\u70ba\u4f55 AI \u9700\u8981 800G InfiniBand<\/strong><\/h2>\n

AI \u8cc7\u6599\u4e2d\u5fc3\u7684\u74f6\u9838<\/h3>\n

AI \u6a21\u578b\u6b63\u4ee5\u9a5a\u4eba\u7684\u901f\u5ea6\u6210\u9577\u3002\u81ea 2010 \u5e74\u4ee5\u4f86\uff0c\u9019\u4e9b\u6a21\u578b\u7684\u53c3\u6578\u6578\u91cf\u5927\u7d04\u6bcf\u5e74\u7ffb\u500d\u3002\u65e9\u671f\u7684\u6a21\u578b\uff08\u5982 Theseus\uff09\u53ea\u6709 40 \u500b\u53c3\u6578\uff0cGPT-3 \u5df2\u9054 1,750 \u5104\uff0c\u800c QMoE \u6a21\u578b\u66f4\u662f\u8d85\u904e 1.6 \u5146\u3002\u9019\u6a23\u7684\u6307\u6578\u6210\u9577\u9700\u8981\u80fd\u8ddf\u4e0a\u8173\u6b65\u7684\u7db2\u8def\u57fa\u790e\u8a2d\u65bd\u3002<\/p>\n

\u5728\u5206\u6563\u5f0f\u8a13\u7df4\u904e\u7a0b\u4e2d\uff0c\u4f60\u7684 GPU \u53e2\u96c6\u6703\u9762\u81e8\u95dc\u9375\u6311\u6230\u3002\u68af\u5ea6\u540c\u6b65\u4ee5\u53ca All-Reduce\u3001All-Gather \u7b49\u96c6\u5408\u901a\u8a0a\u64cd\u4f5c\uff0c\u53ef\u80fd\u8017\u8cbb\u7e3d\u8a13\u7df4\u6642\u9593\u7684 40\u201360%\u3002\u9019\u4e9b\u64cd\u4f5c\u5c0d\u983b\u5bec\u3001\u5ef6\u9072\u3001\u6296\u52d5\u8207\u58c5\u585e\u63a7\u5236\u6975\u70ba\u654f\u611f\u3002\u5373\u4fbf\u662f\u8f15\u5fae\u7684\u9000\u5316\uff0c\u4e5f\u6703\u9020\u6210 GPU \u4f7f\u7528\u7387\u5927\u5e45\u4e0b\u6ed1\u3002\u5fae\u79d2\u7d1a\u7684\u6296\u52d5\u5c31\u53ef\u80fd\u8b93\u5206\u6563\u5f0f\u8a13\u7df4\u6548\u80fd\u964d\u4f4e 30% \u751a\u81f3\u66f4\u591a\u3002<\/p>\n

\u6a19\u6e96\u7684 400G \u4e59\u592a\u7db2\u8def\u5728\u9019\u65b9\u9762\u986f\u5f97\u529b\u4e0d\u5f9e\u5fc3\u3002\u74f6\u9838\u4e0d\u5728\u540d\u7fa9\u901f\u7387\uff0c\u800c\u5728\u5354\u5b9a\u958b\u92b7\u3002\u5982\u679c\u6c92\u6709 RDMA \u652f\u63f4\uff0c\u9019\u985e\u7db2\u8def\u5f80\u5f80\u6703\u6709\u8f03\u9ad8\u5ef6\u9072\u3001\u5c01\u5305\u907a\u5931\u8207\u4ea4\u63db\u5668\u58c5\u585e\u7b49\u554f\u984c\u3002\u6bcf\u4e00\u9805\u554f\u984c\u90fd\u6703\u76f4\u63a5\u62d6\u6162\u5206\u6563\u5f0f\u8a13\u7df4\u4e2d\u7684\u540c\u6b65\u6b65\u9a5f\uff0c\u6700\u5f8c\u5c0e\u81f4 GPU \u7a7a\u8f49\uff0c\u7b49\u5f85\u9072\u9072\u672a\u5230\u7684\u68af\u5ea6\u8207\u53c3\u6578\u3002<\/p>\n

\u4ee5 InfiniBand \u5c07\u983b\u5bec\u63d0\u5347\u4e00\u500d<\/h3>\n

InfiniBand \u5f9e\u8a2d\u8a08\u7406\u5ff5\u4e0a\u4fbf\u91dd\u5c0d\u9019\u4e9b\u554f\u984c\u7d66\u51fa\u4e86\u7b54\u6848\u3002\u5b83\u63d0\u4f9b\u5206\u6563\u5f0f AI \u5de5\u4f5c\u8ca0\u8f09\u6240\u9700\u7684\u8d85\u4f4e\u5ef6\u9072\u8207\u7121\u640d\u901a\u8a0a\u3002\u63a1\u7528 RoCEv2 \u7684\u73fe\u4ee3 400G \u4e59\u592a\u7db2\uff0c\u5728\u67d0\u4e9b\u4efb\u52d9\u4e0a\u53ef\u4ee5\u8207 InfiniBand \u4e00\u8f03\u9ad8\u4e0b\uff1bMeta \u4f7f\u7528 24,000 \u9846 GPU \u8a13\u7df4 Llama 3 \u5373\u662f\u5178\u578b\u6848\u4f8b\u3002\u4e0d\u904e\uff0c\u5c0d\u7f3a\u4e4f RDMA \u7684\u6a19\u6e96\u4e59\u592a\u7db2\u800c\u8a00\uff0c\u5176\u5148\u5929\u4e0d\u8db3\u662f\u7d50\u69cb\u6027\u7684\u2014\u201410 GbE \u6bd4 InfiniBand NDR \u6162\u4e0a 40 \u500d\u3002<\/p>\n

800G InfiniBand \u4ea4\u63db\u5668\u5c07\u4f60\u7684 AI \u57fa\u790e\u8a2d\u65bd\u53ef\u7528\u983b\u5bec\u63d0\u5347\u4e86\u4e00\u500d\u3002\u9019\u9805\u98db\u8e8d\u6975\u5177\u610f\u7fa9\uff0c\u56e0\u70ba\u96c6\u5408\u901a\u8a0a\u6548\u80fd\u5e7e\u4e4e\u8207\u7db2\u8def\u5bb9\u91cf\u6210\u6b63\u6bd4\u3002\u66f4\u9ad8\u7684\u983b\u5bec\u4ee3\u8868\u68af\u5ea6\u80fd\u66f4\u5feb\u5730\u5728\u6578\u5343\u9846 GPU \u4e4b\u9593\u4ea4\u63db\uff0c\u800c\u66f4\u4f4e\u7684\u5ef6\u9072\u5247\u610f\u5473\u8457\u8655\u7406\u5668\u9592\u7f6e\u6642\u9593\u66f4\u5c11\u3002<\/p>\n

\u9577\u4e45\u4ee5\u4f86\uff0cInfiniBand \u4e00\u76f4\u662f\u9ad8\u6548\u80fd\u904b\u7b97\uff08HPC\uff09\u74b0\u5883\u4e2d\u7684\u9996\u9078\u4e92\u9023\u6280\u8853\u3002\u5176\u4f4e\u5ef6\u9072\u8207\u9ad8\u541e\u5410\u91cf\uff0c\u4f7f\u5176\u975e\u5e38\u9069\u5408\u5927\u898f\u6a21\u5e73\u884c\u8655\u7406\u3002800G InfiniBand \u6280\u8853\u5c07\u9019\u9805\u512a\u52e2\u5ef6\u4f38\u81f3 AI \u6642\u4ee3\u3002NVIDIA \u91dd\u5c0d\u5927\u898f\u6a21 AI \u8a13\u7df4\u5de5\u4f5c\u8ca0\u8f09\uff0c\u5c0d\u6b64\u5e73\u53f0\u9032\u884c\u5c08\u9580\u8abf\u6821\u3002<\/p>\n

\u5c0d HPC \u53e2\u96c6\u8207 AI \u8cc7\u6599\u4e2d\u5fc3\u800c\u8a00\uff0c\u9078\u64c7\u5df2\u76f8\u7576\u660e\u78ba\uff1a\u4f60\u9700\u8981\u7684\u662f\u80fd\u6d88\u9664\u74f6\u9838\uff0c\u800c\u975e\u88fd\u9020\u65b0\u74f6\u9838\u7684\u4e92\u9023\u6280\u8853\u3002Quantum-X800 \u63d0\u4f9b 144 \u500b 800 Gb\/s \u9023\u63a5\u57e0\uff0c\u6b63\u662f\u70ba\u6b64\u800c\u751f\u3002\u9019\u6b3e\u4ea4\u63db\u5668\u8b93\u4f60\u7684\u57fa\u790e\u8a2d\u65bd\u70ba\u4e0b\u4e00\u6ce2 AI \u5275\u65b0\u505a\u597d\u6e96\u5099\u3002<\/p>\n

NVIDIA 800G InfiniBand \u4ea4\u63db\u5668\u7279\u6027<\/strong><\/h2>\n

\u7121\u8207\u502b\u6bd4\u7684\u541e\u5410\u91cf\u8207\u9023\u63a5\u57e0\u5bc6\u5ea6<\/h3>\n

Quantum-X800 \u5e73\u53f0\u63d0\u4f9b\u5169\u7a2e\u4e0d\u540c\u6a5f\u7bb1\uff0c\u4ee5\u7b26\u5408\u4f60\u7684\u90e8\u7f72\u9700\u6c42\u3002Q3200-RA \u63a1\u7528 2U \u6a5f\u578b\uff0c\u5167\u5efa\u5169\u9846\u7368\u7acb\u7684 28.8 Tb\/s \u4ea4\u63db\u6676\u7247\u3002\u6b64\u578b\u865f\u900f\u904e 36 \u500b OSFP \u63d2\u69fd\u63d0\u4f9b 72 \u500b XDR \u9023\u63a5\u57e0\uff0c\u7e3d\u541e\u5410\u91cf\u6700\u9ad8\u53ef\u9054 57.6 Tb\/s\u3002Q3400-RA \u5247\u63a1\u7528 4U \u6a5f\u7bb1\uff0c\u900f\u904e 72 \u500b OSFP \u63d2\u69fd\u63d0\u4f9b 144 \u500b XDR \u9023\u63a5\u57e0\uff0c\u7e3d\u541e\u5410\u91cf\u9ad8\u9054 115.2 Tb\/s\u3002\u5169\u6b3e\u6a5f\u578b\u7686\u57fa\u65bc Quantum-3 \u4ea4\u63db\u6676\u7247\uff0c\u4e26\u642d\u8f09 Intel CFL i3-8100H \u8655\u7406\u5668\u3002<\/p>\n

\u4f60\u53ef\u4f9d\u9700\u6c42\u5f48\u6027\u8a2d\u5b9a\u9023\u63a5\u57e0\u901f\u7387\uff1a\u5728 Q3200-RA \u4e0a\uff0c\u6bcf\u500b\u9023\u63a5\u57e0\u901f\u7387\u7bc4\u570d\u81ea 40 Gb\/s \u81f3 800 Gb\/s\uff1bQ3400-RA \u5247\u652f\u63f4\u6bcf\u9023\u63a5\u57e0 400 Gb\/s \u6216 800 Gb\/s\u3002\u9019\u9805 800Gbps \u7db2\u8def\u80fd\u529b\u5c07\u524d\u4e00\u4ee3\u7522\u54c1\u7684\u983b\u5bec\u76f4\u63a5\u7ffb\u500d\u3002800G InfiniBand \u4ea4\u63db\u5668\u6280\u8853\u8b93\u7aef\u5230\u7aef\u9023\u7dda\u9054\u5230\u524d\u6240\u672a\u6709\u7684\u901f\u5ea6\u3002<\/p>\n

\u5c0d\u5927\u898f\u6a21 AI \u53e2\u96c6\u4f86\u8aaa\uff0c\u9023\u63a5\u57e0\u5bc6\u5ea6\u540c\u6a23\u81f3\u95dc\u91cd\u8981\u3002Q3400-RA \u8207\u65e9\u671f\u4ea4\u63db\u5668\u76f8\u6bd4\uff0c\u5177\u5099\u986f\u8457\u512a\u52e2\uff1a<\/p>\n

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

\u4ea4\u63db\u5668\u578b\u865f<\/th>\n	\u9023\u63a5\u57e0\u6578\u91cf<\/th>\n	\u9023\u63a5\u57e0\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 \u7531\u6b64\u5c0d\u6bd4\u53ef\u898b\uff0c\u5176\u9023\u63a5\u57e0\u6578\u91cf\u8f03 QM9700 \u63d0\u5347 2.25 \u500d\uff0c\u8f03 QM8700 \u5247\u63d0\u5347 3.6 \u500d\u3002Q3400-RA \u7684\u9ad8 radix\uff08\u9ad8\u9023\u63a5\u57e0\u6578\uff09\u8a2d\u8a08\uff0c\u8b93\u4f60\u80fd\u6253\u9020\u5169\u5c64 fat tree \u62d3\u64b2\uff0c\u5c07\u591a\u9054 10,368 \u5f35\u7db2\u8def\u4ecb\u9762\u5361\u5728\u6700\u4f4e\u5ef6\u9072\u689d\u4ef6\u4e0b\u4e92\u806f\uff0c\u70ba\u6700\u56b4\u82db\u7684\u5de5\u4f5c\u8ca0\u8f09\u63d0\u4f9b\u6700\u4f73\u4f5c\u696d\u5340\u57df\u6027\u3002<\/p>\n Q3400-RA \u4ea6\u63a1\u7528\u98a8\u51b7\u8a2d\u8a08\uff0c\u53ef\u76f4\u63a5\u90e8\u7f72\u65bc\u73fe\u6709\u8cc7\u6599\u4e2d\u5fc3\u4e2d\uff0c\u7121\u9700\u9032\u884c\u5927\u898f\u6a21\u57fa\u790e\u8a2d\u65bd\u6539\u9020\u3002\u6a5f\u7bb1\u914d\u7f6e 8 \u500b\u53ef\u66f4\u63db\u96fb\u6e90\u6a21\u7d44\u8207 10 \u500b\u53ef\u66f4\u63db\u98a8\u6247\uff0c\u65b9\u4fbf\u7dad\u8b77\u3002\u5169\u6b3e\u6a5f\u578b\u7686\u63d0\u4f9b USB 2.0 Type-A \u9023\u63a5\u57e0\u3001\u7ba1\u7406\u9023\u63a5\u57e0\u8207\u4e3b\u63a7\u53f0\u9023\u63a5\u57e0\uff0c\u4fbf\u65bc\u65e5\u5e38\u71df\u904b\u7ba1\u7406\u3002<\/p>\n \u642d\u8f09 SHARP v4 \u7684\u5148\u9032\u7db2\u8def\u5167\u904b\u7b97\u80fd\u529b<\/h3>\n 800G InfiniBand \u5e73\u53f0\u5c0e\u5165\u65b0\u4e00\u4ee3\u7db2\u8def\u5167\u904b\u7b97\u6280\u8853 SHARP v4\u3002\u6b64\u5354\u5b9a\u900f\u904e\u5c07\u96c6\u5408\u64cd\u4f5c\u5378\u8f09\u5230\u4ea4\u63db\u7db2\u8def\u672c\u8eab\uff0c\u52a0\u901f\u76f8\u95dc\u904b\u7b97\u6d41\u7a0b\uff0c\u5f9e\u800c\u964d\u4f4e\u8cc7\u6599\u5728\u7db2\u8def\u4e2d\u7684\u5f80\u8fd4\uff0c\u540c\u6642\u63d0\u5347\u6574\u9ad4\u6548\u7387\u3002<\/p>\n \u8207\u524d\u4e00\u4ee3 SHARP \u6280\u8853\u76f8\u6bd4\uff0cSHARP v4 \u5e36\u4f86 9 \u500d\u7684\u7db2\u8def\u5167\u904b\u7b97\u6548\u80fd\u63d0\u5347\u3002\u9019\u9805\u6539\u9032\u76f4\u63a5\u7e2e\u77ed AI \u8a13\u7df4\u9031\u671f\u3002\u5c0d\u65bc\u57fa\u65bc NCCL \u7684\u5de5\u4f5c\u8ca0\u8f09\uff0c\u63a1\u7528 SHARP \u52a0\u901f\u7684\u96c6\u5408\u901a\u8a0a\u53ef\u5e36\u4f86\u6700\u9ad8 2.5 \u500d\u7684\u6548\u80fd\u63d0\u5347\u3002\u7531\u65bc\u68af\u5ea6\u540c\u6b65\u4e0d\u518d\u6210\u70ba\u74f6\u9838\uff0c\u4f60\u7684\u5927\u578b\u8a9e\u8a00\u6a21\u578b\u8a13\u7df4\u901f\u5ea6\u56e0\u800c\u5927\u5e45\u63d0\u5347\u3002<\/p>\n \u6b64\u4ea4\u63db\u5668\u4ea6\u6574\u5408\u589e\u5f37\u578b\u81ea\u9069\u61c9\u8def\u7531\u529f\u80fd\uff0c\u53ef\u91dd\u5c0d\u52d5\u614b\u6d41\u91cf\u578b\u614b\u512a\u5316\u8def\u5f91\u9078\u64c7\u3002\u4ee5\u9059\u6e2c\u70ba\u57fa\u790e\u7684\u58c5\u585e\u63a7\u5236\u5229\u7528\u5373\u6642\u7db2\u8def\u8cc7\u6599\uff0c\u5728\u58c5\u585e\u767c\u751f\u524d\u4fbf\u9810\u5148\u9632\u7bc4\u3002\u9019\u4e9b\u80fd\u529b\u901a\u529b\u5408\u4f5c\uff0c\u5373\u4fbf\u5728\u9ad8\u8ca0\u8f09\u60c5\u6cc1\u4e0b\u4ecd\u80fd\u7dad\u6301\u4f4e\u65bc 100 \u5948\u79d2\u7684\u9023\u63a5\u57e0\u5230\u9023\u63a5\u57e0\u5ef6\u9072\u3002<\/p>\n NVIDIA \u4e5f\u70ba Quantum-X800 \u5e73\u53f0\u8a2d\u8a08\u4e86\u512a\u7570\u7684\u53ef\u9760\u6027\u6a5f\u5236\u3002\u81ea\u6211\u4fee\u5fa9\u7684\u7db2\u8def\u7e54\u69cb\u63d0\u4f9b\u76ee\u524d\u6700\u5feb\u7684\u5fa9\u539f\u6a5f\u5236\uff0c\u5728\u93c8\u8def\u6545\u969c\u6642\u53ef\u9054\u5230 1,000 \u500d\u66f4\u9ad8\u7684\u5f48\u6027\u3002\u5373\u4f7f\u500b\u5225\u9023\u7dda\u51fa\u73fe\u554f\u984c\uff0c\u4f60\u7684\u57fa\u790e\u8a2d\u65bd\u4f9d\u7136\u80fd\u7dad\u6301\u7a69\u5b9a\u904b\u4f5c\u3002<\/p>\n \u8d85\u4f4e\u5ef6\u9072\u8207\u9ad8\u541e\u5410\u91cf\u7684\u7d44\u5408\uff0c\u4f7f\u9019\u6b3e\u4ea4\u63db\u5668\u975e\u5e38\u9069\u5408\u9ad8\u6548\u80fd\u904b\u7b97\u74b0\u5883\u3002HPC \u53e2\u96c6\u5fc5\u9808\u5728\u6210\u5343\u4e0a\u842c\u500b\u7bc0\u9ede\u4e4b\u9593\u7dad\u6301\u4e00\u81f4\u7684\u6548\u80fd\u8868\u73fe\u3002Quantum-X800 \u5e73\u53f0\u53ef\u5728\u5169\u5c64 fat tree \u62d3\u64b2\u4e2d\u652f\u63f4\u8d85\u904e 10,000 \u500b\u7bc0\u9ede\uff0c\u9019\u4e00\u898f\u6a21\u9060\u9060\u8d85\u51fa\u524d\u4e00\u4ee3\u7522\u54c1\u6240\u80fd\u627f\u8f09\u7684\u53e2\u96c6\u5927\u5c0f\u3002<\/p>\n \u5c0d AI \u5de5\u4f5c\u8ca0\u8f09\u800c\u8a00\uff0c\u6536\u7a6b\u4e0d\u50c5\u6b62\u65bc\u901f\u5ea6\u3002\u8207\u65e9\u671f\u578b\u865f\u76f8\u6bd4\uff0c\u6b64\u4ea4\u63db\u5668\u964d\u4f4e\u4e86\u6bcf bit \u80fd\u8017\u3002\u5728\u7dad\u6301\u5353\u8d8a\u6548\u80fd\u7684\u540c\u6642\uff0c\u4f60\u80fd\u5be6\u73fe\u66f4\u9ad8\u7684\u6bcf\u74e6\u904b\u7b97\u80fd\u529b\u3002Quantum-X800 \u5e73\u53f0\u4ea6\u63d0\u4f9b Quantum-X Photonics \u9078\u9805\uff0c\u5c07\u5149\u5b78\u5143\u4ef6\u76f4\u63a5\u6574\u5408\u81f3\u4ea4\u63db\u6676\u7247\u4e2d\uff0c\u9032\u4e00\u6b65\u964d\u4f4e\u529f\u8017\u8207\u5ef6\u9072\u3002<\/p>\n \u9019\u6b3e 800G InfiniBand \u4ea4\u63db\u5668\u70ba\u73fe\u4ee3 AI \u57fa\u790e\u8a2d\u65bd\u63d0\u4f9b\u4e00\u5957\u5b8c\u6574\u89e3\u6c7a\u65b9\u6848\u3002\u4f60\u5c07\u7372\u5f97\u63a8\u52d5\u5c16\u7aef\u7814\u7a76\u8207\u751f\u7522\u7cfb\u7d71\u6240\u9700\u7684\u983b\u5bec\u3001\u667a\u6167\u8207\u53ef\u9760\u6027\u3002<\/p>\n 800G InfiniBand \u5c0d AI \u7684\u512a\u52e2<\/strong><\/h2>\n \u52a0\u901f AI \u8a13\u7df4\u8207\u63a8\u8ad6<\/h3>\n Quantum-X800 InfiniBand \u4ea4\u63db\u5668\u5728\u5927\u578b\u8a9e\u8a00\u6a21\u578b\u8a13\u7df4\u6548\u7387\u65b9\u9762\u5be6\u73fe 9 \u500d\u63d0\u5347\u3002\u9019\u6b78\u529f\u65bc NVIDIA \u7684 SHARP v4 \u5354\u5b9a\uff0c\u5b83\u5728\u4ea4\u63db\u5668\u4e0a\u76f4\u63a5\u57f7\u884c\u7db2\u8def\u5167\u8cc7\u6599\u5f59\u7e3d\u8207\u96c6\u5408\u64cd\u4f5c\u3002\u85c9\u7531\u5c07\u9019\u4e9b\u901a\u8a0a\u4efb\u52d9\u5f9e\u4f3a\u670d\u5668\u7aef\u5378\u8f09\uff0c\u4ea4\u63db\u5668\u964d\u4f4e\u4e86\u901a\u8a0a\u74f6\u9838\u8207\u4f3a\u670d\u5668\u8ca0\u8f09\uff0c\u5f9e\u800c\u5927\u5e45\u52a0\u5feb\u5146\u5104\u53c3\u6578\u6a21\u578b\u7684\u8a13\u7df4\u9031\u671f\u3002<\/p>\n \u4f60\u540c\u6642\u4e5f\u80fd\u7372\u5f97\u986f\u8457\u7684\u80fd\u6548\u6539\u5584\u3002\u6b64\u4ea4\u63db\u5668\u4f7f\u7528\u5171\u5c01\u88dd\u5149\u5b78\u6280\u8853\uff0c\u53d6\u6d88\u4e86\u53ef\u63d2\u62d4\u5149\u6a21\u7d44\u3002\u6b64\u4e00\u8a2d\u8a08\u7c21\u5316\u4e86\u7269\u6599\u6e05\u55ae\u4e26\u964d\u4f4e\u6210\u672c\uff0c\u5176\u80fd\u6548\u8f03\u50b3\u7d71\u53ef\u63d2\u62d4\u5149\u6a21\u7d44\u9ad8\u51fa 5 \u500d\u3002\u7531\u65bc\u6bcb\u9808 DSP \u91cd\u5b9a\u6642\u5668\uff0c\u7db2\u8def\u5074\u96fb\u529b\u6d88\u8017\u5927\u5e45\u4e0b\u964d\u3002\u9019\u9805\u964d\u4f4e\u76f4\u63a5\u6e1b\u5c11 AI \u8cc7\u6599\u4e2d\u5fc3\u7684\u6574\u9ad4\u6301\u6709\u6210\u672c\uff0c\u96a8\u8457\u57fa\u790e\u8a2d\u65bd\u751f\u547d\u9031\u671f\u63a8\u9032\uff0c\u5c07\u7d2f\u7a4d\u6210\u53ef\u89c0\u7684\u71df\u904b\u6548\u76ca\u3002<\/p>\n \n\n\n\n<\/colgroup>\n\n\n\n\n\n \u6307\u6a19<\/th>\n \u6539\u5584\u5e45\u5ea6<\/th>\n<\/tr>\n \u80fd\u6548<\/td>\n \u8f03\u53ef\u63d2\u62d4\u5149\u6a21\u7d44\u9ad8\u51fa 5 \u500d<\/td>\n<\/tr>\n AI \u61c9\u7528\u6301\u7e8c\u904b\u884c\u6642\u9593<\/td>\n \u97cc\u6027\u63d0\u5347 5 \u500d<\/td>\n<\/tr>\n \u6d1e\u5bdf\u6642\u9593\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 \u9019\u4e09\u9805\u6307\u6a19\u51f8\u986f Quantum-X800 \u5e73\u53f0\u5728\u71df\u904b\u5c64\u9762\u7684\u512a\u52e2\u3002\u7531\u65bc\u7701\u53bb\u4e86 DSP \u91cd\u5b9a\u6642\u5668\uff0c\u6b64\u4ea4\u63db\u5668\u4e5f\u80fd\u9032\u4e00\u6b65\u964d\u4f4e\u5ef6\u9072\u3002\u5c0d\u5927\u898f\u6a21 AI \u8a13\u7df4\u800c\u8a00\uff0c\u9019\u7a2e\u8d85\u4f4e\u5ef6\u9072\u81f3\u95dc\u91cd\u8981\uff0c\u6bcf\u4e00\u5fae\u79d2\u90fd\u5f4c\u8db3\u73cd\u8cb4\u3002\u4f60\u7684 GPU \u5c07\u82b1\u66f4\u5c11\u6642\u9593\u7b49\u5f85\u8cc7\u6599\uff0c\u800c\u80fd\u628a\u66f4\u591a\u6642\u9593\u6295\u5165\u904b\u7b97\u3002InfiniBand \u67b6\u69cb\u63d0\u4f9b\u6a19\u6e96\u4e59\u592a\u7db2\u96e3\u4ee5\u9054\u6210\u7684\u7121\u640d\u901a\u8a0a\uff0c\u78ba\u4fdd\u8a13\u7df4\u4f5c\u696d\u4e0d\u6703\u56e0\u5c01\u5305\u907a\u5931\u6216\u7db2\u8def\u58c5\u585e\u800c\u88ab\u8feb\u4e2d\u65b7\u3002\u6700\u7d42\u6210\u679c\u662f\uff1a\u66f4\u5feb\u901f\u7684\u6a21\u578b\u958b\u767c\u9031\u671f\uff0c\u4ee5\u53ca\u6574\u500b\u53e2\u96c6\u4e2d\u66f4\u9ad8\u7684 GPU \u4f7f\u7528\u7387\u3002<\/p>\n \u85c9\u7531 NVIDIA Quantum-X800 \u5be6\u73fe\u53ef\u64f4\u5145\u6027<\/h3>\n Quantum-X800 \u5e73\u53f0\u652f\u63f4\u5f9e\u5c0f\u578b\u7814\u7a76\u5be6\u9a57\u5ba4\u5230\u8d85\u5927\u898f\u6a21\u8cc7\u6599\u4e2d\u5fc3\u7684\u5e73\u6ed1\u64f4\u5145\u3002\u8207\u524d\u4e00\u4ee3\u7522\u54c1\u76f8\u6bd4\uff0cNVIDIA \u9019\u6b3e\u4ea4\u63db\u5668\u5728\u53ef\u64f4\u5145\u6027\u65b9\u9762\u5c55\u73fe\u51fa\u660e\u986f\u512a\u52e2\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\u524d\u4e00\u4ee3\uff09<\/th>\n<\/tr>\n \u6bcf\u53f0\u4ea4\u63db\u5668\u9023\u63a5\u57e0\u6578<\/td>\n 144\uff08800 Gb\/s\uff09<\/td>\n 64\uff08400 Gb\/s\uff09<\/td>\n<\/tr>\n \u6700\u5927\u4e3b\u6a5f\u9023\u7dda\u6578\uff08\u5169\u5c64 Fat-Tree\uff09<\/td>\n >10,000\uff08800 Gb\/s\uff09<\/td>\n –<\/td>\n<\/tr>\n \u7db2\u8def\u5167\u904b\u7b97<\/td>\n \u7b2c\u56db\u4ee3 SHARP\uff08\u652f\u63f4 FP8\uff09<\/td>\n –<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n \u9ad8\u9023\u63a5\u57e0\u5bc6\u5ea6\u8b93\u4f60\u5f97\u4ee5\u5728\u5169\u5c64 fat tree \u62d3\u64b2\u4e2d\u9023\u63a5\u8d85\u904e 10,000 \u500b\u7bc0\u9ede\u3002\u5982\u6b64\u898f\u6a21\u8db3\u4ee5\u627f\u8f09\u6700\u5927\u578b\u7684 AI \u5de5\u4f5c\u8ca0\u8f09\uff0c\u4e14\u7121\u9700\u69cb\u5efa\u8907\u96dc\u7684\u591a\u5c64\u7db2\u8def\u8a2d\u8a08\u3002\u96a8\u8457\u904b\u7b97\u9700\u6c42\u589e\u52a0\uff0c\u4f60\u7684\u96f2\u7aef\u57fa\u790e\u8a2d\u65bd\u53ef\u4ee5\u9806\u66a2\u64f4\u5f35\uff0c\u6bcf\u4e00\u500b\u65b0\u7bc0\u9ede\u90fd\u80fd\u5728\u4e0d\u6253\u4e82\u65e2\u6709\u5de5\u4f5c\u6d41\u7a0b\u7684\u60c5\u6cc1\u4e0b\u52a0\u5165\u53e2\u96c6\u3002<\/p>\n Q3400-RA \u63a1\u7528\u98a8\u51b7\u8a2d\u8a08\uff0c\u53ef\u8207\u73fe\u6709\u8cc7\u6599\u4e2d\u5fc3\u7684\u6563\u71b1\u57fa\u790e\u8a2d\u65bd\u76f8\u5bb9\u3002<\/p>\n \u98a8\u51b7\u5e73\u53f0\u50c5\u652f\u63f4\u5177\u6563\u71b1\u9c2d\u7247\uff08Finned Top\uff09\u7248\u672c\u7684 1.6T \u6536\u767c\u5668\u3002\u5e73\u9802\uff08Flat Top\uff09\u7248\u672c\u4ef0\u8cf4\u51b7\u677f\u5c0e\u71b1\uff0c\u53ea\u9069\u7528\u65bc\u6db2\u51b7\u67b6\u69cb\u3002Q3400-RA \u4f7f\u7528\u7684\u662f Finned Top \u6536\u767c\u5668\uff0c\u9019\u4e00\u9ede\u4e5f\u8b49\u5be6\u4e86\u5176\u98a8\u51b7\u8a2d\u8a08\u53ca\u8207\u65e2\u6709\u8cc7\u6599\u4e2d\u5fc3\u6563\u71b1\u7cfb\u7d71\u7684\u76f8\u5bb9\u6027\u3002<\/p>\n \u5f97\u76ca\u65bc\u6b64\u98a8\u51b7\u65b9\u6848\uff0c\u4f60\u53ef\u4ee5\u5728\u6a19\u6e96\u6a5f\u67b6\u4e2d\u90e8\u7f72 Q3400-RA\uff0c\u800c\u4e0d\u5fc5\u5efa\u7f6e\u6db2\u51b7\u7cfb\u7d71\u30024U \u6a5f\u7bb1\u53ef\u76f4\u63a5\u653e\u5165\u65e2\u6709\u8cc7\u6599\u4e2d\u5fc3\u6a5f\u6ac3\u4e2d\u3002\u4f60\u4e5f\u56e0\u6b64\u7121\u9808\u70ba\u8d85\u5927\u898f\u6a21\u8cc7\u6599\u4e2d\u5fc3\u6539\u9020\u6563\u71b1\u7cfb\u7d71\uff0c\u5f9e\u800c\u907f\u514d\u984d\u5916\u6210\u672c\u8207\u8907\u96dc\u5ea6\u3002\u9019\u9805\u8a2d\u8a08\u6289\u64c7\u8b93\u90e8\u7f72\u6d41\u7a0b\u65e2\u7c21\u55ae\u53c8\u53ef\u9810\u671f\u3002<\/p>\n InfiniBand \u6280\u8853\u8207 NVIDIA \u5e73\u53f0\u8a2d\u8a08\u7684\u7d50\u5408\uff0c\u4f7f\u9019\u6b3e\u4ea4\u63db\u5668\u6210\u70ba HPC \u74b0\u5883\u7684\u7406\u60f3\u9078\u64c7\u3002\u5176\u8d85\u4f4e\u5ef6\u9072\u8207\u9ad8\u541e\u5410\u91cf\u8db3\u4ee5\u652f\u6490\u6700\u56b4\u82db\u7684 HPC \u5de5\u4f5c\u8ca0\u8f09\u3002\u4f60\u53ef\u4ee5\u5728\u540c\u4e00\u5f35\u7db2\u8def\u4e0a\u540c\u6642\u57f7\u884c AI \u8a13\u7df4\u4efb\u52d9\u8207\u50b3\u7d71 HPC \u6a21\u64ec\u3002\u96a8\u8457\u6a21\u578b\u898f\u6a21\u4e0d\u65b7\u81a8\u8139\uff0c\u9019\u6b3e InfiniBand \u4ea4\u63db\u5668\u4ea6\u80fd\u6eff\u8db3\u6301\u7e8c\u6210\u9577\u7684 GPU \u904b\u7b97\u9700\u6c42\u3002800 Gb\/s \u7684\u9023\u63a5\u57e0\u983b\u5bec\u78ba\u4fdd GPU \u4e0d\u6703\u56e0\u8cc7\u6599\u4e0d\u8db3\u800c\u300c\u9913\u6b7b\u300d\u3002\u9019\u7a2e\u53ef\u64f4\u5145\u6027\u8b93\u4f60\u7684\u57fa\u790e\u8a2d\u65bd\u70ba\u672a\u4f86 AI \u7a81\u7834\u9810\u5148\u92ea\u8def\u3002<\/p>\n \n <\/div>\n<\/div>\n NVIDIA Quantum-X800 \u5e73\u53f0\u8c61\u5fb5 AI \u7db2\u8def\u9818\u57df\u7684\u4e00\u9805\u95dc\u9375\u9032\u5c55\u3002\u5176 800G InfiniBand \u6280\u8853\u70ba\u4f60\u7684 AI \u5de5\u4f5c\u8ca0\u8f09\u5e36\u4f86\u524d\u6240\u672a\u6709\u7684\u901f\u5ea6\u3001\u66f4\u5f37\u7684\u53ef\u64f4\u5145\u6027\u4ee5\u53ca\u66f4\u9ad8\u7684\u6548\u7387\u3002\u9019\u6b3e 800G InfiniBand \u4ea4\u63db\u5668\u5fb9\u5e95\u6d88\u9664\u4e86\u62d6\u6162 GPU \u8a13\u7df4\u8207\u63a8\u8ad6\u7684\u7db2\u8def\u74f6\u9838\u3002<\/p>\n NVIDIA \u7684\u6280\u8853\u85cd\u5716\u4e5f\u5c07\u9019\u9805\u5275\u65b0\u6301\u7e8c\u63a8\u5411\u65b0\u9ad8\u3002\u4e0b\u4e00\u4ee3\u5e73\u53f0 Quantum-X Photonics \u5c07\u77fd\u5149\u5b50\u6280\u8853\u76f4\u63a5\u6574\u5408\u81f3\u4ea4\u63db\u5668\u5c01\u88dd\uff0c\u53ef\u5be6\u73fe 3.5 \u500d\u7684\u80fd\u6548\u63d0\u5347\u8207 10 \u500d\u66f4\u9ad8\u7684\u97cc\u6027\u3002\u61c9\u7528\u7a0b\u5f0f\u53ef\u5728\u4e0d\u4e2d\u65b7\u7684\u60c5\u6cc1\u4e0b\u904b\u884c\u66f4\u9577\u6642\u9593\uff0c\u6700\u9577\u53ef\u63d0\u5347 5 \u500d\u3002\u672a\u4f86\u7684 InfiniBand \u9023\u7dda\u5c07\u9081\u5411 1.6T \u901f\u7387\uff0c\u4ee5\u652f\u63f4\u5146\u7d1a\u53c3\u6578\u6a21\u578b\u3002<\/p>\n \u4e0d\u59a8\u601d\u8003\u9019\u9805\u6280\u8853\u5c07\u5982\u4f55\u91cd\u5851\u4f60\u7684\u8cc7\u6599\u4e2d\u5fc3\u7b56\u7565\u3002Quantum-X800 \u5e73\u53f0\u8b93\u4f60\u7684\u57fa\u790e\u8a2d\u65bd\u5982\u4eca\u4fbf\u70ba\u660e\u65e5\u7684 AI \u7a81\u7834\u505a\u597d\u6e96\u5099\u3002<\/p>\n \u5e38\u898b\u554f\u984c\uff08FAQ\uff09<\/strong><\/h2>\n Quantum-X800 \u8207\u820a\u6b3e\u4ea4\u63db\u5668\u6709\u4f55\u4e0d\u540c\uff1f<\/h3>\n \u5b83\u63d0\u4f9b 144 \u500b\u9023\u63a5\u57e0\uff0c\u6bcf\u500b\u9023\u63a5\u57e0\u652f\u63f4 800 Gb\/s \u7684 InfiniBand \u901f\u7387\u3002\u9019\u6b3e NVIDIA \u4ea4\u63db\u5668\u63a1\u7528 SHARP v4 \u52a0\u901f\u96c6\u5408\u64cd\u4f5c\uff0c\u8b93\u5927\u578b AI \u6a21\u578b\u7684\u8a13\u7df4\u6548\u7387\u63d0\u5347 9 \u500d\u3002<\/p>\n SHARP v4 \u5982\u4f55\u63d0\u5347 AI \u8a13\u7df4\u6548\u7387\uff1f<\/h3>\n NVIDIA \u7684 SHARP v4 \u5c07\u96c6\u5408\u901a\u8a0a\u5f9e GPU \u5378\u8f09\u5230\u4ea4\u63db\u7db2\u8def\u4e2d\u3002\u9019\u9805 NVIDIA \u6280\u8853\u6e1b\u5c11\u8cc7\u6599\u79fb\u52d5\uff0c\u5728\u8a13\u7df4\u5de5\u4f5c\u8ca0\u8f09\u4e2d\u53ef\u5e36\u4f86\u6700\u9ad8 9 \u500d\u7684\u6548\u80fd\u63d0\u5347\u3002<\/p>\n Quantum-X800 \u662f\u5426\u80fd\u5728\u73fe\u6709\u8cc7\u6599\u4e2d\u5fc3\u4e2d\u4f7f\u7528\uff1f<\/h3>\n \u53ef\u4ee5\u3002Q3400-RA \u63a1\u7528\u98a8\u51b7\u8a2d\u8a08\uff0c\u53ef\u76f4\u63a5\u90e8\u7f72\u5728\u6a19\u6e96\u6a5f\u67b6\u4e2d\uff0c\u7121\u9808\u6db2\u51b7\u7cfb\u7d71\u3002\u9019\u8b93\u5b83\u80fd\u8f15\u9b06\u6574\u5408\u81f3\u8d85\u5927\u898f\u6a21\u8cc7\u6599\u4e2d\u5fc3\u8207\u73fe\u6709\u96f2\u7aef\u57fa\u790e\u8a2d\u65bd\u4e2d\u3002<\/p>\n \u9019\u6b3e\u4ea4\u63db\u5668\u5982\u4f55\u652f\u63f4\u5927\u578b AI \u53e2\u96c6\uff1f<\/h3>\n Quantum-X800 \u80fd\u5728\u5169\u5c64 fat tree \u62d3\u64b2\u4e2d\u9023\u63a5\u8d85\u904e 10,000 \u500b\u7bc0\u9ede\uff0c\u6bcf\u500b\u7bc0\u9ede\u64c1\u6709 800 Gb\/s \u983b\u5bec\u3002\u9019\u6b3e NVIDIA \u4ea4\u63db\u5668\u53ef\u5f9e\u7814\u7a76\u5be6\u9a57\u5ba4\u4e00\u8def\u64f4\u5c55\u81f3\u8d85\u5927\u898f\u6a21 GPU \u53e2\u96c6\u3002<\/p>\n \u6b64\u4ea4\u63db\u5668\u80fd\u61c9\u5c0d\u672a\u4f86\u66f4\u5927\u898f\u6a21\u7684 AI \u6a21\u578b\u55ce\uff1f<\/h3>\n \u53ef\u4ee5\u3002800 Gb\/s \u7684 InfiniBand \u983b\u5bec\u53ef\u96a8\u4f60\u7684\u9700\u6c42\u64f4\u5145\u3002\u9019\u9805 NVIDIA \u5e73\u53f0\u652f\u63f4\u8d85\u904e 10,000 \u500b\u7bc0\u9ede\uff0c\u8db3\u4ee5\u56e0\u61c9\u5146\u7d1a\u53c3\u6578\u6a21\u578b\u3002\u4f60\u7684\u57fa\u790e\u8a2d\u65bd\u5f97\u4ee5\u5728\u4e0d\u4e2d\u65b7\u696d\u52d9\u7684\u60c5\u6cc1\u4e0b\u6301\u7e8c\u64f4\u5bb9\u3002<\/p>\n","protected":false},"excerpt":{"rendered":" NVIDIA \u63a8\u51fa\u4e86 Quantum-X800 InfiniBand \u4ea4\u63db\u5668\uff0c\u9019\u662f AI \u57fa\u790e\u8a2d\u65bd \u9818\u57df\u7684\u4e00\u9805 […]<\/p>\n Read More…<\/a><\/p>\n","protected":false},"author":2,"featured_media":34277,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[4815],"tags":[12960,12961,12042,12043,1729],"class_list":["post-34280","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-america-dedicated-server-tc","tag-800g-infiniband-switch","tag-gpu-cluster","tag-nvidia-quantum-x800","tag-ai-networking","tag-high-performance-computing-tc"],"acf":[],"yoast_head":"\n