\u4e3a\u4f55 AI \u9700\u8981 800G InfiniBand<\/strong><\/h2>\n

AI \u6570\u636e\u4e2d\u5fc3\u7684\u74f6\u9888<\/h3>\n

AI \u6a21\u578b\u4ee5\u60ca\u4eba\u7684\u901f\u5ea6\u589e\u957f\u3002\u81ea 2010 \u5e74\u4ee5\u6765\uff0c\u8fd9\u4e9b\u6a21\u578b\u7684\u53c2\u6570\u6570\u91cf\u5927\u7ea6\u6bcf\u5e74\u7ffb\u4e00\u756a\u3002\u65e9\u671f\u7684\u6a21\u578b\uff08\u5982 Theseus\uff09\u53ea\u6709 40 \u4e2a\u53c2\u6570\uff0cGPT-3 \u5df2\u7ecf\u8fbe\u5230 1750 \u4ebf\uff0c\u800c QMoE \u6a21\u578b\u5219\u8d85\u8fc7 1.6 \u4e07\u4ebf\u3002\u8fd9\u6837\u7684\u6307\u6570\u7ea7\u589e\u957f\u9700\u8981\u80fd\u8ddf\u4e0a\u8282\u594f\u7684\u7f51\u7edc\u57fa\u7840\u8bbe\u65bd\u3002<\/p>\n

\u5728\u5206\u5e03\u5f0f\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0c\u4f60\u7684 GPU \u96c6\u7fa4\u4f1a\u9762\u4e34\u5173\u952e\u6311\u6218\u3002\u68af\u5ea6\u540c\u6b65\u4ee5\u53ca All-Reduce\u3001All-Gather \u7b49\u96c6\u5408\u901a\u4fe1\u64cd\u4f5c\uff0c\u53ef\u80fd\u6d88\u8017\u603b\u8bad\u7ec3\u65f6\u95f4\u7684 40\u201360%\u3002\u8fd9\u4e9b\u64cd\u4f5c\u5bf9\u5e26\u5bbd\u3001\u65f6\u5ef6\u3001\u6296\u52a8\u548c\u62e5\u585e\u63a7\u5236\u6781\u4e3a\u654f\u611f\u3002\u5373\u4fbf\u662f\u8f7b\u5fae\u52a3\u5316\uff0c\u4e5f\u4f1a\u9020\u6210 GPU \u5229\u7528\u7387\u5927\u5e45\u4e0b\u964d\u3002\u5fae\u79d2\u7ea7\u7684\u6296\u52a8\u5c31\u53ef\u80fd\u8ba9\u5206\u5e03\u5f0f\u8bad\u7ec3\u6027\u80fd\u964d\u4f4e 30% \u751a\u81f3\u66f4\u591a\u3002<\/p>\n

\u6807\u51c6\u7684 400G \u4ee5\u592a\u7f51\u7f51\u7edc\u5728\u8fd9\u91cc\u663e\u5f97\u529b\u4e0d\u4ece\u5fc3\u3002\u74f6\u9888\u4e0d\u5728\u4e8e\u540d\u4e49\u901f\u7387\uff0c\u800c\u5728\u4e8e\u534f\u8bae\u5f00\u9500\u3002\u5982\u679c\u6ca1\u6709 RDMA \u652f\u6301\uff0c\u8fd9\u7c7b\u7f51\u7edc\u5f80\u5f80\u5b58\u5728\u8f83\u9ad8\u65f6\u5ef6\u3001\u4e22\u5305\u548c\u4ea4\u6362\u673a\u62e5\u585e\u7b49\u95ee\u9898\u3002\u6bcf\u4e00\u4e2a\u95ee\u9898\u90fd\u4f1a\u76f4\u63a5\u62d6\u6162\u5206\u5e03\u5f0f\u8bad\u7ec3\u4e2d\u7684\u540c\u6b65\u6b65\u9aa4\uff0c\u6700\u7ec8\u5bfc\u81f4 GPU \u7a7a\u8f6c\uff0c\u7b49\u5f85\u8fdf\u8fdf\u5230\u8fbe\u7684\u68af\u5ea6\u548c\u53c2\u6570\u3002<\/p>\n

\u7528 InfiniBand \u5c06\u5e26\u5bbd\u63d0\u5347\u4e00\u500d<\/h3>\n

InfiniBand \u4ece\u8bbe\u8ba1\u7406\u5ff5\u4e0a\u5c31\u9488\u5bf9\u8fd9\u4e9b\u95ee\u9898\u7ed9\u51fa\u4e86\u7b54\u6848\u3002\u5b83\u63d0\u4f9b\u5206\u5e03\u5f0f AI \u5de5\u4f5c\u8d1f\u8f7d\u6240\u9700\u7684\u8d85\u4f4e\u65f6\u5ef6\u548c\u65e0\u635f\u901a\u4fe1\u3002\u73b0\u4ee3\u91c7\u7528 RoCEv2 \u7684 400G \u4ee5\u592a\u7f51\u5728\u90e8\u5206\u4efb\u52a1\u4e0a\u53ef\u4ee5\u4e0e InfiniBand \u5ab2\u7f8e\uff0cMeta \u4f7f\u7528 24,000 \u5757 GPU \u8bad\u7ec3 Llama 3 \u5373\u662f\u5178\u578b\u6848\u4f8b\u3002\u4e0d\u8fc7\uff0c\u5bf9\u7f3a\u5c11 RDMA \u7684\u6807\u51c6\u4ee5\u592a\u7f51\u800c\u8a00\uff0c\u5176\u4e0d\u8db3\u662f\u6839\u672c\u6027\u7684\u2014\u201410 GbE \u6bd4 InfiniBand NDR \u6162\u4e86 40 \u500d\u3002<\/p>\n

800G InfiniBand \u4ea4\u6362\u673a\u5c06\u4f60\u7684 AI \u57fa\u7840\u8bbe\u65bd\u53ef\u7528\u5e26\u5bbd\u63d0\u5347\u4e86\u4e00\u500d\u3002\u8fd9\u4e00\u98de\u8dc3\u610f\u4e49\u91cd\u5927\uff0c\u56e0\u4e3a\u96c6\u5408\u901a\u4fe1\u6027\u80fd\u51e0\u4e4e\u4e0e\u7f51\u7edc\u5bb9\u91cf\u6210\u6b63\u6bd4\u3002\u66f4\u9ad8\u5e26\u5bbd\u610f\u5473\u7740\u66f4\u5feb\u7684\u68af\u5ea6\u5728\u6570\u5343\u5757 GPU \u95f4\u5b8c\u6210\u4ea4\u6362\uff0c\u66f4\u4f4e\u65f6\u5ef6\u5219\u610f\u5473\u7740\u5904\u7406\u5668\u7684\u7a7a\u95f2\u65f6\u95f4\u66f4\u5c11\u3002<\/p>\n

\u957f\u671f\u4ee5\u6765\uff0cInfiniBand \u4e00\u76f4\u662f\u9ad8\u6027\u80fd\u8ba1\u7b97\uff08HPC\uff09\u73af\u5883\u4e2d\u9996\u9009\u7684\u4e92\u8fde\u6280\u672f\u3002\u5176\u4f4e\u65f6\u5ef6\u548c\u9ad8\u541e\u5410\uff0c\u4f7f\u5176\u975e\u5e38\u9002\u5408\u5927\u89c4\u6a21\u5e76\u884c\u8ba1\u7b97\u3002800G InfiniBand \u6280\u672f\u5c06\u8fd9\u4e00\u4f18\u52bf\u6269\u5c55\u5230\u4e86 AI \u65f6\u4ee3\u3002NVIDIA \u9488\u5bf9\u5927\u89c4\u6a21 AI \u8bad\u7ec3\u5de5\u4f5c\u8d1f\u8f7d\u5bf9\u8be5\u5e73\u53f0\u8fdb\u884c\u4e86\u4e13\u95e8\u4f18\u5316\u3002<\/p>\n

\u5bf9\u4e8e HPC \u96c6\u7fa4\u548c AI \u6570\u636e\u4e2d\u5fc3\u800c\u8a00\uff0c\u9009\u62e9\u975e\u5e38\u660e\u786e\uff1a\u4f60\u9700\u8981\u7684\u662f\u4e00\u79cd\u80fd\u6d88\u9664\u74f6\u9888\uff0c\u800c\u4e0d\u662f\u5236\u9020\u65b0\u74f6\u9888\u7684\u4e92\u8fde\u3002Quantum-X800 \u63d0\u4f9b 144 \u4e2a 800 Gb\/s \u7aef\u53e3\uff0c\u6b63\u662f\u4e3a\u6b64\u800c\u751f\u3002\u8fd9\u6b3e\u4ea4\u6362\u673a\u8ba9\u4f60\u7684\u57fa\u7840\u8bbe\u65bd\u4e3a\u4e0b\u4e00\u6ce2 AI \u521b\u65b0\u505a\u597d\u4e86\u51c6\u5907\u3002<\/p>\n

NVIDIA 800G InfiniBand \u4ea4\u6362\u673a\u7279\u6027<\/strong><\/h2>\n

\u65e0\u4e0e\u4f26\u6bd4\u7684\u541e\u5410\u548c\u7aef\u53e3\u5bc6\u5ea6<\/h3>\n

Quantum-X800 \u5e73\u53f0\u63d0\u4f9b\u4e24\u79cd\u4e0d\u540c\u673a\u7bb1\uff0c\u4ee5\u5339\u914d\u4f60\u7684\u90e8\u7f72\u9700\u6c42\u3002Q3200-RA \u91c7\u7528 2U \u673a\u578b\uff0c\u5185\u7f6e\u4e24\u4e2a\u72ec\u7acb\u7684 28.8 Tb\/s \u4ea4\u6362\u82af\u7247\u3002\u8be5\u578b\u53f7\u901a\u8fc7 36 \u4e2a OSFP \u63d2\u69fd\u63d0\u4f9b 72 \u4e2a XDR \u7aef\u53e3\uff0c\u603b\u541e\u5410\u91cf\u6700\u9ad8\u53ef\u8fbe 57.6 Tb\/s\u3002Q3400-RA \u4f7f\u7528 4U \u673a\u7bb1\uff0c\u901a\u8fc7 72 \u4e2a OSFP \u63d2\u69fd\u63d0\u4f9b 144 \u4e2a XDR \u7aef\u53e3\uff0c\u5176\u603b\u541e\u5410\u91cf\u9ad8\u8fbe 115.2 Tb\/s\u3002\u4e24\u6b3e\u673a\u578b\u5747\u57fa\u4e8e Quantum-3 \u4ea4\u6362\u82af\u7247\uff0c\u5e76\u642d\u8f7d Intel CFL i3-8100H \u5904\u7406\u5668\u3002<\/p>\n

\u4f60\u53ef\u6839\u636e\u9700\u6c42\u7075\u6d3b\u914d\u7f6e\u7aef\u53e3\u901f\u7387\uff1a\u5728 Q3200-RA \u4e0a\uff0c\u6bcf\u4e2a\u7aef\u53e3\u7684\u901f\u7387\u8303\u56f4\u4e3a 40 Gb\/s \u81f3 800 Gb\/s\uff1bQ3400-RA \u652f\u6301\u6bcf\u7aef\u53e3 400 Gb\/s \u6216 800 Gb\/s\u3002\u8fd9\u4e00 800Gbps \u7f51\u7edc\u80fd\u529b\u5c06\u4e0a\u4e00\u4ee3\u4ea7\u54c1\u7684\u5e26\u5bbd\u76f4\u63a5\u7ffb\u500d\u3002800G InfiniBand \u4ea4\u6362\u673a\u6280\u672f\u4f7f\u7aef\u5230\u7aef\u8fde\u63a5\u8fbe\u5230\u4e86\u524d\u6240\u672a\u6709\u7684\u901f\u5ea6\u3002<\/p>\n

\u5bf9\u4e8e\u5927\u89c4\u6a21 AI \u96c6\u7fa4\u800c\u8a00\uff0c\u7aef\u53e3\u5bc6\u5ea6\u540c\u6837\u81f3\u5173\u91cd\u8981\u3002Q3400-RA \u76f8\u6bd4\u65e9\u671f\u4ea4\u6362\u673a\u5177\u6709\u660e\u663e\u4f18\u52bf\uff1a<\/p>\n

\n\n\n\n\n<\/colgroup>\n\n\n\n\n\n

\u4ea4\u6362\u673a\u578b\u53f7<\/th>\n	\u7aef\u53e3\u6570\u91cf<\/th>\n	\u7aef\u53e3\u901f\u7387<\/th>\n<\/tr>\n
Q3400-RA<\/td>\n	144<\/td>\n	800 Gb\/s<\/td>\n<\/tr>\n
QM9700<\/td>\n	64<\/td>\n	400 Gb\/s<\/td>\n<\/tr>\n
QM8700<\/td>\n	40<\/td>\n	200 Gb\/s<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n \u5bf9\u6bd4\u53ef\u4ee5\u770b\u51fa\uff0c\u7aef\u53e3\u6570\u91cf\u76f8\u8f83 QM9700 \u63d0\u5347\u4e86 2.25 \u500d\uff0c\u76f8\u8f83 QM8700 \u5219\u63d0\u5347\u4e86 3.6 \u500d\u3002Q3400-RA \u7684\u9ad8 radix\uff08\u9ad8\u7aef\u53e3\u6570\uff09\u8bbe\u8ba1\uff0c\u5141\u8bb8\u4f60\u6784\u5efa\u4e24\u5c42 fat tree \u62d3\u6251\uff0c\u5c06\u591a\u8fbe 10,368 \u4e2a\u7f51\u5361\u5728\u6700\u4f4e\u65f6\u5ef6\u6761\u4ef6\u4e0b\u4e92\u8054\uff0c\u4e3a\u6700\u82db\u523b\u7684\u5de5\u4f5c\u8d1f\u8f7d\u63d0\u4f9b\u6700\u4f73\u4f5c\u4e1a\u5c40\u90e8\u6027\u3002<\/p>\n Q3400-RA \u8fd8\u91c7\u7528\u98ce\u51b7\u8bbe\u8ba1\uff0c\u53ef\u76f4\u63a5\u90e8\u7f72\u5728\u73b0\u6709\u6570\u636e\u4e2d\u5fc3\u4e2d\uff0c\u65e0\u9700\u8fdb\u884c\u91cd\u5927\u57fa\u7840\u8bbe\u65bd\u6539\u9020\u3002\u673a\u7bb1\u914d\u7f6e 8 \u4e2a\u53ef\u66f4\u6362\u7535\u6e90\u6a21\u5757\u548c 10 \u4e2a\u53ef\u66f4\u6362\u98ce\u6247\uff0c\u4fbf\u4e8e\u7ef4\u62a4\u3002\u4e24\u6b3e\u673a\u578b\u5747\u63d0\u4f9b USB 2.0 Type-A \u63a5\u53e3\u3001\u7ba1\u7406\u7aef\u53e3\u548c\u63a7\u5236\u53f0\u63a5\u53e3\uff0c\u4fbf\u4e8e\u65e5\u5e38\u8fd0\u7ef4\u7ba1\u7406\u3002<\/p>\n \u642d\u8f7d SHARP v4 \u7684\u5148\u8fdb\u7f51\u7edc\u5185\u8ba1\u7b97\u80fd\u529b<\/h3>\n 800G InfiniBand \u5e73\u53f0\u5f15\u5165\u4e86\u65b0\u4e00\u4ee3\u7f51\u7edc\u5185\u8ba1\u7b97\u6280\u672f SHARP v4\u3002\u8be5\u534f\u8bae\u901a\u8fc7\u5c06\u96c6\u5408\u64cd\u4f5c\u5378\u8f7d\u5230\u4ea4\u6362\u7f51\u7edc\u672c\u8eab\uff0c\u52a0\u901f\u4e86\u76f8\u5173\u8ba1\u7b97\u8fc7\u7a0b\uff0c\u4ece\u800c\u51cf\u5c11\u6570\u636e\u5728\u7f51\u7edc\u4e2d\u7684\u5f80\u8fd4\uff0c\u540c\u65f6\u63d0\u5347\u6574\u4f53\u6548\u7387\u3002<\/p>\n \u4e0e\u524d\u4e00\u4ee3 SHARP \u6280\u672f\u76f8\u6bd4\uff0cSHARP v4 \u5e26\u6765\u4e86 9 \u500d\u7684\u7f51\u7edc\u5185\u8ba1\u7b97\u6027\u80fd\u63d0\u5347\u3002\u8fd9\u4e00\u6539\u8fdb\u76f4\u63a5\u52a0\u5feb\u4e86 AI \u8bad\u7ec3\u5468\u671f\u3002\u5bf9\u4e8e\u57fa\u4e8e NCCL \u7684\u5de5\u4f5c\u8d1f\u8f7d\uff0c\u4f7f\u7528 SHARP \u52a0\u901f\u7684\u96c6\u5408\u901a\u4fe1\u53ef\u5e26\u6765\u6700\u9ad8 2.5 \u500d\u7684\u6027\u80fd\u63d0\u5347\u3002\u7531\u4e8e\u68af\u5ea6\u540c\u6b65\u4e0d\u518d\u6210\u4e3a\u74f6\u9888\uff0c\u4f60\u7684\u5927\u578b\u8bed\u8a00\u6a21\u578b\u8bad\u7ec3\u901f\u5ea6\u4e5f\u968f\u4e4b\u63d0\u5347\u3002<\/p>\n \u8be5\u4ea4\u6362\u673a\u8fd8\u96c6\u6210\u4e86\u589e\u5f3a\u578b\u81ea\u9002\u5e94\u8def\u7531\u529f\u80fd\uff0c\u53ef\u9488\u5bf9\u52a8\u6001\u6d41\u91cf\u6a21\u5f0f\u4f18\u5316\u8def\u5f84\u9009\u62e9\u3002\u57fa\u4e8e\u9065\u6d4b\u7684\u62e5\u585e\u63a7\u5236\u5229\u7528\u5b9e\u65f6\u7f51\u7edc\u6570\u636e\uff0c\u5728\u62e5\u585e\u53d1\u751f\u524d\u8fdb\u884c\u9884\u9632\u3002\u8fd9\u4e9b\u80fd\u529b\u534f\u540c\u5de5\u4f5c\uff0c\u5373\u4f7f\u5728\u9ad8\u8d1f\u8f7d\u60c5\u51b5\u4e0b\u4e5f\u80fd\u4fdd\u6301\u4e9a 100 \u7eb3\u79d2\u7684\u7aef\u53e3\u5230\u7aef\u53e3\u65f6\u5ef6\u3002<\/p>\n NVIDIA \u4e5f\u4e3a Quantum-X800 \u5e73\u53f0\u8bbe\u8ba1\u4e86\u51fa\u8272\u7684\u53ef\u9760\u6027\u673a\u5236\u3002\u81ea\u6108\u5f0f\u7f51\u7edc\u7ed3\u6784\u63d0\u4f9b\u76ee\u524d\u6700\u5feb\u7684\u6062\u590d\u673a\u5236\uff0c\u5728\u94fe\u8def\u6545\u969c\u65f6\u53ef\u5b9e\u73b0 1,000 \u500d\u66f4\u9ad8\u7684\u5f39\u6027\u3002\u5373\u4fbf\u5355\u4e2a\u8fde\u63a5\u51fa\u73b0\u95ee\u9898\uff0c\u4f60\u7684\u57fa\u7840\u8bbe\u65bd\u4ecd\u80fd\u4fdd\u6301\u7a33\u5b9a\u8fd0\u884c\u3002<\/p>\n \u8d85\u4f4e\u65f6\u5ef6\u4e0e\u9ad8\u541e\u5410\u7684\u7ec4\u5408\uff0c\u4f7f\u8fd9\u6b3e\u4ea4\u6362\u673a\u975e\u5e38\u9002\u5408\u9ad8\u6027\u80fd\u8ba1\u7b97\u73af\u5883\u3002HPC \u96c6\u7fa4\u9700\u8981\u5728\u6210\u5343\u4e0a\u4e07\u8282\u70b9\u4e4b\u95f4\u7ef4\u6301\u4e00\u81f4\u7684\u6027\u80fd\u8868\u73b0\u3002Quantum-X800 \u5e73\u53f0\u80fd\u591f\u5728\u4e24\u5c42 fat tree \u62d3\u6251\u4e2d\u652f\u6491\u8d85\u8fc7 10,000 \u4e2a\u8282\u70b9\uff0c\u8fd9\u4e00\u89c4\u6a21\u8fdc\u8d85\u4e0a\u4e00\u4ee3\u4ea7\u54c1\u53ef\u652f\u6301\u7684\u96c6\u7fa4\u5927\u5c0f\u3002<\/p>\n \u5bf9\u4e8e AI \u5de5\u4f5c\u8d1f\u8f7d\u800c\u8a00\uff0c\u6536\u76ca\u4e0d\u4ec5\u5728\u4e8e\u901f\u5ea6\u3002\u4e0e\u65e9\u671f\u578b\u53f7\u76f8\u6bd4\uff0c\u8be5\u4ea4\u6362\u673a\u964d\u4f4e\u4e86\u6bcf\u6bd4\u7279\u80fd\u8017\u3002\u5728\u4fdd\u6301\u5353\u8d8a\u6027\u80fd\u7684\u540c\u65f6\uff0c\u4f60\u53ef\u4ee5\u5b9e\u73b0\u66f4\u9ad8\u7684\u6bcf\u74e6\u8ba1\u7b97\u80fd\u529b\u3002Quantum-X800 \u5e73\u53f0\u8fd8\u63d0\u4f9b Quantum-X Photonics \u9009\u9879\uff0c\u5c06\u5149\u5b66\u6a21\u5757\u76f4\u63a5\u96c6\u6210\u5230\u4ea4\u6362\u82af\u7247\u4e2d\uff0c\u8fdb\u4e00\u6b65\u964d\u4f4e\u80fd\u8017\u548c\u65f6\u5ef6\u3002<\/p>\n \u8fd9\u6b3e 800G InfiniBand \u4ea4\u6362\u673a\u4e3a\u73b0\u4ee3 AI \u57fa\u7840\u8bbe\u65bd\u63d0\u4f9b\u4e86\u5b8c\u6574\u89e3\u51b3\u65b9\u6848\u3002\u4f60\u5c06\u83b7\u5f97\u6ee1\u8db3\u524d\u6cbf\u7814\u7a76\u548c\u751f\u4ea7\u7cfb\u7edf\u6240\u9700\u7684\u5e26\u5bbd\u3001\u667a\u80fd\u4e0e\u53ef\u9760\u6027\u3002<\/p>\n 800G InfiniBand \u5bf9 AI \u7684\u4f18\u52bf<\/strong><\/h2>\n \u52a0\u901f AI \u8bad\u7ec3\u548c\u63a8\u7406<\/h3>\n Quantum-X800 InfiniBand \u4ea4\u6362\u673a\u5728\u5927\u578b\u8bed\u8a00\u6a21\u578b\u8bad\u7ec3\u6548\u7387\u65b9\u9762\u5b9e\u73b0\u4e86 9 \u500d\u63d0\u5347\u3002\u8fd9\u5f97\u76ca\u4e8e NVIDIA \u7684 SHARP v4 \u534f\u8bae\uff0c\u8be5\u534f\u8bae\u5728\u4ea4\u6362\u673a\u4e0a\u76f4\u63a5\u6267\u884c\u7f51\u7edc\u5185\u6570\u636e\u805a\u5408\u548c\u96c6\u5408\u64cd\u4f5c\u3002\u901a\u8fc7\u5c06\u8fd9\u4e9b\u901a\u4fe1\u4efb\u52a1\u4ece\u670d\u52a1\u5668\u4fa7\u5378\u8f7d\u51fa\u53bb\uff0c\u4ea4\u6362\u673a\u964d\u4f4e\u4e86\u901a\u4fe1\u74f6\u9888\u4e0e\u670d\u52a1\u5668\u8d1f\u8f7d\uff0c\u4ece\u800c\u663e\u8457\u52a0\u5feb\u4e07\u4ebf\u53c2\u6570\u6a21\u578b\u7684\u8bad\u7ec3\u5faa\u73af\u3002<\/p>\n \u4f60\u8fd8\u80fd\u83b7\u5f97\u663e\u8457\u7684\u80fd\u6548\u6539\u5584\u3002\u8be5\u4ea4\u6362\u673a\u4f7f\u7528\u5171\u5c01\u88c5\u5149\u5b66\u6280\u672f\uff0c\u53d6\u6d88\u4e86\u53ef\u63d2\u62d4\u5149\u6a21\u5757\u3002\u8fd9\u79cd\u8bbe\u8ba1\u7b80\u5316\u4e86\u7269\u6599\u6e05\u5355\u5e76\u964d\u4f4e\u4e86\u6210\u672c\uff0c\u5176\u80fd\u6548\u662f\u4f20\u7edf\u53ef\u63d2\u62d4\u5149\u6a21\u5757\u7684 5 \u500d\u3002\u7531\u4e8e\u65e0\u9700 DSP \u91cd\u5b9a\u65f6\u82af\u7247\uff0c\u7f51\u7edc\u4fa7\u7684\u529f\u8017\u5927\u5e45\u964d\u4f4e\u3002\u8fd9\u4e00\u964d\u8017\u76f4\u63a5\u51cf\u5c11\u4e86 AI \u6570\u636e\u4e2d\u5fc3\u7684\u603b\u4f53\u62e5\u6709\u6210\u672c\uff0c\u5728\u6574\u4e2a\u57fa\u7840\u8bbe\u65bd\u751f\u547d\u5468\u671f\u5185\u4f1a\u7d2f\u79ef\u6210\u53ef\u89c2\u7684\u8fd0\u8425\u6536\u76ca\u3002<\/p>\n \n\n\n\n<\/colgroup>\n\n\n\n\n\n \u6307\u6807<\/th>\n \u6539\u8fdb\u5e45\u5ea6<\/th>\n<\/tr>\n \u80fd\u6548<\/td>\n \u6bd4\u53ef\u63d2\u62d4\u5149\u6a21\u5757\u9ad8 5 \u500d<\/td>\n<\/tr>\n \u6301\u7eed AI \u5e94\u7528\u8fd0\u884c\u65f6\u95f4<\/td>\n \u5f39\u6027\u63d0\u5347 5 \u500d<\/td>\n<\/tr>\n \u6d1e\u5bdf\u65f6\u95f4\uff08Time to insight\uff09<\/td>\n \u90e8\u7f72\u901f\u5ea6\u63d0\u5347 1.3 \u500d<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n \u8fd9\u4e09\u4e2a\u6307\u6807\u51f8\u663e\u4e86 Quantum-X800 \u5e73\u53f0\u5728\u8fd0\u8425\u5c42\u9762\u7684\u4f18\u52bf\u3002\u7531\u4e8e\u7701\u53bb\u4e86 DSP \u91cd\u5b9a\u65f6\u82af\u7247\uff0c\u8be5\u4ea4\u6362\u673a\u4e5f\u80fd\u8fdb\u4e00\u6b65\u964d\u4f4e\u65f6\u5ef6\u3002\u5bf9\u4e8e\u5927\u89c4\u6a21 AI \u8bad\u7ec3\u6765\u8bf4\uff0c\u8fd9\u79cd\u8d85\u4f4e\u65f6\u5ef6\u81f3\u5173\u91cd\u8981\uff0c\u6bcf\u4e00\u5fae\u79d2\u90fd\u5f88\u5b9d\u8d35\u3002\u4f60\u7684 GPU \u5c06\u82b1\u66f4\u5c11\u65f6\u95f4\u7b49\u5f85\u6570\u636e\uff0c\u800c\u628a\u66f4\u591a\u65f6\u95f4\u7528\u4e8e\u8ba1\u7b97\u3002InfiniBand \u67b6\u6784\u63d0\u4f9b\u4e86\u6807\u51c6\u4ee5\u592a\u7f51\u96be\u4ee5\u5b9e\u73b0\u7684\u65e0\u635f\u901a\u4fe1\uff0c\u786e\u4fdd\u8bad\u7ec3\u4f5c\u4e1a\u4e0d\u4f1a\u56e0\u4e22\u5305\u6216\u7f51\u7edc\u62e5\u585e\u800c\u88ab\u6253\u65ad\u3002\u6700\u7ec8\u6210\u679c\u662f\uff1a\u66f4\u5feb\u7684\u6a21\u578b\u5f00\u53d1\u5468\u671f\u4ee5\u53ca\u66f4\u9ad8\u7684\u6574\u96c6\u7fa4 GPU \u5229\u7528\u7387\u3002<\/p>\n \u501f\u52a9 NVIDIA Quantum-X800 \u5b9e\u73b0\u53ef\u6269\u5c55\u6027<\/h3>\n Quantum-X800 \u5e73\u53f0\u652f\u6301\u4ece\u5c0f\u578b\u7814\u7a76\u5b9e\u9a8c\u5ba4\u5230\u8d85\u5927\u89c4\u6a21\u6570\u636e\u4e2d\u5fc3\u7684\u5e73\u6ed1\u6269\u5c55\u3002\u4e0e\u4e0a\u4e00\u4ee3\u4ea7\u54c1\u76f8\u6bd4\uff0cNVIDIA \u7684\u4ea4\u6362\u673a\u5728\u53ef\u6269\u5c55\u6027\u65b9\u9762\u5177\u6709\u663e\u8457\u4f18\u52bf\u3002<\/p>\n \n\n\n\n\n<\/colgroup>\n\n\n\n\n\n \u7279\u6027<\/th>\n Quantum X800<\/th>\n QM9700\uff08\u4e0a\u4e00\u4ee3\uff09<\/th>\n<\/tr>\n \u6bcf\u53f0\u4ea4\u6362\u673a\u7aef\u53e3\u6570<\/td>\n 144\uff08800 Gb\/s\uff09<\/td>\n 64\uff08400 Gb\/s\uff09<\/td>\n<\/tr>\n \u6700\u5927\u4e3b\u673a\u8fde\u63a5\u6570\uff08\u4e24\u5c42 Fat-Tree\uff09<\/td>\n >10,000\uff08800 Gb\/s\uff09<\/td>\n –<\/td>\n<\/tr>\n \u7f51\u7edc\u5185\u8ba1\u7b97<\/td>\n \u7b2c\u56db\u4ee3 SHARP\uff08\u652f\u6301 FP8\uff09<\/td>\n –<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n \u9ad8\u7aef\u53e3\u5bc6\u5ea6\u4f7f\u4f60\u80fd\u591f\u5728\u4e24\u5c42 fat tree \u62d3\u6251\u4e2d\u8fde\u63a5\u8d85\u8fc7 10,000 \u4e2a\u8282\u70b9\u3002\u8fd9\u6837\u7684\u89c4\u6a21\u8db3\u4ee5\u627f\u8f7d\u6700\u5927\u578b AI \u5de5\u4f5c\u8d1f\u8f7d\uff0c\u800c\u65e0\u9700\u6784\u5efa\u590d\u6742\u7684\u591a\u5c42\u7f51\u7edc\u8bbe\u8ba1\u3002\u968f\u7740\u8ba1\u7b97\u9700\u6c42\u7684\u589e\u957f\uff0c\u4f60\u7684\u4e91\u57fa\u7840\u8bbe\u65bd\u53ef\u4ee5\u987a\u7545\u6269\u5c55\uff0c\u6bcf\u4e2a\u65b0\u589e\u8282\u70b9\u90fd\u80fd\u5728\u4e0d\u6253\u4e71\u65e2\u6709\u5de5\u4f5c\u6d41\u7684\u524d\u63d0\u4e0b\u878d\u5165\u96c6\u7fa4\u3002<\/p>\n Q3400-RA \u91c7\u7528\u98ce\u51b7\u8bbe\u8ba1\uff0c\u53ef\u4e0e\u73b0\u6709\u6570\u636e\u4e2d\u5fc3\u7684\u6563\u70ed\u57fa\u7840\u8bbe\u65bd\u517c\u5bb9\u3002<\/p>\n \u98ce\u51b7\u5e73\u53f0\u4ec5\u652f\u6301\u6563\u70ed\u9ccd\u7247\uff08Finned Top\uff09\u7248\u672c\u7684 1.6T \u6536\u53d1\u5668\u3002\u5e73\u9876\uff08Flat Top\uff09\u7248\u672c\u4f9d\u8d56\u51b7\u677f\u5bfc\u70ed\uff0c\u53ea\u9002\u7528\u4e8e\u6db2\u51b7\u67b6\u6784\u3002Q3400-RA \u91c7\u7528\u7684\u662f Finned Top \u6536\u53d1\u5668\uff0c\u8fd9\u4e00\u70b9\u5370\u8bc1\u4e86\u5176\u98ce\u51b7\u8bbe\u8ba1\u4ee5\u53ca\u4e0e\u73b0\u6709\u6570\u636e\u4e2d\u5fc3\u6563\u70ed\u7cfb\u7edf\u7684\u517c\u5bb9\u6027\u3002<\/p>\n \u5f97\u76ca\u4e8e\u8fd9\u79cd\u98ce\u51b7\u65b9\u6848\uff0c\u4f60\u53ef\u4ee5\u5728\u6807\u51c6\u673a\u67b6\u4e2d\u90e8\u7f72 Q3400-RA\uff0c\u800c\u65e0\u9700\u6db2\u51b7\u7cfb\u7edf\u30024U \u673a\u7bb1\u53ef\u76f4\u63a5\u653e\u5165\u73b0\u6709\u7684\u6570\u636e\u4e2d\u5fc3\u673a\u67dc\u4e2d\u3002\u4f60\u4e5f\u5c31\u65e0\u9700\u4e3a\u8d85\u5927\u89c4\u6a21\u6570\u636e\u4e2d\u5fc3\u6539\u9020\u6563\u70ed\u7cfb\u7edf\uff0c\u4ece\u800c\u907f\u514d\u4e86\u989d\u5916\u7684\u6210\u672c\u4e0e\u590d\u6742\u5ea6\u3002\u8fd9\u4e00\u8bbe\u8ba1\u9009\u62e9\u4f7f\u90e8\u7f72\u8fc7\u7a0b\u7b80\u5355\u4e14\u53ef\u9884\u671f\u3002<\/p>\n InfiniBand \u6280\u672f\u4e0e NVIDIA \u5e73\u53f0\u8bbe\u8ba1\u7684\u7ed3\u5408\uff0c\u4f7f\u8fd9\u6b3e\u4ea4\u6362\u673a\u975e\u5e38\u9002\u5408 HPC \u73af\u5883\u3002\u5176\u8d85\u4f4e\u65f6\u5ef6\u548c\u9ad8\u541e\u5410\u80fd\u591f\u652f\u6491\u6700\u82db\u523b\u7684 HPC \u5de5\u4f5c\u8d1f\u8f7d\u3002\u4f60\u53ef\u4ee5\u5728\u540c\u4e00\u7f51\u7edc\u4e0a\u540c\u65f6\u8fd0\u884c AI \u8bad\u7ec3\u4efb\u52a1\u548c\u4f20\u7edf HPC \u4eff\u771f\u3002\u968f\u7740\u6a21\u578b\u89c4\u6a21\u4e0d\u65ad\u81a8\u80c0\uff0c\u8fd9\u6b3e InfiniBand \u4ea4\u6362\u673a\u4e5f\u80fd\u6ee1\u8db3\u4e0d\u65ad\u589e\u957f\u7684 GPU \u8ba1\u7b97\u9700\u6c42\u3002800 Gb\/s \u7684\u7aef\u53e3\u5e26\u5bbd\u786e\u4fdd GPU \u4e0d\u4f1a\u56e0\u4e3a\u6570\u636e\u532e\u4e4f\u800c\u201c\u997f\u6b7b\u201d\u3002\u8fd9\u79cd\u53ef\u6269\u5c55\u6027\u8ba9\u4f60\u7684\u57fa\u7840\u8bbe\u65bd\u4e3a\u672a\u6765\u7684 AI \u7a81\u7834\u505a\u597d\u51c6\u5907\u3002<\/p>\n \n <\/div>\n<\/div>\n NVIDIA Quantum-X800 \u5e73\u53f0\u6807\u5fd7\u7740 AI \u7f51\u7edc\u9886\u57df\u7684\u5173\u952e\u8fdb\u6b65\u3002\u5176 800G InfiniBand \u6280\u672f\u4e3a\u4f60\u7684 AI \u5de5\u4f5c\u8d1f\u8f7d\u63d0\u4f9b\u524d\u6240\u672a\u6709\u7684\u901f\u5ea6\u3001\u589e\u5f3a\u7684\u53ef\u6269\u5c55\u6027\u4ee5\u53ca\u66f4\u9ad8\u7684\u6548\u7387\u3002\u8fd9\u6b3e 800G InfiniBand \u4ea4\u6362\u673a\u5f7b\u5e95\u6d88\u9664\u4e86\u62d6\u6162 GPU \u8bad\u7ec3\u548c\u63a8\u7406\u7684\u7f51\u7edc\u74f6\u9888\u3002<\/p>\n NVIDIA \u7684\u6280\u672f\u8def\u7ebf\u56fe\u8fd8\u5c06\u8fd9\u4e00\u521b\u65b0\u8fdb\u4e00\u6b65\u63a8\u5411\u524d\u6cbf\u3002\u4e0b\u4e00\u4ee3\u5e73\u53f0 Quantum-X Photonics \u5c06\u7845\u5149\u5b50\u6280\u672f\u76f4\u63a5\u96c6\u6210\u5230\u4ea4\u6362\u673a\u5c01\u88c5\u4e2d\uff0c\u53ef\u5b9e\u73b0 3.5 \u500d\u7684\u80fd\u6548\u63d0\u5347\u548c 10 \u500d\u66f4\u9ad8\u7684\u5f39\u6027\u3002\u5e94\u7528\u53ef\u4ee5\u5728\u4e0d\u4e2d\u65ad\u60c5\u51b5\u4e0b\u8fd0\u884c\u66f4\u957f\u65f6\u95f4\uff0c\u6700\u957f\u53ef\u63d0\u5347 5 \u500d\u3002\u672a\u6765\u7684 InfiniBand \u8fde\u63a5\u5c06\u5411 1.6T \u901f\u7387\u6f14\u8fdb\uff0c\u4ee5\u652f\u6491\u4e07\u4ebf\u7ea7\u53c2\u6570\u6a21\u578b\u3002<\/p>\n \u601d\u8003\u4e00\u4e0b\u8fd9\u9879\u6280\u672f\u5c06\u5982\u4f55\u91cd\u5851\u4f60\u7684\u6570\u636e\u4e2d\u5fc3\u7b56\u7565\u3002Quantum-X800 \u5e73\u53f0\u8ba9\u4f60\u7684\u57fa\u7840\u8bbe\u65bd\u5728\u4eca\u5929\u5c31\u4e3a\u660e\u5929\u7684 AI \u7a81\u7834\u505a\u597d\u51c6\u5907\u3002<\/p>\n \u5e38\u89c1\u95ee\u9898\uff08FAQ\uff09<\/strong><\/h2>\n Quantum-X800 \u4e0e\u65e7\u6b3e\u4ea4\u6362\u673a\u6709\u4f55\u4e0d\u540c\uff1f<\/h3>\n \u5b83\u63d0\u4f9b 144 \u4e2a\u7aef\u53e3\uff0c\u6bcf\u4e2a\u7aef\u53e3\u652f\u6301 800 Gb\/s \u7684 InfiniBand \u901f\u7387\u3002\u8fd9\u6b3e NVIDIA \u4ea4\u6362\u673a\u91c7\u7528 SHARP v4 \u52a0\u901f\u96c6\u5408\u64cd\u4f5c\uff0c\u8ba9\u5927\u578b AI \u6a21\u578b\u8bad\u7ec3\u6548\u7387\u63d0\u5347 9 \u500d\u3002<\/p>\n SHARP v4 \u5982\u4f55\u63d0\u5347 AI \u8bad\u7ec3\u6548\u7387\uff1f<\/h3>\n NVIDIA \u7684 SHARP v4 \u5c06\u96c6\u5408\u901a\u4fe1\u4ece GPU \u5378\u8f7d\u5230\u4ea4\u6362\u7f51\u7edc\u4e2d\u3002\u8fd9\u9879 NVIDIA \u6280\u672f\u51cf\u5c11\u4e86\u6570\u636e\u79fb\u52a8\uff0c\u5728\u8bad\u7ec3\u5de5\u4f5c\u8d1f\u8f7d\u4e2d\u53ef\u5b9e\u73b0\u9ad8\u8fbe 9 \u500d\u7684\u6027\u80fd\u63d0\u5347\u3002<\/p>\n Quantum-X800 \u80fd\u5426\u5728\u73b0\u6709\u6570\u636e\u4e2d\u5fc3\u4e2d\u4f7f\u7528\uff1f<\/h3>\n \u53ef\u4ee5\u3002Q3400-RA \u91c7\u7528\u98ce\u51b7\u8bbe\u8ba1\uff0c\u53ef\u76f4\u63a5\u90e8\u7f72\u5728\u6807\u51c6\u673a\u67b6\u4e2d\uff0c\u65e0\u9700\u6db2\u51b7\u7cfb\u7edf\u3002\u8fd9\u4f7f\u5176\u80fd\u591f\u8f7b\u677e\u96c6\u6210\u5230\u8d85\u5927\u89c4\u6a21\u6570\u636e\u4e2d\u5fc3\u548c\u73b0\u6709\u4e91\u57fa\u7840\u8bbe\u65bd\u4e2d\u3002<\/p>\n \u8fd9\u6b3e\u4ea4\u6362\u673a\u5982\u4f55\u652f\u6491\u5927\u578b AI \u96c6\u7fa4\uff1f<\/h3>\n Quantum-X800 \u80fd\u5728\u4e24\u5c42 fat tree \u62d3\u6251\u4e2d\u8fde\u63a5\u8d85\u8fc7 10,000 \u4e2a\u8282\u70b9\uff0c\u6bcf\u4e2a\u8282\u70b9\u62e5\u6709 800 Gb\/s \u5e26\u5bbd\u3002\u8fd9\u6b3e NVIDIA \u4ea4\u6362\u673a\u53ef\u4ece\u7814\u7a76\u5b9e\u9a8c\u5ba4\u6269\u5c55\u5230\u8d85\u5927\u89c4\u6a21 GPU \u96c6\u7fa4\u3002<\/p>\n \u8be5\u4ea4\u6362\u673a\u80fd\u5e94\u5bf9\u672a\u6765\u66f4\u5927\u89c4\u6a21\u7684 AI \u6a21\u578b\u5417\uff1f<\/h3>\n \u53ef\u4ee5\u3002800 Gb\/s \u7684 InfiniBand \u5e26\u5bbd\u53ef\u968f\u4f60\u7684\u9700\u6c42\u6269\u5c55\u3002\u8fd9\u4e00 NVIDIA \u5e73\u53f0\u652f\u6301\u8d85\u8fc7 10,000 \u4e2a\u8282\u70b9\uff0c\u9002\u914d\u4e07\u4ebf\u53c2\u6570\u6a21\u578b\u3002\u4f60\u7684\u57fa\u7840\u8bbe\u65bd\u53ef\u4ee5\u5728\u4e0d\u4e2d\u65ad\u4e1a\u52a1\u7684\u60c5\u51b5\u4e0b\u6301\u7eed\u6269\u5bb9\u3002<\/p>\n","protected":false},"excerpt":{"rendered":" NVIDIA \u63a8\u51fa\u4e86 Quantum-X800 InfiniBand \u4ea4\u6362\u673a\uff0c\u8fd9\u662f AI \u57fa\u7840\u8bbe\u65bd \u9886\u57df\u7684\u4e00\u9879\u7a81\u7834\u3002\u8fd9\u6b3e\u9ad8\u6027\u80fd\u4ea4\u6362\u673a\u63d0\u4f9b 144 \u4e2a 800 Gb\/s \u7aef\u53e3\uff0c\u4e13\u4e3a\u8fde\u63a5\u73b0\u4ee3 \u7f8e\u56fd\u670d\u52a1\u5668 \u90e8\u7f72\u4e2d\u7684\u5927\u89c4\u6a21 GPU \u96c6\u7fa4\u800c\u8bbe\u8ba1\u3002\u4f60\u53ef\u4ee5\u6d88\u9664 AI \u8bad\u7ec3\u548c\u63a8\u7406\u4e2d\u7684\u6570\u636e\u74f6\u9888\uff0c\u4ece\u800c\u52a0\u5feb\u6a21\u578b\u5f00\u53d1\u901f\u5ea6\uff0c\u5b9e\u73b0\u66f4\u5927\u89c4\u6a21\u7684\u90e8\u7f72\u3002 800G InfiniBand \u4ea4\u6362\u673a\u4ee3\u8868\u4e86\u7f51\u7edc\u6027\u80fd\u7684\u4e00\u6b21\u98de\u8dc3\u3002\u51ed\u501f\u4e9a 100 \u7eb3\u79d2\u7684\u7aef\u53e3\u5230\u7aef\u53e3\u65f6\u5ef6\uff0c\u4ee5\u53ca\u901a\u8fc7 SHARP v4 \u5b9e\u73b0\u7684 14.4 TFLOPS\uff08\u4e07\u4ebf\u6b21\u6d6e\u70b9\u8fd0\u7b97\uff09\u7684\u7f51\u7edc\u5185\u8ba1\u7b97\u80fd\u529b\uff0cQuantum-X800 \u5e73\u53f0\u6811\u7acb\u4e86\u65b0\u7684\u6807\u51c6\u3002800G InfiniBand \u6280\u672f\u76f4\u63a5\u652f\u6491\u7740\u65e5\u76ca\u589e\u957f\u7684 GPU \u8ba1\u7b97\u9700\u6c42\u3002 \u8fd9\u6b3e\u4ea4\u6362\u673a\u901a\u8fc7\u4e3a\u4e0b\u4e00\u4ee3\u6570\u636e\u4e2d\u5fc3\u63d0\u4f9b\u9aa8\u5e72\u7f51\u7edc\uff0c\u63a8\u52a8\u4e86 AI \u65f6\u4ee3\u7684\u6f14\u8fdb\u3002NVIDIA \u7684\u521b\u65b0\uff0c\u4f7f\u4f60\u80fd\u591f\u6784\u5efa\u4e3a\u672a\u6765 AI \u5de5\u4f5c\u8d1f\u8f7d\u505a\u597d\u51c6\u5907\u7684\u57fa\u7840\u8bbe\u65bd\u3002 \u8981\u70b9\u901f\u89c8 \u5168\u65b0\u7684 NVIDIA \u4ea4\u6362\u673a\u91c7\u7528 800G InfiniBand\uff0c\u6d88\u9664 AI \u8bad\u7ec3\u4e2d\u7684\u6570\u636e\u74f6\u9888\u3002 SHARP v4 \u6280\u672f\u4ece GPU \u4e2d\u5378\u8f7d\u90e8\u5206\u5de5\u4f5c\uff0c\u4f7f\u8bad\u7ec3\u901f\u5ea6\u63d0\u5347 9 \u500d\u3002 […]<\/p>\n Read More…<\/a><\/p>\n","protected":false},"author":2,"featured_media":34278,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[4816],"tags":[12962,12963,12047,12048,1734],"class_list":["post-34281","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-america-dedicated-server-sc","tag-800g-infiniband-switch","tag-gpu-cluster","tag-nvidia-quantum-x800","tag-ai-networking","tag-high-performance-computing-sc"],"acf":[],"yoast_head":"\n