\u67b6\u69cb\u6838\u5fc3\uff1aNVIDIA H200 \u7684 HBM3e \u8a18\u61b6\u9ad4\u5347\u7d1a<\/strong><\/h2>\n

\u8981\u5728\u8a13\u7df4\u8d85\u5927\u898f\u6a21\u4eba\u5de5\u667a\u6167\u6a21\u578b\u6642\u907f\u514d\u60f1\u4eba\u7684\u7b97\u529b\u74f6\u9838\uff0c\u4f60\u9700\u8981\u5f37\u5927\u7684\u786c\u9ad4\u5143\u4ef6\u3002NVIDIA H200 \u900f\u904e\u5c0e\u5165 141GB \u7684\u5148\u9032 HBM3e \u8a18\u61b6\u9ad4\uff0c\u89e3\u6c7a\u50b3\u7d71\u7684\u8a18\u61b6\u9ad4\u9650\u5236\u3002\u8207 H100 SXM \u8655\u7406\u5668\u76f8\u6bd4\uff0c\u9019\u9805\u5347\u7d1a\u5e36\u4f86 76% \u7684\u5bb9\u91cf\u589e\u52a0\u8207 43% \u7684\u983b\u5bec\u63d0\u5347\u3002<\/p>\n

4.8TB\/s \u8a18\u61b6\u9ad4\u983b\u5bec\u8207 Tensor Core \u5229\u7528\u7387<\/h3>\n

H200 \u63d0\u4f9b 141 GB HBM3e \u8a18\u61b6\u9ad4\u8207 4.8 TB\/s \u983b\u5bec\uff0c\u662f H100 \u8a18\u61b6\u9ad4\u5bb9\u91cf\u7684 1.44 \u500d\u3001\u983b\u5bec\u7684 1.8 \u500d\u3002<\/p>\n

\u9019\u689d\u8d85\u5bec\u8a18\u61b6\u9ad4\u901a\u9053\u5fb9\u5e95\u6539\u8b8a\u7cfb\u7d71\u8655\u7406\u8907\u96dc\u6578\u5b78\u8cc7\u6599\u7684\u65b9\u5f0f\u3002\u7576\u4f60\u8a13\u7df4\u8d85\u5927\u898f\u6a21\u4eba\u5de5\u667a\u6167\u6a21\u578b\u6642\uff0cTensor Core \u5f80\u5f80\u6703\u56e0\u7b49\u5f85\u8cc7\u6599\u800c\u66ab\u505c\u3002\u9ad8\u983b\u5bec\u80fd\u4ee5\u5feb\u65bc\u6578\u5b78\u904b\u7b97\u5f15\u64ce\u7684\u901f\u5ea6\u50b3\u8f38\u8cc7\u6599\uff0c\u5f9e\u800c\u6d88\u9664\u9019\u4e9b\u505c\u9813\u3002<\/p>\n

\u5728\u81ea\u56de\u6b78\u6587\u5b57\u751f\u6210\u7684\u89e3\u78bc\u968e\u6bb5\uff0c4.8 TB\/s \u7684\u8a18\u61b6\u9ad4\u983b\u5bec\u6709\u52a9\u65bc\u9632\u6b62 GPU \u5728\u8a18\u61b6\u9ad4 I\/O \u64cd\u4f5c\u4e0b\u767c\u751f\u505c\u6eef\u3002\u9019\u80fd\u76f4\u63a5\u78ba\u4fdd Tensor Core \u4fdd\u6301\u6d3b\u8e8d\uff0c\u7279\u5225\u662f\u5728\u5927\u6279\u91cf\u8a13\u7df4\u60c5\u5883\u4e2d\uff0c\u539f\u672c\u786c\u9ad4\u53ef\u80fd\u958b\u59cb\u963b\u585e\u30024.8 TB\/s \u7684\u8a18\u61b6\u9ad4\u983b\u5bec\u7e2e\u77ed\u4e86\u8cc7\u6599\u5bc6\u96c6\u578b\u4efb\u52d9\u6240\u9700\u6642\u9593\uff0c\u8b93\u6a21\u578b\u8a13\u7df4\u66f4\u5feb\u901f\uff0c\u4e26\u63d0\u5347 Tensor Core \u5728\u8907\u96dc\u904b\u7b97\u4e2d\u7684\u4f7f\u7528\u6548\u7387\u3002<\/p>\n

\u5728\u90e8\u7f72\u9019\u5957\u8a18\u61b6\u9ad4\u7cfb\u7d71\u6642\uff0c\u4f60\u53ef\u4ee5\u770b\u5230\u4e00\u7cfb\u5217\u5be6\u969b\u512a\u52e2\uff1a<\/p>\n

\n
• \u7576\u6279\u6b21\u5927\u5c0f\u589e\u52a0\u5230\u7531\u8a18\u61b6\u9ad4\u50b3\u8f38\u4e3b\u5c0e\u8655\u7406\u6642\u9593\u6642\uff0c\u786c\u9ad4\u904b\u4f5c\u901f\u5ea6\u660e\u986f\u63d0\u9ad8\u3002<\/li>\n
• \u7576\u8a18\u61b6\u9ad4\u5bb9\u91cf\u6216\u983b\u5bec\u6210\u70ba\u4e3b\u8981\u7cfb\u7d71\u74f6\u9838\u6642\uff0c\u6548\u80fd\u512a\u52e2\u5c24\u5176\u660e\u986f\u3002<\/li>\n
• \u55ae\u9846 NVIDIA H200 \u5c31\u80fd\u627f\u64d4\u539f\u672c\u9700\u8981\u5169\u9846 H100 \u624d\u80fd\u8655\u7406\u7684\u6c89\u91cd\u5de5\u4f5c\u8ca0\u8f09\uff0c\u964d\u4f4e\u591a GPU \u90e8\u7f72\u7684\u8907\u96dc\u5ea6\u3002<\/li>\n
• \u5c0d\u65bc\u80fd\u653e\u9032 80 GB \u8a18\u61b6\u9ad4\u7684\u5c0f\u6a21\u578b\uff0c\u7531\u65bc\u74f6\u9838\u4e0d\u5728\u8a18\u61b6\u9ad4\uff0c\u56e0\u6b64\u6548\u80fd\u5927\u81f4\u76f8\u7576\u3002<\/li>\n<\/ul>\n

  \u4f01\u696d\u56de\u5831\u986f\u793a\uff0c\u76f8\u8f03 H100 \u914d\u7f6e\uff0cH200 \u53e2\u96c6\u53ef\u5c07\u591a\u5341\u5104\u53c3\u6578\u6a21\u578b\u7684\u5fae\u8abf\u6642\u9593\u7e2e\u77ed\u591a\u9054 40%\u3002<\/p>\n

  \u7576\u6a21\u578b\u5df2\u7121\u6cd5\u5728 H100 \u4e0a\u300c\u8212\u9069\u300d\u5bb9\u7d0d\u6642\uff0cH200 \u7684\u512a\u52e2\u5c24\u5176\u660e\u986f\u3002\u5c0d\u65bc\u5927\u6279\u6b21\u8a13\u7df4\u3001\u8d85\u9577\u8108\u7d61\u9577\u5ea6\u6216\u9ad8\u5ea6\u5403\u8a18\u61b6\u9ad4\u7684\u5de5\u4f5c\u8ca0\u8f09\uff0cH200 \u80fd\u5e36\u4f86\u66f4\u6d41\u66a2\u3001\u66f4\u5feb\u901f\u7684\u9ad4\u9a57\u3002<\/p>\n

  \u964d\u4f4e\u591a GPU \u53e2\u96c6\u4e2d\u7684\u901a\u8a0a\u5ef6\u9072<\/h3>\n

  \u591a\u7bc0\u9ede\u8a13\u7df4\u9700\u8981\u5728\u4e0d\u540c\u4f3a\u670d\u5668\u7bc0\u9ede\u4e4b\u9593\u6301\u7e8c\u5171\u4eab\u8cc7\u6599\u3002\u7576\u4f60\u8a13\u7df4\u66f4\u5927\u898f\u6a21\u7684\u6a21\u578b\u6642\uff0c\u7db2\u8def\u540c\u6b65\u958b\u92b7\u5e38\u5e38\u6703\u5728 All-Reduce \u7b49\u96c6\u5408\u901a\u8a0a\u64cd\u4f5c\u4e2d\u5f62\u6210\u56b4\u5cfb\u7684\u6548\u80fd\u74f6\u9838\u3002<\/p>\n

  \u73fe\u4ee3\u53e2\u96c6\u7db2\u8def\u67b6\u69cb\u53ef\u4ee5\u5728\u5206\u6563\u5f0f\u57fa\u790e\u8a2d\u65bd\u4e2d\u6d88\u9664\u9019\u4e9b\u5ef6\u9072\uff1a<\/p>\n

  \n
  • \u57fa\u65bc RDMA over Converged Ethernet \u7684\u7db2\u8def\u74b0\u5883\uff0c\u53ef\u4ee5\u5728 82 \u5206\u9418\u5167\u5b8c\u6210\u5927\u578b\u8a9e\u8a00\u6a21\u578b\u5fae\u8abf\uff0c\u70ba Llama 3.1 70B \u7b49\u5927\u578b\u6a21\u578b\u5e36\u4f86\u986f\u8457\u52a0\u901f\u3002<\/li>\n
  • \u64f4\u5145\u6548\u7387\u5927\u5e45\u63d0\u5347\uff0c\u7279\u5225\u662f\u5728 70B \u7b49\u7d1a\u7684\u5927\u578b\u6a21\u578b\u4e2d\uff0c\u82e5\u7f3a\u5c11\u6700\u4f73\u5316\u7684\u7db2\u8def\u5354\u5b9a\uff0c\u8a13\u7df4\u6642\u9593\u6703\u6025\u907d\u62c9\u9577\u3002<\/li>\n
  • \u66f4\u4f4e\u5ef6\u9072\u7684\u8a2d\u8a08\u4f9d\u8cf4\u4e59\u592a\u7db2\u8def\u4e0a\u7684\u76f4\u63a5\u8a18\u61b6\u9ad4\u5b58\u53d6\uff0c\u6d88\u9664\u4e0d\u5fc5\u8981\u7684 CPU \u8655\u7406\uff0c\u964d\u4f4e\u7bc0\u9ede\u9593\u7aef\u5230\u7aef\u901a\u8a0a\u5ef6\u9072\u3002<\/li>\n
  • \u57fa\u790e\u8a2d\u65bd\u6210\u672c\u7bc0\u7701\u6301\u7e8c\u7d2f\u7a4d\uff0c\u56e0\u70ba\u66f4\u5feb\u7684\u5fae\u8abf\u4efb\u52d9\u9700\u8981\u66f4\u5c11\u7684\u8a08\u7b97\u5c0f\u6642\u6578\uff0c\u5f9e\u800c\u76f4\u63a5\u964d\u4f4e\u71df\u904b\u652f\u51fa\u3002<\/li>\n<\/ul>\n

    \u900f\u904e\u79fb\u9664 CPU \u8655\u7406\u5ef6\u9072\uff0c\u4f60\u7684\u591a\u7bc0\u9ede\u53e2\u96c6\u80fd\u7dad\u6301\u9ad8\u8cc7\u6599\u541e\u5410\u3002\u5f37\u5927\u7684\u8a18\u61b6\u9ad4\u983b\u5bec\u8207\u66f4\u4f4e\u7684\u7db2\u8def\u5ef6\u9072\u7d50\u5408\uff0c\u8b93\u6240\u6709\u9023\u7dda\u7684\u8655\u7406\u5668\u5728\u9577\u6642\u9593\u8a13\u7df4\u671f\u9593\u59cb\u7d42\u7dad\u6301\u9ad8\u8f09\u904b\u4f5c\u3002<\/p>\n

    \u8a13\u7df4\u57fa\u6e96\u8207\u57fa\u790e\u6a21\u578b\u541e\u5410\u91cf<\/strong><\/h2>\n

    \u4f60\u53ef\u4ee5\u900f\u904e\u5be6\u8b49\u57fa\u6e96\u6e2c\u8a66\u4f86\u8a55\u4f30\u8655\u7406\u5668\u7684\u771f\u5be6\u901f\u5ea6\u3002\u6a19\u6e96\u5316\u7684\u57fa\u6e96\u6e2c\u8a66\u5957\u4ef6\uff0c\u7528\u65bc\u8861\u91cf\u8907\u96dc\u786c\u9ad4\u7cfb\u7d71\u5728\u5be6\u969b\u904b\u884c\u4e2d\u7684\u6548\u80fd\u8868\u73fe\u3002\u7368\u7acb\u7684\u57fa\u6e96\u8a18\u9304\u6e05\u695a\u5c55\u793a\u4e86\u5927\u578b\u57fa\u790e\u6a21\u578b\u8ca0\u8f09\u4e0b\u7684\u8655\u7406\u80fd\u529b\u63d0\u5347\u3002<\/p>\n

    MLPerf v4.0 \u6307\u6a19\u8207\u8a13\u7df4\u52a0\u901f<\/h3>\n

    \u6a19\u6e96\u5316\u57fa\u6e96\u6e2c\u8a66\u5728\u9ad8\u5f37\u5ea6\u4f01\u696d\u8a13\u7df4\u4efb\u52d9\u4e2d\u5c55\u793a\u51fa\u660e\u986f\u7684\u6548\u80fd\u589e\u76ca\u3002\u7368\u7acb\u7522\u696d\u7d44\u7e54\u900f\u904e\u56b4\u8b39\u7684\u6e2c\u8a66\u6d41\u7a0b\uff0c\u9a57\u8b49\u786c\u9ad4\u5728\u771f\u5be6\u5de5\u4f5c\u8ca0\u8f09\u4e0b\u7684\u8868\u73fe\u3002MLPerf Training v4.0 \u6e2c\u8a66\u5728\u56b4\u683c\u7684\u91cf\u6e2c\u6a19\u6e96\u4e0b\uff0c\u9a57\u8b49\u4e86\u771f\u5be6\u7684\u8a13\u7df4\u52a0\u901f\u6578\u64da\u3002<\/p>\n

    \u5728 MLPerf Training v4.0 \u4e2d\uff0cH200 \u5728\u55ae\u500b 8 GPU \u7bc0\u9ede\u4e0a\uff0c\u4ee5 24.7 \u5206\u9418\u5b8c\u6210 Llama 2 70B \u7684 LoRA \u5fae\u8abf\uff0c\u76f8\u8f03 H100 \u63d0\u5347 14%\u3002<\/p>\n

    \u4f60\u7121\u9700\u66f4\u52d5\u73fe\u6709\u8a13\u7df4\u7a0b\u5f0f\u78bc\u6216\u57fa\u790e\u8a2d\u65bd\u8a2d\u5b9a\uff0c\u5c31\u80fd\u7372\u5f97\u66f4\u5feb\u7684\u8a13\u7df4\u5b8c\u6210\u6642\u9593\u3002\u8a18\u61b6\u9ad4\u5bb9\u91cf\u7684\u64f4\u5145\uff0c\u6e1b\u5c11\u4e86\u5728\u5927\u898f\u6a21\u6578\u5b78\u9663\u5217\u904b\u7b97\u904e\u7a0b\u4e2d\u7684\u8655\u7406\u505c\u9813\u3002\u904b\u7b97\u5f15\u64ce\u53ef\u4ee5\u66f4\u5feb\u901f\u5730\u8655\u7406\u8907\u96dc\u6a21\u578b\u6b0a\u91cd\uff0c\u540c\u6642\u7dad\u6301\u7cbe\u78ba\u7684\u6d6e\u9ede\u904b\u7b97\u7d50\u679c\u3002\u66f4\u9ad8\u7684\u8a13\u7df4\u541e\u5410\u91cf\u76f4\u63a5\u6e1b\u5c11\u6574\u500b\u904b\u7b97\u8cc7\u6e90\u6c60\u4e2d\u7684\u9592\u7f6e\u6642\u9593\u3002<\/p>\n

    \u52a0\u901f\u7684\u57f7\u884c\u901f\u5ea6\u5c07\u5fb9\u5e95\u6539\u8b8a\u4f60\u7684\u8edf\u9ad4\u958b\u767c\u7bc0\u594f\u3002\u5de5\u7a0b\u5718\u968a\u53ef\u4ee5\u5728\u76f8\u540c\u7684\u5de5\u4f5c\u9031\u671f\u5167\u57f7\u884c\u66f4\u591a\u8a13\u7df4\u8fed\u4ee3\u3002\u66f4\u77ed\u7684\u8a13\u7df4\u9031\u671f\uff0c\u8b93\u8cc7\u6599\u79d1\u5b78\u5bb6\u80fd\u66f4\u5feb\u6e2c\u8a66\u65b0\u7684\u8d85\u53c3\u6578\u914d\u7f6e\u3002\u4f60\u7684\u7d44\u7e54\u53ef\u4ee5\u5728\u66f4\u5c11\u7684\u5929\u6578\u5167\uff0c\u5c07\u57fa\u790e\u6a21\u578b\u5f9e\u5be6\u9a57\u968e\u6bb5\u63a8\u9032\u5230\u6b63\u5f0f\u4e0a\u7dda\u3002\u8a13\u7df4\u6642\u9593\u7684\u7e2e\u77ed\u52a0\u901f\u4e86\u7522\u54c1\u767c\u4f48\uff0c\u70ba\u4f60\u7684\u696d\u52d9\u5e36\u4f86\u7af6\u722d\u512a\u52e2\u3002<\/p>\n

    \u52a0\u901f\u5982 Meta Llama 2 \u7b49\u5927\u578b\u6a21\u578b<\/h3>\n

    \u5927\u578b\u8a9e\u8a00\u6a21\u578b\u5728\u9810\u8a13\u7df4\u8207\u5fae\u8abf\u6642\uff0c\u9700\u8981\u6975\u9ad8\u7684\u8a18\u61b6\u9ad4\u983b\u5bec\uff0c\u4ee5\u6301\u7e8c\u70ba\u6578\u5b78\u904b\u7b97\u5f15\u64ce\u8f38\u9001\u77e9\u9663\u8cc7\u6599\uff0c\u907f\u514d\u7b97\u529b\u9592\u7f6e\u3002\u516b\u5361 HGX H200 \u7cfb\u7d71\u5728\u5fae\u8abf Llama 2 70B \u6642\uff0c\u541e\u5410\u91cf\u53ef\u8d85\u904e 15,000 tokens\/\u79d2\u3002\u8207\u4f7f\u7528\u540c\u4e00 NeMo \u6846\u67b6\u7684 A100 GPU \u76f8\u6bd4\uff0c\u8a72\u7cfb\u7d71\u63d0\u4f9b\u6700\u9ad8 4.2 \u500d\u7684\u541e\u5410\u63d0\u5347\u3002<\/p>\n

    \u9f90\u5927\u7684\u53c3\u6578\u898f\u6a21\u8b93\u6df1\u5ea6\u5b78\u7fd2\u4efb\u52d9\u5c0d\u786c\u9ad4\u8a18\u61b6\u9ad4\u7522\u751f\u6975\u5927\u58d3\u529b\u3002\u64f4\u5145\u8a18\u61b6\u9ad4\u5bb9\u91cf\uff0c\u53ef\u4ee5\u5728\u4e0d\u540c\u53c3\u6578\u898f\u6a21\u4e0b\u89e3\u9396\u66f4\u5927\u7684\u6279\u6b21\u5927\u5c0f\uff1a<\/p>\n

    \n
    • \u5728\u5168\u7cbe\u5ea6\u6a21\u5f0f\u4e0b\uff0cLlama-70B \u8a13\u7df4\u7684\u6bcf GPU \u6279\u6b21\u5927\u5c0f\uff0c\u5f9e H100 \u4e0a\u7684 4 \u63d0\u5347\u70ba H200 \u4e0a\u7684 8\u3002<\/li>\n
    • NVIDIA H200 \u652f\u63f4\u5c0d 180B \u53c3\u6578\u6a21\u578b\u9032\u884c LoRA \u5fae\u8abf\uff0c\u800c H100 \u50c5\u80fd\u652f\u63f4\u5230 70B\u3002<\/li>\n
    • Mixtral 8x22B MoE \u6a21\u578b\u5728 2 \u9846 H200 \u4e0a\u5373\u53ef\u8f09\u5165\uff0c\u800c\u5728 H100 \u4e0a\u5247\u9700\u8981 5 \u9846\uff0ctoken \u541e\u5410\u91cf\u63d0\u5347 2.3 \u500d\u3002<\/li>\n
    • \u76f8\u8f03 H100 \u7684 80 GB\uff0c141 GB \u8a18\u61b6\u9ad4\u70ba 70B \u6a21\u578b\u63d0\u4f9b\u7d04 7 \u500d\u7684 KV Cache \u5bb9\u91cf\u3002<\/li>\n
    • 4.8 TB\/s \u983b\u5bec\u76f8\u8f03 H100 \u7684 3.35 TB\/s\uff0c\u53ef\u5be6\u73fe\u7d04 1.4 \u500d\u66f4\u5feb\u7684 token \u751f\u6210\u3002<\/li>\n
    • \u55ae\u9846 H200 \u5c31\u80fd\u57f7\u884c 130B \u6a21\u578b\uff0c\u7701\u53bb\u5169\u9846 H100 \u6240\u9700\u7684\u5f35\u91cf\u5e73\u884c\u958b\u92b7\u3002<\/li>\n<\/ul>\n

      \u5728\u8a13\u7df4\u8d85\u5927\u898f\u6a21\u57fa\u790e\u6a21\u578b\u6642\uff0c\u4f60\u53ef\u4ee5\u964d\u4f4e\u6240\u9700\u7684\u4f3a\u670d\u5668\u786c\u9ad4\u6578\u91cf\u3002\u5c07\u8907\u96dc\u7684 Mixture-of-Experts \u6a21\u578b\u57f7\u884c\u5728\u66f4\u5c11\u7684\u5be6\u9ad4\u7cfb\u7d71\u4e0a\uff0c\u80fd\u6e1b\u5c11\u7db2\u8def\u540c\u6b65\u74f6\u9838\u3002\u66f4\u5927\u7684\u8a18\u61b6\u9ad4\u5bb9\u91cf\u4e5f\u964d\u4f4e\u4e86\u5728\u53cd\u5411\u50b3\u64ad\u968e\u6bb5\u7bc0\u9ede\u9593\u50b3\u905e\u4e2d\u4ecb\u5f35\u91cf\u72c0\u614b\u7684\u8907\u96dc\u5ea6\u3002\u4f60\u7684\u904b\u7b97\u53e2\u96c6\u5728\u7bc0\u9ede\u9593\u8cc7\u6599\u50b3\u8f38\u4e0a\u82b1\u8cbb\u7684\u6642\u9593\u66f4\u5c11\u3002<\/p>\n

      \u6279\u6b21\u5927\u5c0f\u7684\u64f4\u5145\u70ba\u4f60\u7684\u6a5f\u5668\u5b78\u7fd2\u5de5\u4f5c\u6d41\u7a0b\u5e36\u4f86\u5de8\u5927\u7684\u7b97\u529b\u6548\u7387\u3002\u66f4\u9ad8\u7684 token \u541e\u5410\u91cf\u8b93 GPU \u6838\u5fc3\u5728\u9577\u6642\u9593\u8a13\u7df4\u671f\u9593\u59cb\u7d42\u63a5\u8fd1\u6eff\u8f09\u904b\u4f5c\u3002\u4f60\u53ef\u4ee5\u5728\u512a\u5316\u8cc7\u6599\u4e2d\u5fc3\u80fd\u8017\u7684\u540c\u6642\uff0c\u66f4\u5feb\u901f\u5730\u8a13\u7df4\u6700\u5148\u9032\u7684\u57fa\u790e\u6a21\u578b\u3002\u9019\u4e9b\u541e\u5410\u91cf\u7684\u6574\u9ad4\u63d0\u5347\uff0c\u8b93\u4f60\u7684\u786c\u9ad4\u57fa\u790e\u8a2d\u65bd\u6295\u8cc7\u5728\u9577\u671f\u5167\u7372\u5f97\u6700\u5927\u5316\u5831\u916c\u3002<\/p>\n

      \u90e8\u7f72\u901f\u5ea6\u8207\u4e0b\u4e00\u4ee3\u66ff\u4ee3\u65b9\u6848\u6bd4\u8f03<\/strong><\/h2>\n

      \u76f8\u5bb9\u98a8\u51b7 HGX \u6a5f\u67b6\u7684\u5373\u63d2\u5373\u7528\u5347\u7d1a<\/h3>\n

      \u4f60\u53ef\u4ee5\u904b\u7528\u73fe\u6709\u4f3a\u670d\u5668\u6a5f\u7bb1\uff0c\u5feb\u901f\u5347\u7d1a\u4f01\u696d\u904b\u7b97\u57fa\u790e\u8a2d\u65bd\u3002NVIDIA H200 \u53ef\u76f4\u63a5\u5b89\u88dd\u5728\u6a19\u6e96\u98a8\u51b7 20\u201330 kW \u6a5f\u67b6\u4e2d\uff0c\u7121\u9700\u5c0d\u6a5f\u623f\u9032\u884c\u5fb9\u5e95\u6539\u9020\u3002\u50cf Dell PowerEdge\u3001HPE ProLiant \u548c Aivres KR6288-X2 6U \u7b49\u4f01\u696d\u4f3a\u670d\u5668\u5e73\u53f0\uff0c\u90fd\u53ef\u7acb\u5373\u652f\u63f4\u9019\u4e9b\u8655\u7406\u5668\u3002\u5177\u5099 N+1 \u5197\u9918\u7684\u96d9\u5340\u98a8\u6247\uff0c\u53ef\u5728\u6eff\u8f09\u904b\u4f5c\u6642\u7dad\u6301\u7a69\u5b9a\u7684\u904b\u884c\u74b0\u5883\u3002<\/p>\n

      \u900f\u904e\u4fdd\u7559\u73fe\u6709\u7684\u4f9b\u96fb\u8207\u9001\u98a8\u8a2d\u8a08\uff0c\u4f60\u53ef\u4ee5\u907f\u514d\u6602\u8cb4\u7684\u8cc7\u6599\u4e2d\u5fc3\u91cd\u5efa\u6210\u672c\u3002\u9ad8\u5bc6\u5ea6\u4f3a\u670d\u5668\u90e8\u7f72\u8981\u6c42\u7cbe\u6e96\u7684\u6c23\u6d41\u7ba1\u7406\uff0c\u540c\u6642\u5c07\u9032\u98a8\u6eab\u5ea6\u7dad\u6301\u5728 18\u00b0C\u201325\u00b0C \u4e4b\u9593\u3002\u51b7\u71b1\u901a\u9053\u5c01\u9589\u7cfb\u7d71\u6709\u52a9\u65bc\u963b\u6b62\u71b1\u7a7a\u6c23\u5728\u76f8\u9130\u6a5f\u7bb1\u4e4b\u9593\u56de\u6d41\u3002\u5c07\u6a21\u578b\u5b8c\u6574\u88dd\u5165 141 GB HBM3e \u8a18\u61b6\u9ad4\u4e2d\uff0c\u53ef\u6d88\u9664\u8de8\u591a\u7bc0\u9ede\u53e2\u96c6\u9032\u884c\u6a21\u578b\u5207\u5206\u7684\u8907\u96dc\u5ea6\u3002\u5de5\u7a0b\u5718\u968a\u5728\u4fee\u6539\u7a0b\u5f0f\u78bc\u4e0a\u82b1\u8cbb\u7684\u6642\u9593\u66f4\u5c11\uff0c\u9032\u800c\u76f4\u63a5\u63d0\u5347\u8edf\u9ad4\u958b\u767c\u6548\u7387\u3002<\/p>\n

      \u6d88\u9664\u6db2\u51b7\u57fa\u790e\u8a2d\u65bd\u5e36\u4f86\u7684\u5ef6\u9072<\/h3>\n

      \u65b0\u4e00\u4ee3 GPU \u67b6\u69cb\u7684\u71b1\u8a2d\u8a08\u529f\u8017\u6975\u9ad8\uff0c\u5f80\u5f80\u5fc5\u9808\u4f9d\u8cf4\u6db2\u51b7\u8a2d\u65bd\u3002\u9077\u79fb\u5230\u9019\u4e9b\u5e73\u53f0\u6703\u8feb\u4f7f\u6a5f\u623f\u5718\u968a\u5b89\u88dd\u8907\u96dc\u7684\u7ba1\u7dda\u3001\u6db2\u51b7\u71b1\u4ea4\u63db\u5668\u53ca\u9ad8\u5bb9\u91cf\u4f9b\u96fb\u55ae\u5143\u3002\u9019\u4e9b\u5be6\u9ad4\u57fa\u790e\u8a2d\u65bd\u6539\u9020\u4e0d\u50c5\u65bd\u5de5\u671f\u9577\uff0c\u9084\u6703\u589e\u52a0\u71df\u904b\u98a8\u96aa\u3002<\/p>\n

      \u9078\u64c7\u98a8\u51b7\u4f3a\u670d\u5668\u5e73\u53f0\uff0c\u800c\u975e\u9ad8\u5ea6\u4f9d\u8cf4\u6db2\u51b7\u7684\u67b6\u69cb\uff0c\u53ef\u4ee5\u8b93\u4f60\u66f4\u5feb\u901f\u90e8\u7f72\u4f3a\u670d\u5668\u786c\u9ad4\u3002\u4e0b\u5217\u8868\u683c\u6bd4\u8f03\u4e86\u4e0d\u540c GPU \u4e16\u4ee3\u5728\u71b1\u8a2d\u8a08\u529f\u8017\u8207\u6563\u71b1\u9700\u6c42\u4e0a\u7684\u4e3b\u8981\u5dee\u7570\uff1a<\/p>\n

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

      GPU \u578b\u865f<\/th>\n	\u71b1\u8a2d\u8a08\u529f\u8017\uff08TDP\uff09<\/th>\n	\u6563\u71b1\u9700\u6c42<\/th>\n<\/tr>\n
      H200<\/td>\n	700W<\/td>\n	\u98a8\u51b7\u90e8\u7f72<\/td>\n<\/tr>\n
      B200<\/td>\n	1000W<\/td>\n	\u9700\u6db2\u51b7<\/td>\n<\/tr>\n
      B300<\/td>\n	1400W<\/td>\n	\u5fc5\u9808\u6db2\u51b7<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n \u9078\u64c7\u98a8\u51b7\u786c\u9ad4\uff0c\u53ef\u4ee5\u8b93\u4f60\u7acb\u5373\u5728\u6210\u719f\u7684\u8edf\u9ad4\u5806\u758a\u4e0a\u904b\u4f5c\uff0c\u800c\u4e0d\u6703\u4e2d\u65b7\u73fe\u6709\u5de5\u4f5c\u6d41\u7a0b\u3002\u4f60\u4e5f\u80fd\u907f\u514d\u5ba2\u88fd\u5316\u6db2\u51b7\u5143\u4ef6\u5728\u4f9b\u61c9\u93c8\u4e0a\u7684\u74f6\u9838\u3002\u8cc7\u6599\u79d1\u5b78\u5718\u968a\u53ef\u4ee5\u66f4\u5feb\u901f\u90e8\u7f72\u6700\u5148\u9032\u7684\u57fa\u790e\u6a21\u578b\uff0c\u52a0\u901f\u7d44\u7e54\u7684\u4e0a\u5e02\u6642\u7a0b\u3002\u66f4\u5feb\u7684\u7cfb\u7d71\u90e8\u7f72\u53ef\u7acb\u5373\u5e36\u4f86\u71df\u904b\u50f9\u503c\uff0c\u540c\u6642\u6709\u6548\u5b88\u4f4f\u6574\u9ad4\u64c1\u6709\u6210\u672c\u3002<\/p>\n \u4f01\u696d TCO \u8207\u80fd\u6e90\u6548\u7387\u6700\u4f73\u5316<\/strong><\/h2>\n \u7576\u4f60\u90e8\u7f72\u5728\u65e2\u5b9a\u6a5f\u623f\u96fb\u529b\u9650\u5236\u4e0b\uff0c\u4ecd\u80fd\u63d0\u4f9b\u66f4\u5f37\u7b97\u529b\u7684\u8655\u7406\u5668\u6642\uff0c\u5c31\u80fd\u6700\u5927\u5316\u8cc7\u6599\u4e2d\u5fc3\u6548\u7387\u3002\u73fe\u4ee3 AI \u52a0\u901f\u5668\u5728\u63a7\u5236\u9ad8\u80fd\u6e90\u9700\u6c42\u7684\u540c\u6642\uff0c\u63d0\u4f9b\u66f4\u9ad8\u7684\u904b\u7b97\u5bc6\u5ea6\u8207\u66f4\u512a\u7684\u6bcf\u74e6\u6548\u80fd\u3002<\/p>\n \u66f4\u9ad8\u7684\u904b\u7b97\u5bc6\u5ea6\u8207\u6548\u80fd\u529f\u8017\u6bd4<\/h3>\n H200 \u4e0d\u50c5\u6548\u80fd\u66f4\u5f37\uff0c\u800c\u4e14\u529f\u8017\u8207 H100 \u6301\u5e73\u3002\u5728 LLM \u4efb\u52d9\u4e0a\u5be6\u73fe 50% \u7684\u80fd\u6e90\u4f7f\u7528\u964d\u4f4e\uff0c\u540c\u6642\u8a18\u61b6\u9ad4\u983b\u5bec\u52a0\u500d\uff0c\u4f7f\u5176\u6574\u9ad4\u64c1\u6709\u6210\u672c\uff08TCO\uff09\u6700\u591a\u964d\u4f4e 50%\u3002<\/p>\n \u5728\u5927\u898f\u6a21\u8a13\u7df4\u9031\u671f\u4e2d\uff0c\u4f60\u53ef\u4ee5\u900f\u904e\u570d\u7e5e\u80fd\u6548\u6700\u4f73\u5316\u7684\u786c\u9ad4\u5275\u65b0\uff0c\u7372\u5f97\u53ef\u89c0\u7684\u71df\u904b\u6536\u76ca\u3002\u5148\u9032\u7cfb\u7d71\u67b6\u69cb\u5728\u53e2\u96c6\u5c64\u7d1a\u512a\u5316\u4e86\u96fb\u529b\u5206\u914d\uff1a<\/p>\n \n \u900f\u904e\u9ad8\u6548 HBM3e \u8a18\u61b6\u9ad4\u6280\u8853\u63d0\u5347\u6548\u80fd\u529f\u8017\u6bd4\u3002<\/li>\n \u81ea\u9069\u61c9\u5de5\u4f5c\u8ca0\u8f09\u6392\u7a0b\u53ef\u6700\u4f73\u5316 GPU \u5229\u7528\u7387\uff0c\u6e1b\u5c11\u4e0d\u5fc5\u8981\u7684\u8017\u96fb\u3002<\/li>\n \u8207\u6db2\u51b7\u4f3a\u670d\u5668\u6574\u5408\u5f8c\uff0c\u53ef\u964d\u4f4e\u8cc7\u6599\u4e2d\u5fc3\u7684 PUE\uff08\u96fb\u6e90\u4f7f\u7528\u6548\u7387\uff09\u3002<\/li>\n \u66f4\u77ed\u7684\u8a13\u7df4\u6642\u9593\u76f4\u63a5\u6e1b\u5c11\u6bcf\u6b21\u8a13\u7df4\u4f5c\u696d\u7684\u7e3d\u80fd\u8017\u3002<\/li>\n \u66f4\u9ad8\u7684\u53e2\u96c6\u5bc6\u5ea6\u63d0\u5347\u6574\u9ad4\u6548\u7387\uff0c\u964d\u4f4e\u6bcf\u500b\u5de5\u4f5c\u8ca0\u8f09\u7684\u7e3d\u80fd\u8017\u3002<\/li>\n<\/ul>\n \u9019\u4e9b\u6548\u7387\u63d0\u5347\u80fd\u5728\u4f60\u6bcf\u5929\u57f7\u884c\u91cd\u5ea6\u8ca0\u8f09\u6642\uff0c\u7acb\u5373\u5e36\u4f86\u8ca1\u52d9\u4e0a\u7684\u7de9\u89e3\u3002<\/p>\n \u7e2e\u77ed\u53e2\u96c6\u5be6\u969b\u904b\u8f49\u6642\u6578\u4ee5\u964d\u4f4e\u71df\u904b\u6210\u672c<\/h3>\n \u5728\u63a1\u7528 H200 \u4e4b\u5f8c\uff0c\u6709\u5831\u544a\u6307\u51fa\u5176\u96fb\u529b\u7bc0\u7701\u53ef\u9054 28%\uff0c\u4e14\u6bcf\u6a5f\u67b6\u541e\u5410\u91cf\u66f4\u9ad8\u3002<\/p>\n \u4f60\u53ef\u4ee5\u7528\u66f4\u5c11\u7684\u7e3d\u904b\u8f49\u6642\u9593\u5b8c\u6210\u5927\u578b\u6a21\u578b\u8a13\u7df4\u4f5c\u696d\uff0c\u5f9e\u800c\u76f4\u63a5\u964d\u4f4e\u8cc7\u6599\u4e2d\u5fc3\u96fb\u8cbb\u3002\u66f4\u77ed\u7684\u8a13\u7df4\u6642\u9593\u6e1b\u5c11\u4e86\u6240\u6709\u914d\u5957\u8a2d\u5099\uff08\u5982\u4f3a\u670d\u5668\u98a8\u6247\u8207\u6a5f\u623f\u51b7\u537b\u7cfb\u7d71\uff09\u7684\u6574\u9ad4\u80fd\u8017\u3002\u7531\u65bc\u66f4\u5feb\u7684\u786c\u9ad4\u8655\u7406\u80fd\u529b\u63d0\u5347\u4e86\u6bcf\u65e5\u7e3d\u8f38\u51fa\uff0c\u4f60\u7684\u786c\u9ad4\u6295\u8cc7\u4e5f\u80fd\u66f4\u5feb\u56de\u672c\u3002<\/p>\n \u5728\u8655\u7406\u5927\u898f\u6a21\u57fa\u790e\u6a21\u578b\u6642\uff0c\u4f60\u7684\u4f01\u696d\u53ef\u4ee5\u5728\u6bcf\u6708\u96fb\u8cbb\u4e0a\u7bc0\u7701\u5927\u91cf\u6210\u672c\u3002\u55ae\u7b46\u524d\u671f\u786c\u9ad4\u6295\u8cc7\uff0c\u5f80\u5f80\u53ef\u5728\u4e00\u5e74\u5167\u56de\u672c\uff1b\u5728\u9054\u5230\u56de\u672c\u9ede\u5f8c\uff0c\u4f60\u7684\u6301\u7e8c\u71df\u904b\u6210\u672c\u5c07\u986f\u8457\u4e0b\u964d\u3002\u66f4\u5feb\u7684\u6a21\u578b\u958b\u767c\u7bc0\u594f\u80fd\u52a0\u901f\u8edf\u9ad4\u767c\u4f48\uff0c\u540c\u6642\u7dad\u6301\u4f01\u696d\u71df\u904b\u9810\u7b97\u3002<\/p>\n \n <\/div>\n<\/div>\n NVIDIA H200 \u5fb9\u5e95\u91cd\u5851\u4e86\u4f60\u7684\u4eba\u5de5\u667a\u6167\u6a21\u578b\u8a13\u7df4\u80fd\u529b\u3002\u5176 141GB HBM3e \u5bb9\u91cf\u8207 4.8TB\/s \u8a18\u61b6\u9ad4\u983b\u5bec\uff0c\u76f4\u63a5\u6d88\u9664\u4e86 GPU \u8a08\u7b97\u9592\u7f6e\u6642\u9593\u3002\u4f60\u7684 Tensor Core \u80fd\u5728\u7121\u9700\u7b49\u5f85\u8cc7\u6599\u50b3\u8f38\u7684\u60c5\u6cc1\u4e0b\uff0c\u6301\u7e8c\u8655\u7406\u6d77\u91cf\u8cc7\u6599\u96c6\u3002<\/p>\n \u900f\u904e\u8207\u73fe\u6709\u4f3a\u670d\u5668\u7684\u5373\u63d2\u5373\u7528\u76f8\u5bb9\u6027\uff0c\u4f60\u5728 2026 \u5e74\u7372\u5f97\u4e86\u986f\u8457\u7684\u4e0a\u5e02\u6642\u9593\u512a\u52e2\u3002\u4f60\u7684\u5718\u968a\u53ef\u4ee5\u5c07\u9ad8\u5bc6\u5ea6\u904b\u7b97\u7bc0\u9ede\u76f4\u63a5\u90e8\u7f72\u5230\u6a19\u6e96\u98a8\u51b7\u6a5f\u67b6\u4e2d\uff0c\u7121\u9700\u6602\u8cb4\u7684\u6a5f\u623f\u6539\u9020\u6216\u6db2\u51b7\u5de5\u7a0b\u5ef6\u9072\u3002<\/p>\n \u5c0d\u65bc\u4f01\u696d\u7d1a AI \u5de5\u4f5c\u8ca0\u8f09\u800c\u8a00\uff0cNVIDIA H200 \u5728\u901f\u5ea6\u3001\u80fd\u6e90\u6548\u7387\u4ee5\u53ca\u6295\u8cc7\u56de\u6536\u901f\u5ea6\u65b9\u9762\uff0c\u63d0\u4f9b\u4e86\u6975\u5177\u7af6\u722d\u529b\u7684\u6574\u9ad4\u8868\u73fe\u3002<\/p>\n \u5e38\u898b\u554f\u984c\uff08FAQ\uff09<\/strong><\/h2>\n NVIDIA H200 \u5728 AI \u8a13\u7df4\u65b9\u9762\u6bd4 H100 \u5feb\u591a\u5c11\uff1f<\/h3>\n \u5c0d\u65bc\u50cf Llama 2 \u9019\u6a23\u7684\u5927\u578b\u6a21\u578b\uff0cNVIDIA H200 \u53ef\u63d0\u4f9b\u6700\u9ad8 2 \u500d\u7684\u541e\u5410\u91cf\u63d0\u5347\u3002MLPerf v4.0 \u57fa\u6e96\u6e2c\u8a66\u986f\u793a\uff0c\u5728\u55ae\u7bc0\u9ede\u4e0a\u9032\u884c Llama 2 70B \u7684 LoRA \u5fae\u8abf\u6642\uff0cH200 \u76f8\u8f03 H100 \u63d0\u5347 14%\uff0c\u5c07\u8a13\u7df4\u6642\u9593\u7e2e\u77ed\u81f3 24.7 \u5206\u9418\u3002<\/p>\n \u90e8\u7f72 NVIDIA H200 \u662f\u5426\u9700\u8981\u6db2\u51b7\uff1f<\/h3>\n \u4e0d\u9700\u8981\uff0c\u4f60\u7121\u9700\u70ba\u5176\u5efa\u7f6e\u6db2\u51b7\u57fa\u790e\u8a2d\u65bd\u3002NVIDIA H200 \u7684\u71b1\u8a2d\u8a08\u529f\u8017\u70ba 700W\uff0c\u53ef\u76f4\u63a5\u90e8\u7f72\u5230\u73fe\u6709\u7684\u4f01\u696d\u98a8\u51b7\u4f3a\u670d\u5668\u6a5f\u67b6\u4e2d\uff0c\u7121\u9700\u6602\u8cb4\u7684\u6a5f\u623f\u5347\u7d1a\u6216\u65bd\u5de5\u5ef6\u9072\u3002<\/p>\n \u70ba\u4ec0\u9ebc HBM3e \u8a18\u61b6\u9ad4\u53ef\u4ee5\u63d0\u5347 LLM \u8a13\u7df4\u901f\u5ea6\uff1f<\/h3>\n H200 \u63d0\u4f9b 141GB \u7684 HBM3e \u8a18\u61b6\u9ad4\u8207 4.8TB\/s \u7684\u8a18\u61b6\u9ad4\u983b\u5bec\uff0c\u5f62\u6210\u4e00\u689d\u300c\u8d85\u5bec\u8cc7\u6599\u7ba1\u7dda\u300d\uff0c\u53ef\u4ee5\u6301\u7e8c\u70ba Tensor Core \u4f9b\u61c9\u8cc7\u6599\u3002\u9ad8\u983b\u5bec\u53ef\u4ee5\u9632\u6b62\u904b\u7b97\u5f15\u64ce\u5728\u5927\u898f\u6a21\u6578\u5b78\u904b\u7b97\u4e2d\u983b\u7e41\u66ab\u505c\uff0c\u8b93\u8655\u7406\u5668\u5728\u5927\u6279\u6b21\u8a13\u7df4\u4efb\u52d9\u4e2d\u59cb\u7d42\u4fdd\u6301\u6d3b\u8e8d\u3002<\/p>\n NVIDIA H200 \u662f\u5426\u53ef\u4ee5\u964d\u4f4e\u4f01\u696d\u71df\u904b\u6210\u672c\uff1f<\/h3>\n \u53ef\u4ee5\uff0c\u4f60\u53ef\u4ee5\u5c07\u6574\u9ad4\u64c1\u6709\u6210\u672c\u6700\u9ad8\u964d\u4f4e 50%\u3002H200 \u5728\u7dad\u6301\u76f8\u540c\u6027\u80fd\u7684\u524d\u63d0\u4e0b\uff0c\u53ef\u7bc0\u7701\u7d04 28% \u7684\u8017\u96fb\u91cf\uff0c\u4e26\u7e2e\u77ed\u53e2\u96c6\u7684\u5be6\u969b\u8a13\u7df4\u6642\u9593\u3002\u66f4\u5feb\u7684\u8a13\u7df4\u4f5c\u696d\u5728\u4e0d\u63d0\u9ad8\u6a5f\u623f\u96fb\u529b\u4e0a\u9650\u7684\u60c5\u6cc1\u4e0b\uff0c\u986f\u8457\u964d\u4f4e\u6bcf\u6708\u80fd\u6e90\u6d88\u8017\u3002<\/p>\n","protected":false},"excerpt":{"rendered":" \u900f\u904e NVIDIA H200 GPU\uff0c\u4f60\u53ef\u4ee5\u5927\u5e45\u52a0\u901f\u4f01\u696d\u7d1a AI \u6a21\u578b\u8a13\u7df4\u3002\u9019\u6b3e\u8655\u7406\u5668\u900f\u904e\u786c\u9ad4\u5347\u7d1a\uff0c\u6d88\u9664\u56b4\u91cd\u7684 […]<\/p>\n Read More…<\/a><\/p>\n","protected":false},"author":2,"featured_media":34201,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[61],"tags":[12926,12864,4276,3861,3317],"class_list":["post-34208","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-hong-kong-dedicated-server-tc","tag-ai-training-speed","tag-hbm3e-memory","tag-nvidia-h200-tc","tag-data-center-efficiency-tc","tag-gpu-performance-tc"],"acf":[],"yoast_head":"\n