d1\u5c55\u901f\u5ea6\u5c06\u65e0\u6cd5\u6ee1\u8db3\u5e02\u573a\u7684\u9700\u6c42\u3002\u968f\u7740\u670d\u52a1\u5668\u5e02\u573a\u8fc5\u901f\u4ece\u4f20\u7edf\u8ba1\u7b97\u8f6c\u5411 AI\/ML\uff0c\u9884\u8ba1 RDMA \u548c RoCE \u7684\u5e02\u573a\u6f5c\u529b\u5c06\u5728\u672a\u6765\u51e0\u5e74\u663e\u8457\u589e\u957f\u3002\u5c3d\u7ba1\u5ba2\u6237\u548c\u4f9b\u5e94\u5546\u5bf9\u6280\u672f\u7684\u504f\u597d\u5404\u6709\u4e0d\u540c\uff0c\u4f46 RDMA \u6280\u672f\u4ecd\u5c06\u6301\u7eed\u7e41\u8363\u53d1\u5c55\u3002\u4ee5\u592a\u7f51\u548c InfiniBand \u8fd9\u4e24\u79cd\u6280\u672f\u5c06\u5728\u5e02\u573a\u4e0a\u5e76\u5b58\uff0c\u5ba2\u6237\u4e0d\u5e94\u5c06\u5176\u89c6\u4e3a\u4e8c\u5143\u9009\u62e9\u3002\u5927\u591a\u6570\u5ba2\u6237\u5c06\u5728 InfiniBand \u548c\u4ee5\u592a\u7f51\u4e0a\u4f7f\u7528 RDMA\uff0c\u5e76\u6df7\u5408\u4f7f\u7528\u4e0d\u540c\u7684 GPU\u3002\u53ea\u6709\u6781\u5c11\u6570\u5ba2\u6237\u4f1a\u5b8c\u5168\u4f9d\u8d56\u4e00\u79cd\u7f51\u7edc\u6216\u4f9b\u5e94\u5546\u3002AI\/ML \u7684\u5de5\u4f5c\u8d1f\u8f7d\u975e\u5e38\u591a\u6837\uff0c\u5305\u62ec\u57fa\u7840\u8bad\u7ec3\u3001\u5f3a\u5316\u5b66\u4e60\u548c\u63a8\u7406\u7b49\u3002RDMA \u5728\u591a\u79cd\u6280\u672f\u4e0a\u7684\u5e94\u7528\u6709\u52a9\u4e8e\u5ba2\u6237\u6269\u5c55\u5e76\u4e13\u6ce8\u4e8e AI\/ML \u5de5\u4f5c\u8d1f\u8f7d\uff0c\u800c\u975e\u5e95\u5c42\u7f51\u7edc\u3002<\/p>\n\n\n\n<p class=\"has-cyan-bluish-gray-color has-text-color has-small-font-size\"><em>\u4ee5\u4e0a\u5185\u5bb9\u7ffb\u8bd1\u81ea\u300aRDMA Networking and AI Research Report\u300b\uff0c\u5982\u9700\u539f\u6587\uff0c\u8bf7\u4e0e\u6211\u4eec\u8054\u7cfb\u3002<\/em><\/p>\n\n\n\n<p>WF Research \u662f\u4ee5\u7b2c\u4e00\u6027\u539f\u7406\u4e3a\u57fa\u7840\u7684\u4e13\u4e1a\u987e\u95ee\u670d\u52a1\u673a\u6784\uff0c\u6b22\u8fce\u5173\u6ce8\u548c\u7559\u8a00\uff01<\/p>\n\n\n\n<p>\u5fae\u4fe1\u53f7\uff1aAlexqjl<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u4ec0\u4e48\u662f RDMA RDMA \u6280\u672f\u5141\u8bb8\u4e24\u53f0\u670d\u52a1\u5668\u5728\u4e0d\u7ecf\u8fc7\u5904\u7406\u5668\u3001\u7f13\u5b58\u6216\u64cd\u4f5c\u7cfb\u7edf\u7684\u5e72\u9884\u4e0b\uff0c\u76f4\u63a5\u8bfb\u53d6\u6216\u5199\u5165\u5bf9\u65b9\u7684\u5185\u5b58 [&hellip;]<\/p>\n","protected":false},"author":2,"featured_media":3133,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"_links":{"self":[{"href":"http:\/\/blog.cloud500.ai\/index.php?rest_route=\/wp\/v2\/posts\/3132"}],"collection":[{"href":"http:\/\/blog.cloud500.ai\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/blog.cloud500.ai\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/blog.cloud500.ai\/index.php?rest_route=\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"http:\/\/blog.cloud500.ai\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=3132"}],"version-history":[{"count":4,"href":"http:\/\/blog.cloud500.ai\/index.php?rest_route=\/wp\/v2\/posts\/3132\/revisions"}],"predecessor-version":[{"id":3182,"href":"http:\/\/blog.cloud500.ai\/index.php?rest_route=\/wp\/v2\/posts\/3132\/revisions\/3182"}],"wp:featuredmedia":[{"embeddable":true,"href":"http:\/\/blog.cloud500.ai\/index.php?rest_route=\/wp\/v2\/media\/3133"}],"wp:attachment":[{"href":"http:\/\/blog.cloud500.ai\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=3132"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/blog.cloud500.ai\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=3132"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/blog.cloud500.ai\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=3132"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}