\u67b6\u6784\u6838\u5fc3\uff1aNVIDIA H200 \u7684 HBM3e \u5185\u5b58\u5347\u7ea7<\/strong><\/h2>\n

\u8981\u5728\u8bad\u7ec3\u8d85\u5927\u89c4\u6a21\u4eba\u5de5\u667a\u80fd\u6a21\u578b\u65f6\u907f\u514d\u607c\u4eba\u7684\u7b97\u529b\u74f6\u9888\uff0c\u4f60\u9700\u8981\u5f3a\u5927\u7684\u786c\u4ef6\u7ec4\u4ef6\u3002NVIDIA H200 \u901a\u8fc7\u5f15\u5165 141GB \u7684\u5148\u8fdb HBM3e \u5185\u5b58\uff0c\u89e3\u51b3\u4e86\u4f20\u7edf\u7684\u5185\u5b58\u9650\u5236\u3002\u4e0e H100 SXM \u5904\u7406\u5668\u76f8\u6bd4\uff0c\u8fd9\u4e00\u5347\u7ea7\u5e26\u6765 76% \u7684\u5bb9\u91cf\u63d0\u5347\u548c 43% \u7684\u5e26\u5bbd\u63d0\u5347\u3002<\/p>\n

4.8TB\/s \u5185\u5b58\u5e26\u5bbd\u4e0e Tensor Core \u5229\u7528\u7387<\/h3>\n

H200 \u63d0\u4f9b 141 GB HBM3e \u5185\u5b58\u548c 4.8 TB\/s \u5e26\u5bbd\uff0c\u662f H100 \u5185\u5b58\u5bb9\u91cf\u7684 1.44 \u500d\u3001\u5e26\u5bbd\u7684 1.8 \u500d\u3002<\/p>\n

\u8fd9\u6761\u8d85\u5bbd\u5185\u5b58\u901a\u9053\u5f7b\u5e95\u6539\u53d8\u4e86\u7cfb\u7edf\u5904\u7406\u590d\u6742\u6570\u5b66\u6570\u636e\u7684\u65b9\u5f0f\u3002\u5f53\u4f60\u8bad\u7ec3\u8d85\u5927\u89c4\u6a21\u4eba\u5de5\u667a\u80fd\u6a21\u578b\u65f6\uff0cTensor Core \u5f80\u5f80\u4f1a\u56e0\u7b49\u5f85\u6570\u636e\u800c\u6682\u505c\u3002\u9ad8\u5e26\u5bbd\u901a\u8fc7\u4ee5\u5feb\u4e8e\u6570\u5b66\u8ba1\u7b97\u5f15\u64ce\u7684\u901f\u5ea6\u4f20\u8f93\u6570\u636e\uff0c\u6d88\u9664\u4e86\u8fd9\u4e9b\u505c\u987f\u3002<\/p>\n

\u5728\u81ea\u56de\u5f52\u6587\u672c\u751f\u6210\u7684\u89e3\u7801\u9636\u6bb5\uff0c4.8 TB\/s \u7684\u5185\u5b58\u5e26\u5bbd\u6709\u52a9\u4e8e\u9632\u6b62 GPU \u5728\u5185\u5b58 I\/O \u64cd\u4f5c\u4e0b\u53d1\u751f\u505c\u987f\u3002\u8fd9\u76f4\u63a5\u4fdd\u8bc1 Tensor Core \u4fdd\u6301\u6d3b\u8dc3\uff0c\u5c24\u5176\u662f\u5728\u5927\u6279\u91cf\u8bad\u7ec3\u573a\u666f\u4e2d\uff0c\u539f\u672c\u786c\u4ef6\u53ef\u80fd\u4f1a\u5f00\u59cb\u963b\u585e\u30024.8 TB\/s \u7684\u5185\u5b58\u5e26\u5bbd\u7f29\u77ed\u4e86\u6570\u636e\u5bc6\u96c6\u578b\u4efb\u52a1\u6240\u9700\u65f6\u95f4\uff0c\u8ba9\u6a21\u578b\u8bad\u7ec3\u66f4\u5feb\uff0c\u5e76\u63d0\u5347 Tensor Core \u5728\u590d\u6742\u8ba1\u7b97\u4e2d\u7684\u5229\u7528\u6548\u7387\u3002<\/p>\n

\u5728\u90e8\u7f72\u8fd9\u5957\u5185\u5b58\u7cfb\u7edf\u65f6\uff0c\u4f60\u53ef\u4ee5\u770b\u5230\u4e00\u7cfb\u5217\u5b9e\u9645\u4f18\u52bf\uff1a<\/p>\n

\n
• \u5f53\u6279\u5927\u5c0f\u589e\u5927\u5230\u7531\u5185\u5b58\u4f20\u8f93\u4e3b\u5bfc\u5904\u7406\u65f6\u95f4\u65f6\uff0c\u786c\u4ef6\u8fd0\u884c\u901f\u5ea6\u660e\u663e\u63d0\u9ad8\u3002<\/li>\n
• \u5f53\u5185\u5b58\u5bb9\u91cf\u6216\u5e26\u5bbd\u6210\u4e3a\u4e3b\u8981\u7cfb\u7edf\u74f6\u9888\u65f6\uff0c\u6027\u80fd\u4f18\u52bf\u5c24\u4e3a\u663e\u8457\u3002<\/li>\n
• \u5355\u9897 NVIDIA H200 \u5c31\u80fd\u627f\u62c5\u539f\u672c\u9700\u8981\u4e24\u9897 H100 \u624d\u80fd\u627f\u8f7d\u7684\u7e41\u91cd\u5de5\u4f5c\u8d1f\u8f7d\uff0c\u964d\u4f4e\u591a GPU \u90e8\u7f72\u590d\u6742\u5ea6\u3002<\/li>\n
• \u5bf9\u4e8e\u53ef\u653e\u5165 80 GB \u5185\u5b58\u7684\u5c0f\u6a21\u578b\uff0c\u7531\u4e8e\u74f6\u9888\u4e0d\u5728\u5185\u5b58\uff0c\u6027\u80fd\u57fa\u672c\u76f8\u5f53\u3002<\/li>\n<\/ul>\n

  \u4f01\u4e1a\u53cd\u9988\u663e\u793a\uff0c\u76f8\u6bd4 H100 \u914d\u7f6e\uff0cH200 \u96c6\u7fa4\u53ef\u5c06\u591a\u5341\u4ebf\u53c2\u6570\u6a21\u578b\u7684\u5fae\u8c03\u65f6\u95f4\u7f29\u77ed\u591a\u8fbe 40%\u3002<\/p>\n

  \u5f53\u6a21\u578b\u5df2\u7ecf\u65e0\u6cd5\u5728 H100 \u4e0a\u201c\u8212\u9002\u201d\u5bb9\u7eb3\u65f6\uff0cH200 \u7684\u4f18\u52bf\u5c24\u4e3a\u660e\u663e\u3002\u5bf9\u4e8e\u5927\u6279\u91cf\u8bad\u7ec3\u3001\u8d85\u957f\u4e0a\u4e0b\u6587\u7a97\u53e3\u6216\u9ad8\u5ea6\u5403\u5185\u5b58\u7684\u5de5\u4f5c\u8d1f\u8f7d\uff0cH200 \u80fd\u5e26\u6765\u66f4\u5e73\u6ed1\u3001\u66f4\u5feb\u901f\u7684\u4f53\u9a8c\u3002<\/p>\n

  \u964d\u4f4e\u591a GPU \u96c6\u7fa4\u4e2d\u7684\u901a\u4fe1\u5ef6\u8fdf<\/h3>\n

  \u591a\u8282\u70b9\u8bad\u7ec3\u9700\u8981\u5728\u4e0d\u540c\u670d\u52a1\u5668\u8282\u70b9\u4e4b\u95f4\u6301\u7eed\u5171\u4eab\u6570\u636e\u3002\u5f53\u4f60\u8bad\u7ec3\u66f4\u5927\u89c4\u6a21\u7684\u6a21\u578b\u65f6\uff0c\u7f51\u7edc\u540c\u6b65\u5f00\u9500\u5e38\u5e38\u4f1a\u5728 All-Reduce \u7b49\u96c6\u5408\u901a\u4fe1\u64cd\u4f5c\u4e2d\u5f62\u6210\u4e25\u5cfb\u7684\u6027\u80fd\u74f6\u9888\u3002<\/p>\n

  \u73b0\u4ee3\u96c6\u7fa4\u7f51\u7edc\u67b6\u6784\u53ef\u5728\u5206\u5e03\u5f0f\u57fa\u7840\u8bbe\u65bd\u4e2d\u6d88\u9664\u8fd9\u4e9b\u5ef6\u8fdf\uff1a<\/p>\n

  \n
  • \u57fa\u4e8e RDMA over Converged Ethernet \u7684\u7f51\u7edc\u73af\u5883\uff0c\u53ef\u4ee5\u5728 82 \u5206\u949f\u5185\u5b8c\u6210\u5927\u8bed\u8a00\u6a21\u578b\u5fae\u8c03\uff0c\u4e3a Llama 3.1 70B \u7b49\u5927\u578b\u6a21\u578b\u5e26\u6765\u663e\u8457\u52a0\u901f\u3002<\/li>\n
  • \u6269\u5c55\u6548\u7387\u5927\u5e45\u63d0\u5347\uff0c\u5c24\u5176\u662f\u5728 70B \u7ea7\u522b\u7684\u5927\u6a21\u578b\u4e2d\uff0c\u82e5\u7f3a\u5c11\u4f18\u5316\u7684\u7f51\u7edc\u534f\u8bae\uff0c\u8bad\u7ec3\u65f6\u95f4\u4f1a\u6025\u5267\u62c9\u957f\u3002<\/li>\n
  • \u66f4\u4f4e\u5ef6\u8fdf\u7684\u8bbe\u8ba1\u4f9d\u8d56\u4ee5\u592a\u7f51\u4e0a\u7684\u76f4\u63a5\u5185\u5b58\u8bbf\u95ee\uff0c\u6d88\u9664\u4e0d\u5fc5\u8981\u7684 CPU \u5904\u7406\uff0c\u964d\u4f4e\u8282\u70b9\u95f4\u7aef\u5230\u7aef\u901a\u4fe1\u5ef6\u8fdf\u3002<\/li>\n
  • \u57fa\u7840\u8bbe\u65bd\u6210\u672c\u8282\u7ea6\u6301\u7eed\u7d2f\u79ef\uff0c\u56e0\u4e3a\u66f4\u5feb\u7684\u5fae\u8c03\u4efb\u52a1\u9700\u8981\u66f4\u5c11\u7684\u8ba1\u7b97\u5c0f\u65f6\u6570\uff0c\u4ece\u800c\u76f4\u63a5\u964d\u4f4e\u8fd0\u8425\u652f\u51fa\u3002<\/li>\n<\/ul>\n

    \u901a\u8fc7\u79fb\u9664 CPU \u5904\u7406\u5ef6\u8fdf\uff0c\u4f60\u7684\u591a\u8282\u70b9\u96c6\u7fa4\u53ef\u4ee5\u7ef4\u6301\u9ad8\u6570\u636e\u541e\u5410\u3002\u5f3a\u5927\u7684\u5185\u5b58\u5e26\u5bbd\u4e0e\u964d\u4f4e\u7684\u7f51\u7edc\u5ef6\u8fdf\u7ed3\u5408\u5728\u4e00\u8d77\uff0c\u8ba9\u6240\u6709\u8fde\u63a5\u7684\u5904\u7406\u5668\u5728\u957f\u65f6\u95f4\u8bad\u7ec3\u4e2d\u59cb\u7ec8\u4fdd\u6301\u9ad8\u8d1f\u8f7d\u8fd0\u884c\u3002<\/p>\n

    \u8bad\u7ec3\u57fa\u51c6\u4e0e\u57fa\u7840\u6a21\u578b\u541e\u5410\u91cf<\/strong><\/h2>\n

    \u4f60\u53ef\u4ee5\u901a\u8fc7\u5b9e\u8bc1\u57fa\u51c6\u6d4b\u8bd5\u6765\u8bc4\u4f30\u5904\u7406\u5668\u7684\u771f\u5b9e\u901f\u5ea6\u3002\u6807\u51c6\u5316\u7684\u57fa\u51c6\u6d4b\u8bd5\u5957\u4ef6\uff0c\u7528\u4e8e\u8861\u91cf\u590d\u6742\u786c\u4ef6\u7cfb\u7edf\u5728\u5b9e\u9645\u8fd0\u884c\u4e2d\u7684\u6027\u80fd\u8868\u73b0\u3002\u72ec\u7acb\u7684\u57fa\u51c6\u8bb0\u5f55\u6e05\u6670\u5c55\u793a\u4e86\u5927\u578b\u57fa\u7840\u6a21\u578b\u8d1f\u8f7d\u4e0b\u7684\u5904\u7406\u80fd\u529b\u63d0\u5347\u3002<\/p>\n

    MLPerf v4.0 \u6307\u6807\u4e0e\u8bad\u7ec3\u52a0\u901f<\/h3>\n

    \u6807\u51c6\u5316\u57fa\u51c6\u6d4b\u8bd5\u5728\u9ad8\u5f3a\u5ea6\u4f01\u4e1a\u8bad\u7ec3\u4efb\u52a1\u4e2d\u5c55\u793a\u4e86\u660e\u663e\u7684\u6027\u80fd\u589e\u76ca\u3002\u72ec\u7acb\u884c\u4e1a\u7ec4\u7ec7\u901a\u8fc7\u4e25\u683c\u7684\u6d4b\u8bd5\u6d41\u7a0b\uff0c\u9a8c\u8bc1\u786c\u4ef6\u5728\u771f\u5b9e\u8d1f\u8f7d\u4e0b\u7684\u8868\u73b0\u3002MLPerf Training v4.0 \u6d4b\u8bd5\u5728\u4e25\u683c\u8861\u91cf\u6807\u51c6\u4e0b\u9a8c\u8bc1\u4e86\u771f\u5b9e\u7684\u8bad\u7ec3\u52a0\u901f\u6570\u636e\u3002<\/p>\n

    \u5728 MLPerf Training v4.0 \u4e2d\uff0cH200 \u5728\u5355\u4e2a 8 GPU \u8282\u70b9\u4e0a\uff0c\u4ee5 24.7 \u5206\u949f\u5b8c\u6210 Llama 2 70B LoRA \u5fae\u8c03\uff0c\u76f8\u6bd4 H100 \u63d0\u5347 14%\u3002<\/p>\n

    \u4f60\u65e0\u9700\u66f4\u6539\u73b0\u6709\u8bad\u7ec3\u4ee3\u7801\u6216\u57fa\u7840\u8bbe\u65bd\u8bbe\u7f6e\uff0c\u5373\u53ef\u83b7\u5f97\u66f4\u5feb\u7684\u8bad\u7ec3\u5b8c\u6210\u65f6\u95f4\u3002\u5185\u5b58\u5bb9\u91cf\u7684\u6269\u5c55\uff0c\u51cf\u5c11\u4e86\u5728\u5927\u89c4\u6a21\u6570\u5b66\u6570\u7ec4\u8ba1\u7b97\u8fc7\u7a0b\u4e2d\u7684\u5904\u7406\u505c\u987f\u3002\u8ba1\u7b97\u5f15\u64ce\u53ef\u4ee5\u66f4\u5feb\u901f\u5730\u5904\u7406\u590d\u6742\u6a21\u578b\u6743\u91cd\uff0c\u540c\u65f6\u4fdd\u6301\u7cbe\u786e\u7684\u6d6e\u70b9\u8ba1\u7b97\u7ed3\u679c\u3002\u66f4\u9ad8\u7684\u8bad\u7ec3\u541e\u5410\u91cf\u76f4\u63a5\u51cf\u5c11\u4e86\u6574\u4e2a\u8ba1\u7b97\u8d44\u6e90\u6c60\u4e2d\u7684\u7a7a\u95f2\u65f6\u95f4\u3002<\/p>\n

    \u52a0\u901f\u6267\u884c\u901f\u5ea6\u5c06\u5f7b\u5e95\u6539\u53d8\u4f60\u7684\u8f6f\u4ef6\u5f00\u53d1\u8282\u594f\u3002\u5de5\u7a0b\u56e2\u961f\u53ef\u4ee5\u5728\u76f8\u540c\u7684\u5de5\u4f5c\u5468\u671f\u5185\u8fd0\u884c\u66f4\u591a\u8bad\u7ec3\u8fed\u4ee3\u3002\u66f4\u77ed\u7684\u8bad\u7ec3\u5468\u671f\u8ba9\u6570\u636e\u79d1\u5b66\u5bb6\u80fd\u66f4\u5feb\u6d4b\u8bd5\u65b0\u7684\u8d85\u53c2\u6570\u914d\u7f6e\u3002\u4f60\u7684\u7ec4\u7ec7\u53ef\u4ee5\u5728\u66f4\u5c11\u7684\u5929\u6570\u5185\uff0c\u5c06\u57fa\u7840\u6a21\u578b\u4ece\u5b9e\u9a8c\u9636\u6bb5\u63a8\u8fdb\u5230\u751f\u4ea7\u90e8\u7f72\u3002\u8bad\u7ec3\u65f6\u95f4\u7684\u7f29\u77ed\u52a0\u901f\u4e86\u4ea7\u54c1\u53d1\u5e03\uff0c\u4e3a\u4f60\u7684\u4e1a\u52a1\u5e26\u6765\u7ade\u4e89\u4f18\u52bf\u3002<\/p>\n

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

    \u5927\u578b\u8bed\u8a00\u6a21\u578b\u5728\u9884\u8bad\u7ec3\u548c\u5fae\u8c03\u65f6\uff0c\u9700\u8981\u6781\u9ad8\u7684\u5185\u5b58\u5e26\u5bbd\uff0c\u4ee5\u6301\u7eed\u4e3a\u6570\u5b66\u5f15\u64ce\u8f93\u9001\u77e9\u9635\u6570\u636e\uff0c\u4ece\u800c\u907f\u514d\u7b97\u529b\u7a7a\u8f6c\u3002\u516b\u5361 HGX H200 \u7cfb\u7edf\u5728\u5fae\u8c03 Llama 2 70B \u65f6\uff0c\u541e\u5410\u91cf\u53ef\u8d85\u8fc7 15,000 tokens\/\u79d2\u3002\u4e0e\u4f7f\u7528\u540c\u4e00 NeMo \u6846\u67b6\u7684 A100 GPU \u76f8\u6bd4\uff0c\u8be5\u7cfb\u7edf\u63d0\u4f9b\u6700\u9ad8 4.2 \u500d\u7684\u541e\u5410\u63d0\u5347\u3002<\/p>\n

    \u6d77\u91cf\u53c2\u6570\u89c4\u6a21\u8ba9\u6df1\u5ea6\u5b66\u4e60\u4efb\u52a1\u5bf9\u786c\u4ef6\u5185\u5b58\u4ea7\u751f\u5de8\u5927\u538b\u529b\u3002\u6269\u5c55\u5185\u5b58\u5bb9\u91cf\uff0c\u53ef\u4ee5\u5728\u4e0d\u540c\u53c2\u6570\u89c4\u6a21\u4e0b\u89e3\u9501\u66f4\u5927\u7684\u6279\u5927\u5c0f\uff1a<\/p>\n

    \n
    • \u5728\u5168\u7cbe\u5ea6\u6a21\u5f0f\u4e0b\uff0cLlama-70B \u8bad\u7ec3\u7684\u6bcf GPU \u6279\u5927\u5c0f\uff0c\u4ece H100 \u4e0a\u7684 4 \u63d0\u5347\u4e3a H200 \u4e0a\u7684 8\u3002<\/li>\n
    • NVIDIA H200 \u652f\u6301\u5bf9 180B \u53c2\u6570\u6a21\u578b\u8fdb\u884c LoRA \u5fae\u8c03\uff0c\u800c H100 \u4ec5\u80fd\u652f\u6301\u5230 70B\u3002<\/li>\n
    • Mixtral 8x22B MoE \u6a21\u578b\u5728 2 \u9897 H200 \u4e0a\u5373\u53ef\u52a0\u8f7d\uff0c\u800c\u5728 H100 \u4e0a\u5219\u9700\u8981 5 \u9897\uff0ctoken \u541e\u5410\u91cf\u63d0\u5347 2.3 \u500d\u3002<\/li>\n
    • \u76f8\u8f83 H100 \u7684 80 GB\uff0c141 GB \u5185\u5b58\u4e3a 70B \u6a21\u578b\u63d0\u4f9b\u7ea6 7 \u500d\u7684 KV Cache \u5bb9\u91cf\u3002<\/li>\n
    • 4.8 TB\/s \u5e26\u5bbd\u76f8\u8f83 H100 \u7684 3.35 TB\/s\uff0c\u53ef\u5b9e\u73b0\u7ea6 1.4 \u500d\u66f4\u5feb\u7684 token \u751f\u6210\u3002<\/li>\n
    • \u5355\u9897 H200 \u5c31\u80fd\u8fd0\u884c 130B \u6a21\u578b\uff0c\u7701\u53bb\u4e24\u9897 H100 \u6240\u9700\u7684\u5f20\u91cf\u5e76\u884c\u5f00\u9500\u3002<\/li>\n<\/ul>\n

      \u5728\u8bad\u7ec3\u8d85\u5927\u89c4\u6a21\u57fa\u7840\u6a21\u578b\u65f6\uff0c\u4f60\u53ef\u4ee5\u964d\u4f4e\u6240\u9700\u7684\u670d\u52a1\u5668\u786c\u4ef6\u6570\u91cf\u3002\u5c06\u590d\u6742\u7684 Mixture-of-Experts \u6a21\u578b\u8fd0\u884c\u5728\u66f4\u5c11\u7684\u7269\u7406\u7cfb\u7edf\u4e0a\uff0c\u80fd\u51cf\u5c11\u7f51\u7edc\u540c\u6b65\u74f6\u9888\u3002\u66f4\u5927\u7684\u5185\u5b58\u5bb9\u91cf\u4e5f\u51cf\u5c11\u4e86\u5728\u53cd\u5411\u4f20\u64ad\u9636\u6bb5\u8282\u70b9\u95f4\u4f20\u9012\u4e2d\u95f4\u5f20\u91cf\u72b6\u6001\u7684\u590d\u6742\u5ea6\u3002\u4f60\u7684\u8ba1\u7b97\u96c6\u7fa4\u5728\u8282\u70b9\u95f4\u6570\u636e\u4f20\u8f93\u4e0a\u82b1\u8d39\u7684\u65f6\u95f4\u66f4\u5c11\u3002<\/p>\n

      \u6279\u5927\u5c0f\u7684\u6269\u5c55\u4e3a\u4f60\u7684\u673a\u5668\u5b66\u4e60\u5de5\u4f5c\u6d41\u5e26\u6765\u5de8\u5927\u7684\u7b97\u529b\u6548\u7387\u3002\u66f4\u9ad8\u7684 token \u541e\u5410\u91cf\u8ba9 GPU \u6838\u5fc3\u5728\u957f\u65f6\u95f4\u8bad\u7ec3\u671f\u95f4\u59cb\u7ec8\u63a5\u8fd1\u6ee1\u8d1f\u8f7d\u8fd0\u884c\u3002\u4f60\u53ef\u4ee5\u5728\u4f18\u5316\u6570\u636e\u4e2d\u5fc3\u80fd\u8017\u7684\u540c\u65f6\uff0c\u66f4\u5feb\u901f\u5730\u8bad\u7ec3\u6700\u524d\u6cbf\u7684\u57fa\u7840\u6a21\u578b\u3002\u8fd9\u4e9b\u541e\u5410\u91cf\u7684\u6574\u4f53\u63d0\u5347\uff0c\u8ba9\u4f60\u7684\u786c\u4ef6\u57fa\u7840\u8bbe\u65bd\u6295\u8d44\u5728\u957f\u671f\u5185\u83b7\u5f97\u6700\u5927\u5316\u56de\u62a5\u3002<\/p>\n

      \u90e8\u7f72\u901f\u5ea6\u4e0e\u4e0b\u4e00\u4ee3\u66ff\u4ee3\u65b9\u6848\u5bf9\u6bd4<\/strong><\/h2>\n

      \u517c\u5bb9\u98ce\u51b7 HGX \u673a\u67b6\u7684\u5373\u63d2\u5373\u7528\u5347\u7ea7<\/h3>\n

      \u4f60\u53ef\u4ee5\u5229\u7528\u73b0\u6709\u670d\u52a1\u5668\u673a\u7bb1\uff0c\u5feb\u901f\u5347\u7ea7\u4f01\u4e1a\u7b97\u529b\u57fa\u7840\u8bbe\u65bd\u3002NVIDIA H200 \u53ef\u76f4\u63a5\u5b89\u88c5\u5728\u6807\u51c6\u7684\u98ce\u51b7 20\u201330 kW \u673a\u67b6\u4e2d\uff0c\u65e0\u9700\u5bf9\u673a\u623f\u8fdb\u884c\u5f7b\u5e95\u6539\u9020\u3002\u8bf8\u5982 Dell PowerEdge\u3001HPE ProLiant \u548c Aivres KR6288-X2 6U \u7b49\u4f01\u4e1a\u670d\u52a1\u5668\u5e73\u53f0\u90fd\u53ef\u7acb\u5373\u652f\u6301\u8fd9\u4e9b\u5904\u7406\u5668\u3002\u5177\u5907 N+1 \u5197\u4f59\u7684\u53cc\u533a\u98ce\u6247\uff0c\u53ef\u5728\u6ee1\u8d1f\u8f7d\u8fd0\u884c\u65f6\u4fdd\u6301\u7a33\u5b9a\u7684\u5de5\u4f5c\u73af\u5883\u3002<\/p>\n

      \u901a\u8fc7\u4fdd\u7559\u73b0\u6709\u7684\u4f9b\u7535\u4e0e\u9001\u98ce\u8bbe\u8ba1\uff0c\u4f60\u53ef\u4ee5\u907f\u514d\u6602\u8d35\u7684\u6570\u636e\u4e2d\u5fc3\u91cd\u6784\u6210\u672c\u3002\u9ad8\u5bc6\u5ea6\u670d\u52a1\u5668\u90e8\u7f72\u8981\u6c42\u7cbe\u786e\u7684\u6c14\u6d41\u7ba1\u7406\uff0c\u540c\u65f6\u5c06\u8fdb\u98ce\u6e29\u5ea6\u4fdd\u6301\u5728 18\u00b0C\u201325\u00b0C \u4e4b\u95f4\u3002\u51b7\u70ed\u901a\u9053\u5c01\u95ed\u7cfb\u7edf\u6709\u52a9\u4e8e\u963b\u6b62\u70ed\u7a7a\u6c14\u5728\u76f8\u90bb\u673a\u7bb1\u4e4b\u95f4\u56de\u6d41\u3002\u5c06\u6a21\u578b\u5b8c\u6574\u88c5\u5165 141 GB HBM3e \u5185\u5b58\u4e2d\uff0c\u53ef\u6d88\u9664\u8de8\u591a\u8282\u70b9\u96c6\u7fa4\u8fdb\u884c\u6a21\u578b\u5207\u5206\u7684\u590d\u6742\u5ea6\u3002\u5de5\u7a0b\u56e2\u961f\u5728\u4fee\u6539\u4ee3\u7801\u4e0a\u7684\u65f6\u95f4\u66f4\u5c11\uff0c\u4ece\u800c\u76f4\u63a5\u63d0\u5347\u8f6f\u4ef6\u5f00\u53d1\u6548\u7387\u3002<\/p>\n

      \u6d88\u9664\u6db2\u51b7\u57fa\u7840\u8bbe\u65bd\u5e26\u6765\u7684\u5ef6\u8fdf<\/h3>\n

      \u65b0\u4e00\u4ee3 GPU \u67b6\u6784\u7684\u70ed\u8bbe\u8ba1\u529f\u8017\u6781\u9ad8\uff0c\u5f80\u5f80\u5fc5\u987b\u4f9d\u8d56\u6db2\u51b7\u8bbe\u65bd\u3002\u8fc1\u79fb\u5230\u8fd9\u4e9b\u5e73\u53f0\u4f1a\u8feb\u4f7f\u673a\u623f\u56e2\u961f\u5b89\u88c5\u590d\u6742\u7684\u7ba1\u8def\u3001\u6db2\u51b7\u6362\u70ed\u5668\u4ee5\u53ca\u9ad8\u5bb9\u91cf\u4f9b\u7535\u5355\u5143\u3002\u8fd9\u4e9b\u7269\u7406\u57fa\u7840\u8bbe\u65bd\u6539\u9020\u4e0d\u4ec5\u65bd\u5de5\u5468\u671f\u957f\uff0c\u8fd8\u4f1a\u589e\u52a0\u8fd0\u7ef4\u98ce\u9669\u3002<\/p>\n

      \u9009\u62e9\u98ce\u51b7\u670d\u52a1\u5668\u5e73\u53f0\u800c\u975e\u9ad8\u5ea6\u4f9d\u8d56\u6db2\u51b7\u7684\u67b6\u6784\uff0c\u53ef\u4ee5\u8ba9\u4f60\u66f4\u5feb\u90e8\u7f72\u670d\u52a1\u5668\u786c\u4ef6\u3002\u4e0b\u9762\u7684\u8868\u683c\u5bf9\u6bd4\u4e86\u4e0d\u540c GPU \u4ee3\u9645\u5728\u70ed\u8bbe\u8ba1\u529f\u8017\u548c\u6563\u70ed\u9700\u6c42\u65b9\u9762\u7684\u4e3b\u8981\u5dee\u5f02\uff1a<\/p>\n

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

      GPU \u578b\u53f7<\/th>\n	\u70ed\u8bbe\u8ba1\u529f\u8017\uff08TDP\uff09<\/th>\n	\u6563\u70ed\u9700\u6c42<\/th>\n<\/tr>\n
      H200<\/td>\n	700W<\/td>\n	\u98ce\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\u987b\u6db2\u51b7<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n \u9009\u62e9\u98ce\u51b7\u786c\u4ef6\uff0c\u53ef\u4ee5\u8ba9\u4f60\u7acb\u5373\u5728\u6210\u719f\u7684\u8f6f\u4ef6\u6808\u4e0a\u8fd0\u884c\uff0c\u800c\u4e0d\u4f1a\u4e2d\u65ad\u73b0\u6709\u5de5\u4f5c\u6d41\u3002\u4f60\u4e5f\u80fd\u907f\u514d\u5b9a\u5236\u6db2\u51b7\u7ec4\u4ef6\u5728\u4f9b\u5e94\u94fe\u4e0a\u7684\u74f6\u9888\u3002\u6570\u636e\u79d1\u5b66\u56e2\u961f\u53ef\u4ee5\u66f4\u5feb\u90e8\u7f72\u6700\u5148\u8fdb\u7684\u57fa\u7840\u6a21\u578b\uff0c\u52a0\u901f\u7ec4\u7ec7\u7684\u4e0a\u5e02\u8282\u594f\u3002\u66f4\u5feb\u7684\u7cfb\u7edf\u90e8\u7f72\u53ef\u7acb\u5373\u5e26\u6765\u8fd0\u8425\u4ef7\u503c\uff0c\u540c\u65f6\u6709\u6548\u5b88\u4f4f\u603b\u4f53\u62e5\u6709\u6210\u672c\u3002<\/p>\n \u4f01\u4e1a TCO \u4e0e\u80fd\u6548\u4f18\u5316<\/strong><\/h2>\n \u5f53\u4f60\u90e8\u7f72\u5728\u65e2\u5b9a\u673a\u623f\u529f\u7387\u4e0a\u4ecd\u80fd\u63d0\u4f9b\u66f4\u5f3a\u7b97\u529b\u7684\u5904\u7406\u5668\u65f6\uff0c\u5c31\u80fd\u6700\u5927\u5316\u6570\u636e\u4e2d\u5fc3\u6548\u7387\u3002\u73b0\u4ee3 AI \u52a0\u901f\u5668\u5728\u63a7\u5236\u9ad8\u80fd\u8017\u7684\u540c\u65f6\uff0c\u63d0\u4f9b\u66f4\u9ad8\u7684\u8ba1\u7b97\u5bc6\u5ea6\u548c\u66f4\u4f18\u7684\u5355\u4f4d\u529f\u8017\u6027\u80fd\u3002<\/p>\n \u66f4\u9ad8\u7684\u8ba1\u7b97\u5bc6\u5ea6\u4e0e\u6027\u80fd\u529f\u8017\u6bd4<\/h3>\n H200 \u4e0d\u4ec5\u6027\u80fd\u66f4\u5f3a\uff0c\u800c\u4e14\u529f\u8017\u4e0e H100 \u6301\u5e73\u3002\u5728 LLM \u4efb\u52a1\u4e0a\u5b9e\u73b0 50% \u7684\u80fd\u8017\u4e0b\u964d\uff0c\u540c\u65f6\u5185\u5b58\u5e26\u5bbd\u7ffb\u500d\uff0c\u4f7f\u5176\u603b\u4f53\u62e5\u6709\u6210\u672c\uff08TCO\uff09\u6700\u591a\u964d\u4f4e 50%\u3002<\/p>\n \u5728\u5927\u89c4\u6a21\u8bad\u7ec3\u5468\u671f\u4e2d\uff0c\u4f60\u53ef\u4ee5\u901a\u8fc7\u56f4\u7ed5\u80fd\u6548\u4f18\u5316\u7684\u786c\u4ef6\u521b\u65b0\u83b7\u5f97\u53ef\u89c2\u6536\u76ca\u3002\u5148\u8fdb\u7cfb\u7edf\u67b6\u6784\u5728\u96c6\u7fa4\u8303\u56f4\u5185\u4f18\u5316\u4e86\u7535\u529b\u5206\u914d\uff1a<\/p>\n \n \u901a\u8fc7\u9ad8\u6548\u7684 HBM3e \u5185\u5b58\u6280\u672f\u63d0\u5347\u6027\u80fd\u529f\u8017\u6bd4\u3002<\/li>\n \u81ea\u9002\u5e94\u5de5\u4f5c\u8d1f\u8f7d\u8c03\u5ea6\u4f18\u5316 GPU \u5229\u7528\u7387\uff0c\u51cf\u5c11\u4e0d\u5fc5\u8981\u7684\u80fd\u8017\u3002<\/li>\n \u4e0e\u6db2\u51b7\u670d\u52a1\u5668\u96c6\u6210\uff0c\u53ef\u964d\u4f4e\u6570\u636e\u4e2d\u5fc3\u7684 PUE\uff08\u7535\u6e90\u4f7f\u7528\u6548\u7387\uff09\u3002<\/li>\n \u66f4\u77ed\u7684\u8bad\u7ec3\u65f6\u95f4\u76f4\u63a5\u51cf\u5c11\u6bcf\u6b21\u8bad\u7ec3\u8fd0\u884c\u7684\u603b\u80fd\u8017\u3002<\/li>\n \u66f4\u9ad8\u7684\u96c6\u7fa4\u5bc6\u5ea6\u63d0\u9ad8\u6574\u4f53\u6548\u7387\uff0c\u964d\u4f4e\u6bcf\u4e2a\u5de5\u4f5c\u8d1f\u8f7d\u7684\u603b\u80fd\u8017\u3002<\/li>\n<\/ul>\n \u8fd9\u4e9b\u6548\u7387\u63d0\u5347\u80fd\u5728\u4f60\u6bcf\u5929\u8fd0\u884c\u91cd\u8d1f\u8f7d\u4efb\u52a1\u65f6\uff0c\u7acb\u5373\u5e26\u6765\u8d22\u52a1\u5c42\u9762\u7684\u7f13\u89e3\u3002<\/p>\n \u7f29\u77ed\u96c6\u7fa4\u5b9e\u9645\u8fd0\u884c\u65f6\u957f\u4ee5\u964d\u4f4e\u8fd0\u8425\u6210\u672c<\/h3>\n \u5728\u91c7\u7528 H200 \u4e4b\u540e\uff0c\u6709\u62a5\u544a\u663e\u793a\u5176\u529f\u8017\u8282\u7701\u53ef\u8fbe 28%\uff0c\u4e14\u6bcf\u673a\u67b6\u541e\u5410\u91cf\u66f4\u9ad8\u3002<\/p>\n \u4f60\u53ef\u4ee5\u7528\u66f4\u5c11\u7684\u603b\u8fd0\u884c\u65f6\u95f4\u5b8c\u6210\u5927\u6a21\u578b\u8bad\u7ec3\u4efb\u52a1\uff0c\u4ece\u800c\u76f4\u63a5\u964d\u4f4e\u6570\u636e\u4e2d\u5fc3\u7535\u8d39\u3002\u66f4\u77ed\u7684\u8bad\u7ec3\u65f6\u95f4\u51cf\u5c11\u4e86\u6240\u6709\u914d\u5957\u8bbe\u5907\uff08\u5982\u670d\u52a1\u5668\u98ce\u6247\u548c\u673a\u623f\u51b7\u5374\u7cfb\u7edf\uff09\u7684\u6574\u4f53\u80fd\u8017\u3002\u7531\u4e8e\u66f4\u5feb\u7684\u786c\u4ef6\u5904\u7406\u80fd\u529b\u63d0\u5347\u4e86\u6bcf\u65e5\u603b\u8f93\u51fa\uff0c\u4f60\u7684\u786c\u4ef6\u6295\u8d44\u4e5f\u80fd\u66f4\u5feb\u56de\u672c\u3002<\/p>\n \u5728\u5904\u7406\u5927\u89c4\u6a21\u57fa\u7840\u6a21\u578b\u65f6\uff0c\u4f60\u7684\u4f01\u4e1a\u53ef\u4ee5\u5728\u6bcf\u6708\u7535\u8d39\u4e0a\u8282\u7701\u5927\u91cf\u5f00\u652f\u3002\u4e00\u7b14\u524d\u671f\u786c\u4ef6\u6295\u5165\uff0c\u5f80\u5f80\u5728\u4e00\u5e74\u5185\u5c31\u80fd\u56de\u672c\uff1b\u5728\u8fbe\u5230\u56de\u672c\u70b9\u540e\uff0c\u4f60\u7684\u6301\u7eed\u8fd0\u8425\u6210\u672c\u4f1a\u663e\u8457\u4e0b\u964d\u3002\u66f4\u5feb\u7684\u6a21\u578b\u5f00\u53d1\u8282\u594f\u80fd\u52a0\u901f\u8f6f\u4ef6\u53d1\u5e03\u8fdb\u5ea6\uff0c\u540c\u65f6\u5b88\u4f4f\u4f01\u4e1a\u8fd0\u8425\u9884\u7b97\u3002<\/p>\n \n <\/div>\n<\/div>\n NVIDIA H200 \u5f7b\u5e95\u91cd\u5851\u4e86\u4f60\u7684\u4eba\u5de5\u667a\u80fd\u6a21\u578b\u8bad\u7ec3\u80fd\u529b\u3002\u5176 141GB HBM3e \u5bb9\u91cf\u548c 4.8TB\/s \u5185\u5b58\u5e26\u5bbd\uff0c\u76f4\u63a5\u6d88\u9664\u4e86 GPU \u8ba1\u7b97\u7a7a\u8f6c\u65f6\u95f4\u3002\u4f60\u7684 Tensor Core \u80fd\u5728\u65e0\u9700\u7b49\u5f85\u6570\u636e\u4f20\u8f93\u7684\u60c5\u51b5\u4e0b\uff0c\u6301\u7eed\u5904\u7406\u6d77\u91cf\u6570\u636e\u96c6\u3002<\/p>\n \u901a\u8fc7\u4e0e\u73b0\u6709\u670d\u52a1\u5668\u7684\u5373\u63d2\u5373\u7528\u517c\u5bb9\u6027\uff0c\u4f60\u5728 2026 \u5e74\u83b7\u5f97\u4e86\u663e\u8457\u7684\u4e0a\u5e02\u65f6\u95f4\u4f18\u52bf\u3002\u4f60\u7684\u56e2\u961f\u53ef\u4ee5\u5c06\u9ad8\u5bc6\u5ea6\u8ba1\u7b97\u8282\u70b9\u76f4\u63a5\u90e8\u7f72\u5728\u6807\u51c6\u98ce\u51b7\u673a\u67b6\u4e2d\uff0c\u65e0\u9700\u6602\u8d35\u7684\u673a\u623f\u6539\u9020\u6216\u6db2\u51b7\u5efa\u8bbe\u5ef6\u8fdf\u3002<\/p>\n \u5bf9\u4e8e\u4f01\u4e1a\u7ea7 AI \u5de5\u4f5c\u8d1f\u8f7d\u800c\u8a00\uff0cNVIDIA H200 \u5728\u901f\u5ea6\u3001\u80fd\u6548\u4ee5\u53ca\u6295\u8d44\u56de\u62a5\u901f\u5ea6\u65b9\u9762\uff0c\u63d0\u4f9b\u4e86\u6781\u5177\u7ade\u4e89\u529b\u7684\u7efc\u5408\u8868\u73b0\u3002<\/p>\n \u5e38\u89c1\u95ee\u9898\uff08FAQ\uff09<\/strong><\/h2>\n NVIDIA H200 \u5728 AI \u8bad\u7ec3\u65b9\u9762\u6bd4 H100 \u5feb\u591a\u5c11\uff1f<\/h3>\n \u5bf9\u4e8e\u50cf Llama 2 \u8fd9\u6837\u7684\u5927\u6a21\u578b\uff0cNVIDIA H200 \u53ef\u63d0\u4f9b\u9ad8\u8fbe 2 \u500d\u7684\u541e\u5410\u91cf\u63d0\u5347\u3002MLPerf v4.0 \u57fa\u51c6\u6d4b\u8bd5\u663e\u793a\uff0c\u5728\u5355\u8282\u70b9\u4e0a\u8fdb\u884c Llama 2 70B \u7684 LoRA \u5fae\u8c03\u65f6\uff0cH200 \u76f8\u6bd4 H100 \u63d0\u5347 14%\uff0c\u5c06\u8bad\u7ec3\u65f6\u95f4\u7f29\u77ed\u81f3 24.7 \u5206\u949f\u3002<\/p>\n \u90e8\u7f72 NVIDIA H200 \u662f\u5426\u9700\u8981\u6db2\u51b7\uff1f<\/h3>\n \u4e0d\u9700\u8981\uff0c\u4f60\u65e0\u9700\u4e3a\u5176\u5efa\u8bbe\u6db2\u51b7\u57fa\u7840\u8bbe\u65bd\u3002NVIDIA H200 \u7684\u70ed\u8bbe\u8ba1\u529f\u8017\u4e3a 700W\uff0c\u53ef\u76f4\u63a5\u90e8\u7f72\u5230\u73b0\u6709\u7684\u4f01\u4e1a\u98ce\u51b7\u670d\u52a1\u5668\u673a\u67b6\u4e2d\uff0c\u65e0\u9700\u6602\u8d35\u7684\u673a\u623f\u5347\u7ea7\u6216\u65bd\u5de5\u5ef6\u8fdf\u3002<\/p>\n \u4e3a\u4ec0\u4e48 HBM3e \u5185\u5b58\u53ef\u4ee5\u63d0\u5347 LLM \u8bad\u7ec3\u901f\u5ea6\uff1f<\/h3>\n H200 \u63d0\u4f9b 141GB \u7684 HBM3e \u5185\u5b58\u4e0e 4.8TB\/s \u7684\u5185\u5b58\u5e26\u5bbd\uff0c\u6784\u6210\u4e00\u6761\u201c\u8d85\u5bbd\u6570\u636e\u7ba1\u9053\u201d\uff0c\u53ef\u4ee5\u6301\u7eed\u4e3a Tensor Core \u4f9b\u6570\u3002\u9ad8\u5e26\u5bbd\u53ef\u4ee5\u9632\u6b62\u8ba1\u7b97\u5f15\u64ce\u5728\u5927\u89c4\u6a21\u6570\u5b66\u8ba1\u7b97\u4e2d\u9891\u7e41\u6682\u505c\uff0c\u8ba9\u5904\u7406\u5668\u5728\u5927\u6279\u91cf\u8bad\u7ec3\u4efb\u52a1\u4e2d\u59cb\u7ec8\u4fdd\u6301\u6d3b\u8dc3\u3002<\/p>\n NVIDIA H200 \u662f\u5426\u53ef\u4ee5\u964d\u4f4e\u4f01\u4e1a\u8fd0\u8425\u6210\u672c\uff1f<\/h3>\n \u53ef\u4ee5\uff0c\u4f60\u53ef\u4ee5\u5c06\u603b\u4f53\u62e5\u6709\u6210\u672c\u6700\u591a\u964d\u4f4e 50%\u3002H200 \u5728\u4fdd\u6301\u76f8\u540c\u6027\u80fd\u7684\u524d\u63d0\u4e0b\u53ef\u8282\u7701\u7ea6 28% \u7684\u529f\u8017\uff0c\u5e76\u7f29\u77ed\u96c6\u7fa4\u7684\u5b9e\u9645\u8bad\u7ec3\u65f6\u95f4\u3002\u66f4\u5feb\u7684\u8bad\u7ec3\u4efb\u52a1\u5728\u4e0d\u63d0\u5347\u673a\u623f\u529f\u7387\u4e0a\u9650\u7684\u60c5\u51b5\u4e0b\uff0c\u663e\u8457\u964d\u4f4e\u6bcf\u6708\u80fd\u8017\u3002<\/p>\n","protected":false},"excerpt":{"rendered":" \u501f\u52a9 NVIDIA H200 GPU\uff0c\u4f60\u53ef\u4ee5\u5927\u5e45\u52a0\u901f\u4f01\u4e1a\u7ea7 AI \u6a21\u578b\u8bad\u7ec3\u3002\u8fd9\u6b3e\u5904\u7406\u5668\u901a\u8fc7\u786c\u4ef6\u5347\u7ea7\u6d88\u9664\u4e25\u91cd\u7684\u5185\u5b58\u74f6\u9888\uff1a \u5185\u5b58\u5bb9\u91cf\uff1a141 GB HBM3e \u5185\u5b58\u5e26\u5bbd\uff1a4.8 TB\/s MLPerf Training v4.0 \u57fa\u51c6\u6d4b\u8bd5\u5c55\u793a\u4e86\u8fd9\u4e9b\u52a0\u901f\u6548\u679c\uff1a \u57fa\u51c6\u9879\u76ee \u6027\u80fd\u63d0\u5347 \u914d\u7f6e Llama 2 70B LoRA \u5fae\u8c03 \u63d0\u901f 14%\uff08\u8bad\u7ec3\u65f6\u95f4\u7f29\u77ed\u81f3 24.7 \u5206\u949f\uff09 \u5355\u8282\u70b9\uff0cNVIDIA H200 \u8fd9\u4e00\u67b6\u6784\u4e3a Llama 2 \u7b49\u5927\u578b\u6a21\u578b\u63d0\u4f9b\u9ad8\u8fbe 2 \u500d\u7684\u541e\u5410\u91cf\u63d0\u5347\u3002\u5230 2026 \u5e74\uff0c\u4f60\u7684\u7ec4\u7ec7\u53ef\u4ee5\u5728\u6807\u51c6\u98ce\u51b7\u9999\u6e2f\u72ec\u7acb\u670d\u52a1\u5668\u57fa\u7840\u8bbe\u65bd\u53ca\u5176\u4ed6\u5e38\u89c4\u6570\u636e\u4e2d\u5fc3\u4e2d\uff0c\u7acb\u5373\u5b9e\u73b0\u65e0\u7f1d\u90e8\u7f72\u3002\u4f60\u53ef\u4ee5\u907f\u514d\u590d\u6742\u7684\u6db2\u51b7\u673a\u623f\u6539\u9020\u5ef6\u8fdf\uff0c\u540c\u65f6\u6709\u6548\u4f18\u5316\u603b\u4f53\u62e5\u6709\u6210\u672c\uff08TCO\uff09\u3002 \u5173\u952e\u8981\u70b9 NVIDIA H200 GPU \u642d\u8f7d 141 GB HBM3e \u5185\u5b58\u548c 4.8 TB\/s \u5e26\u5bbd\uff0c\u53ef\u663e\u8457\u52a0\u901f AI \u8bad\u7ec3\u3002 \u5de5\u7a0b\u5e08\u5728\u5fae\u8c03 Llama 2 \u7b49\u5927\u578b […]<\/p>\n Read More…<\/a><\/p>\n","protected":false},"author":2,"featured_media":34202,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[62],"tags":[12927,12869,4279,3863,3321],"class_list":["post-34209","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-hong-kong-dedicated-server-sc","tag-ai-training-speed","tag-hbm3e-memory","tag-nvidia-h200-sc","tag-data-center-efficiency-sc","tag-gpu-performance-sc"],"acf":[],"yoast_head":"\n