\u4f60\u53ef\u4ee5\u53d7\u76ca\u65bc\u9ad8\u5ea6\u6574\u5408\u7684\u7cfb\u7d71\uff0c\u5c07 AMD Instinct MI400 \u7cfb\u5217 GPU \u8207 AMD EPYC \u8655\u7406\u5668\u4ee5\u53ca\u5148\u9032\u7db2\u8def\u76f8\u7d50\u5408\u3002AMD \u6b63\u5728\u76f4\u63a5\u6311\u6230 NVIDIA \u7684\u5e02\u5834\u4e3b\u5c0e\u5730\u4f4d\uff0c\u70ba AI \u8a13\u7df4\u3001\u63a8\u7406\u548c\u5927\u578b\u8a9e\u8a00\u6a21\u578b\u7b49\u5c08\u696d\u9700\u6c42\u63d0\u4f9b\u7368\u7279\u7684\u5ba2\u88fd\u5316\u52a0\u901f\u5668\u89e3\u6c7a\u65b9\u6848\u3002<\/p>\n

\u95dc\u9375\u8981\u9ede<\/strong><\/h2>\n

\n
• AMD Instinct MI400 \u7cfb\u5217 GPU \u6191\u85c9\u5148\u9032\u7684 CDNA \u67b6\u69cb\u8207\u5927\u5bb9\u91cf HBM4 \u8a18\u61b6\u9ad4\uff0c\u70ba\u9ad8\u8981\u6c42\u7684 AI \u8207 HPC \u5de5\u4f5c\u8ca0\u8f09\u63d0\u4f9b\u9ad8\u6548\u80fd\u8207\u9ad8\u6548\u7387\u3002<\/li>\n
• \u6191\u85c9\u9ad8\u9054 288 TFLOPS \u7684\u904b\u7b97\u80fd\u529b\u8207 432GB HBM4 \u8a18\u61b6\u9ad4\uff0c\u4f60\u53ef\u4ee5\u66f4\u5feb\u5b8c\u6210\u5927\u578b\u8a9e\u8a00\u6a21\u578b\u8a13\u7df4\u8207\u63a8\u7406\uff0c\u6e1b\u5c11\u74f6\u9838\u4e26\u63d0\u5347\u8cc7\u6e90\u4f7f\u7528\u7387\u3002<\/li>\n
• MI400 \u7cfb\u5217\u652f\u63f4 PyTorch \u548c TensorFlow \u7b49\u4e3b\u6d41 AI \u6846\u67b6\uff0c\u53ef\u8f15\u9b06\u6574\u5408\uff0c\u7121\u9700\u4fee\u6539\u7a0b\u5f0f\u78bc\uff0c\u78ba\u4fdd\u9806\u66a2\u9077\u79fb\u5230 AMD \u786c\u9ad4\u3002<\/li>\n
• \u53ef\u64f4\u5c55\u6027\u662f\u6838\u5fc3\u512a\u52e2\u4e4b\u4e00\uff0c\u4f60\u53ef\u4ee5\u9023\u63a5\u591a\u584a GPU \u4e26\u5efa\u69cb\u53e2\u96c6\uff0c\u7528\u65bc\u5927\u898f\u6a21 AI \u8a13\u7df4\u8207\u63a8\u7406\uff0c\u9069\u7528\u65bc\u4f01\u696d\u8207\u7814\u7a76\u74b0\u5883\u3002<\/li>\n
• \u53ef\u5ba2\u88fd\u5316\u7684 MI400 \u52a0\u901f\u5361\u53ef\u4ee5\u6eff\u8db3\u7279\u5b9a\u5de5\u4f5c\u8ca0\u8f09\u9700\u6c42\uff0c\u70ba\u5f9e AI \u63a8\u7406\u5230\u9ad8\u7cbe\u5ea6 HPC \u4efb\u52d9\u7684\u591a\u7a2e\u61c9\u7528\u63d0\u4f9b\u6700\u4f73\u5316\u6548\u80fd\u3002<\/li>\n<\/ul>\n

  AMD Instinct MI400 \u7cfb\u5217\u7279\u6027<\/strong><\/h2>\n

  CDNA 5 \u67b6\u69cb\u8207 AI \u5f37\u5316<\/h3>\n

  \u501f\u52a9 AMD Instinct MI400 \u7cfb\u5217\uff0c\u4f60\u53ef\u4ee5\u4f7f\u7528\u9818\u5148\u7684\u6280\u8853\u3002CDNA 5 \u67b6\u69cb\u70ba AI \u5de5\u4f5c\u8ca0\u8f09\u5e36\u4f86\u591a\u9805\u6539\u9032\uff0c\u4f60\u5c07\u7372\u5f97\u66f4\u9ad8\u7684\u904b\u7b97\u6548\u80fd\u4ee5\u53ca\u5c0d\u73fe\u4ee3 AI \u8cc7\u6599\u578b\u5225\u7684\u652f\u63f4\u3002<\/p>\n

  \n
  • \u8a18\u61b6\u9ad4\uff1a\u55ae\u5361 432GB HBM4\uff0c\u983b\u5bec 23.3 TB\/s\uff0c\u53ef\u56e0\u61c9\u8d85\u5927\u898f\u6a21\u8cc7\u6599\u96c6\u3002<\/li>\n
  • \u904b\u7b97\u6548\u80fd\uff1a\u63d0\u4f9b 20 PF \u5cf0\u503c MXFP8 \u8207 40 PF \u5cf0\u503c MXFP4\uff0c\u652f\u63f4 FP4\u3001FP6 \u8207 FP8 \u8cc7\u6599\u578b\u5225\u3002<\/li>\n
  • \u67b6\u69cb\u5275\u65b0\uff1a\u66f4\u5927\u7684\u5feb\u53d6\u3001\u591a\u64ad\u8a18\u61b6\u9ad4\u64cd\u4f5c\u3001Tensor Data Mover \u5f37\u5316\u3001Wave32 \u57f7\u884c\u4ee5\u53ca UALoE \u6a6b\u5411\u64f4\u5c55\u4e92\u9023\uff0c\u6709\u52a9\u65bc\u964d\u4f4e\u8cc7\u6599\u642c\u79fb\u8207\u7ba1\u7406\u958b\u92b7\u3002<\/li>\n<\/ul>\n

    \u9019\u4e9b\u7279\u6027\u8b93\u4f60\u53ef\u4ee5\u66f4\u5feb\u8a13\u7df4\u8207\u90e8\u7f72\u524d\u6cbf AI \u6a21\u578b\uff0c\u964d\u4f4e\u7cfb\u7d71\u74f6\u9838\uff0c\u66f4\u6709\u6548\u7387\u5730\u5229\u7528\u8cc7\u6e90\u3002<\/p>\n

    HBM4 \u8a18\u61b6\u9ad4\u8207 L2 \u5feb\u53d6<\/h3>\n

    \u5728\u901f\u5ea6\u8207\u5bb9\u91cf\u65b9\u9762\uff0c\u4f60\u9700\u8981\u4f9d\u8cf4 HBM4 \u8a18\u61b6\u9ad4\u3002MI400 \u63a1\u7528 HBM4\uff0c\u55ae\u5361\u53ef\u63d0\u4f9b\u9ad8\u9054 432GB \u5bb9\u91cf\u8207 23.3 TB\/s \u983b\u5bec\uff0c\u8db3\u4ee5\u652f\u6490\u5927\u578b\u8a9e\u8a00\u6a21\u578b\u548c\u8907\u96dc AI \u4efb\u52d9\u3002<\/p>\n

    
    \n\u4f60\u540c\u6a23\u53d7\u76ca\u65bc\u66f4\u5927\u7684 L2 \u5feb\u53d6\uff0c\u53ef\u5c07\u5e38\u7528\u8cc7\u6599\u5b58\u653e\u5728\u66f4\u9760\u8fd1\u904b\u7b97\u55ae\u5143\u7684\u4f4d\u7f6e\uff0c\u5f9e\u800c\u964d\u4f4e\u5ef6\u9072\u4e26\u63d0\u5347\u541e\u5410\u91cf\u3002
    \n\u4f60\u7684\u7cfb\u7d71\u53ef\u4ee5\u66f4\u5feb\u5730\u8655\u7406\u8cc7\u6599\uff0c\u4e26\u5728 AI \u8207 HPC \u5de5\u4f5c\u8ca0\u8f09\u4e2d\u7372\u5f97\u66f4\u9ad8\u6548\u80fd\u3002<\/span><\/p>\n

    \u53ef\u64f4\u5c55\u7cfb\u7d71\u8a2d\u8a08<\/h3>\n

    \u501f\u52a9 AMD Instinct MI400 \u7cfb\u5217\uff0c\u4f60\u53ef\u4ee5\u8f15\u9b06\u64f4\u5c55\u7cfb\u7d71\u3002\u8a72\u67b6\u69cb\u652f\u63f4\u6700\u591a 72 \u584a GPU \u8207 AMD EPYC \u8655\u7406\u5668\u5354\u540c\u5de5\u4f5c\uff0c\u4e26\u900f\u904e AMD Pensando DPU \u63d0\u4f9b\u9032\u968e\u8cc7\u6599\u8655\u7406\u80fd\u529b\u3002<\/p>\n

    AMD Helios \u6a5f\u67b6\u7d1a\u89e3\u6c7a\u65b9\u6848\u70ba\u73fe\u4ee3 AI \u8cc7\u6599\u4e2d\u5fc3\u63d0\u4f9b\u9748\u6d3b\u3001\u53ef\u64f4\u5c55\u7684\u8a2d\u8a08\u3002\u4f60\u53ef\u4ee5\u7372\u5f97\u9ad8\u9054 1.4 ExaFLOPS \u7684 FP8 \u904b\u7b97\u80fd\u529b\u548c 2.9 ExaFLOPS \u7684 FP4 \u904b\u7b97\u80fd\u529b\uff0c\u4ee5\u53ca 31TB \u7684 HBM4 \u8a18\u61b6\u9ad4\u3002<\/p>\n

    
    \n\u4ee5\u4e0b\u662f\u7cfb\u7d71\u8a2d\u8a08\u5728\u6574\u5408\u65b9\u9762\u7684\u652f\u63f4\u65b9\u5f0f\uff1a<\/span><\/p>\n

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

    \u7279\u6027<\/th>\n	\u8aaa\u660e<\/th>\n<\/tr>\n
    Helios \u6a5f\u67b6\u7d1a\u7cfb\u7d71<\/td>\n	\u63d0\u4f9b\u9810\u5148\u914d\u7f6e\u6a21\u7d44\uff0c\u4fbf\u65bc\u5c07 AMD Instinct GPU \u8207 EPYC CPU \u8f15\u9b06\u6574\u5408\u3002<\/td>\n<\/tr>\n
    \u9762\u5411 AI \u8207 HPC \u6700\u4f73\u5316<\/td>\n	\u5c08\u70ba AI \u8207\u9ad8\u6548\u80fd\u904b\u7b97\u5de5\u4f5c\u8ca0\u8f09\u8a2d\u8a08\uff0c\u4ee5\u63d0\u5347\u6574\u9ad4\u6548\u80fd\u3002<\/td>\n<\/tr>\n
    \u958b\u653e\u8edf\u9ad4\u57fa\u790e<\/td>\n	\u70ba\u5ba2\u6236\u5728\u5927\u898f\u6a21\u74b0\u5883\u4e2d\u5275\u65b0\u548c\u8abf\u6574\u89e3\u6c7a\u65b9\u6848\u63d0\u4f9b\u5f48\u6027\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n \u4f60\u53ef\u4ee5\u57fa\u65bc Meta \u7684 Open Rack Wide \u6a19\u6e96\uff0c\u5728\u96d9\u5bec\u6a5f\u67b6\u4e2d\u90e8\u7f72\u9019\u4e9b\u7cfb\u7d71\u3002\u6b64\u4e00\u8a2d\u8a08\u6709\u52a9\u65bc\u4f60\u5efa\u69cb\u9ad8\u6548\u4e14\u5f37\u5927\u7684 AI \u57fa\u790e\u8a2d\u65bd\u3002<\/p>\n AI \u5de5\u4f5c\u8ca0\u8f09\u6548\u80fd<\/strong><\/h2>\n \u8a13\u7df4\u8207\u63a8\u7406\u57fa\u6e96<\/h3>\n \u4f60\u5e0c\u671b\u770b\u5230\u7cfb\u7d71\u5728\u771f\u5be6 AI \u5de5\u4f5c\u8ca0\u8f09\u4e0b\u7684\u8868\u73fe\u3002AMD Instinct MI400 \u7cfb\u5217\u5728\u8a13\u7df4\u8207\u63a8\u7406\u65b9\u9762\u90fd\u5c55\u73fe\u51fa\u5f37\u52c1\u6548\u80fd\u3002\u6191\u85c9\u5f37\u5927\u7684\u904b\u7b97\u80fd\u529b\u8207\u8a18\u61b6\u9ad4\u983b\u5bec\uff0c\u4f60\u53ef\u4ee5\u66f4\u5feb\u8a13\u7df4\u5927\u578b\u8a9e\u8a00\u6a21\u578b\u3002\u4f8b\u5982\uff0cMI400 \u53ef\u5be6\u73fe\u9ad8\u9054 20 PF \u5cf0\u503c MXFP8 \u8207 40 PF \u5cf0\u503c MXFP4 \u7684\u904b\u7b97\u6548\u80fd\uff0c\u9019\u610f\u5473\u8457\u4f60\u53ef\u4ee5\u8f15\u9b06\u61c9\u5c0d\u8907\u96dc\u6a21\u578b\u8207\u5927\u578b\u8cc7\u6599\u96c6\u3002<\/p>\n \u5728\u57f7\u884c\u63a8\u7406\u4efb\u52d9\u6642\uff0c\u4f60\u6703\u767c\u73fe\u5ef6\u9072\u66f4\u4f4e\u3001\u541e\u5410\u91cf\u66f4\u9ad8\uff0c\u6709\u5229\u65bc\u90e8\u7f72\u56de\u61c9\u8fc5\u901f\u7684 AI \u61c9\u7528\u3002\u8a31\u591a\u6a5f\u69cb\u5df2\u5728\u79d1\u5b78\u904b\u7b97\u3001\u81ea\u7136\u8a9e\u8a00\u8655\u7406\u8207\u5f71\u50cf\u8fa8\u8b58\u7b49\u5834\u666f\u4e2d\u4f7f\u7528 MI400 GPU\uff0c\u4f60\u53ef\u4ee5\u5728\u5c0f\u898f\u6a21\u8207\u5927\u898f\u6a21\u8a13\u7df4\u548c\u63a8\u7406\u4efb\u52d9\u4e2d\u7372\u5f97\u7a69\u5b9a\u53ef\u9760\u7684\u8868\u73fe\u3002<\/p>\n \u63d0\u793a\uff1a<\/strong> \u4f60\u53ef\u4ee5\u900f\u904e\u5c07 Instinct \u52a0\u901f\u5361\u8207 AMD EPYC \u8655\u7406\u5668\u642d\u914d\uff0c\u4e26\u63a1\u7528\u6700\u4f73\u5316\u7684\u8edf\u9ad4\u5806\u758a\uff0c\u9032\u4e00\u6b65\u91cb\u653e\u7cfb\u7d71\u6548\u80fd\u3002<\/p><\/blockquote>\n \u652f\u63f4\u7684 AI \u6846\u67b6<\/h3>\n \u4f60\u9700\u8981\u5e38\u7528\u5de5\u5177\u80fd\u958b\u7bb1\u5373\u7528\u3002AMD Instinct MI400 \u7cfb\u5217\u652f\u63f4\u4e3b\u6d41 AI \u6846\u67b6\uff0c\u4f60\u53ef\u4ee5\u5728\u5c08\u6848\u4e2d\u4f7f\u7528 PyTorch\u3001TensorFlow \u548c JAX\u3002ROCm \u8edf\u9ad4\u751f\u614b\u78ba\u4fdd\u4f60\u7121\u9700\u4fee\u6539\u7a0b\u5f0f\u78bc\uff0c\u5c31\u80fd\u5c07\u73fe\u6709\u5de5\u4f5c\u8ca0\u8f09\u9077\u79fb\u5230 AMD \u786c\u9ad4\u3002<\/p>\n \u4ee5\u4e0b\u662f\u5404\u6846\u67b6\u7684\u6700\u4f73\u5316\u60c5\u6cc1\u4e00\u89bd\uff1a<\/p>\n \n\n\n\n<\/colgroup>\n\n\n\n\n\n AI \u6846\u67b6<\/th>\n \u6700\u4f73\u5316\u7b49\u7d1a<\/th>\n<\/tr>\n PyTorch<\/td>\n \u6548\u80fd\u6301\u5e73<\/td>\n<\/tr>\n TensorFlow<\/td>\n \u6548\u80fd\u6301\u5e73<\/td>\n<\/tr>\n JAX<\/td>\n \u63a5\u8fd1\u6301\u5e73<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n \u4f60\u53ef\u4ee5\u653e\u5fc3\u57f7\u884c AI \u5de5\u4f5c\u8ca0\u8f09\u3002ROCm \u5e73\u53f0\u70ba\u4f60\u63d0\u4f9b\u8c50\u5bcc\u7684\u51fd\u5f0f\u5eab\u8207\u5de5\u5177\uff0c\u5354\u52a9\u4f60\u5145\u5206\u767c\u63ee GPU \u6548\u80fd\uff0c\u8b93\u4f60\u628a\u7cbe\u529b\u66f4\u591a\u653e\u5728\u6a21\u578b\u5efa\u69cb\u8207\u8a13\u7df4\uff0c\u800c\u4e0d\u662f\u76f8\u5bb9\u6027\u6392\u67e5\u3002<\/p>\n \u9762\u5411\u5927\u578b\u6a21\u578b\u7684\u53ef\u64f4\u5c55\u6027<\/h3>\n \u96a8\u8457\u6a21\u578b\u4e0d\u65b7\u8b8a\u5927\uff0c\u4f60\u5f80\u5f80\u9700\u8981\u6301\u7e8c\u64f4\u5c55\u7cfb\u7d71\u898f\u6a21\u3002MI400 \u67b6\u69cb\u5141\u8a31\u4f60\u9023\u63a5\u591a\u584a GPU \u5efa\u69cb\u53e2\u96c6\uff0c\u4ee5\u652f\u6490\u5927\u898f\u6a21\u8a13\u7df4\u8207\u63a8\u7406\uff0c\u9019\u610f\u5473\u8457\u4f60\u53ef\u4ee5\u8a13\u7df4\u64c1\u6709\u6578\u5341\u5104\u53c3\u6578\u7684\u524d\u6cbf AI \u6a21\u578b\u3002<\/p>\n \u4f60\u5c07\u53d7\u76ca\u65bc\u9ad8\u8a18\u61b6\u9ad4\u5bb9\u91cf\u8207\u9ad8\u901f\u4e92\u9023\uff0c\u9019\u4e9b\u7279\u6027\u6709\u52a9\u65bc\u5728\u8655\u7406\u5927\u898f\u6a21\u8cc7\u6599\u96c6\u6642\u907f\u514d\u74f6\u9838\u3002\u7121\u8ad6\u662f\u5728\u8cc7\u6599\u4e2d\u5fc3\u9084\u662f\u7814\u7a76\u5be6\u9a57\u5ba4\uff0c\u4f60\u90fd\u53ef\u4ee5\u90e8\u7f72\u57fa\u65bc AMD \u89e3\u6c7a\u65b9\u6848\u7684\u7cfb\u7d71\uff0c\u6eff\u8db3\u65e2\u8981\u901f\u5ea6\u53c8\u8981\u898f\u6a21\u7684\u5de5\u4f5c\u8ca0\u8f09\u3002<\/p>\n \u6ce8\u610f\uff1a<\/strong> \u4f60\u53ef\u4ee5\u6839\u64da\u9700\u6c42\u8b8a\u5316\u64f4\u5c55\u57fa\u790e\u8a2d\u65bd\u3002AMD Instinct \u786c\u9ad4\u7684\u5f48\u6027\u8a2d\u8a08\u652f\u63f4\u672a\u4f86\u5728 AI \u8207 HPC \u5de5\u4f5c\u8ca0\u8f09\u4e0a\u7684\u6301\u7e8c\u6210\u9577\u3002<\/p><\/blockquote>\n \u5ba2\u88fd\u5316 AMD Instinct MI400 \u52a0\u901f\u5668\u8207\u80fd\u6548<\/strong><\/h2>\n \u57fa\u65bc MI450 \u7684\u5ba2\u88fd\u5316\u65b9\u6848<\/h3>\n \u4f60\u53ef\u4ee5\u5f9e\u591a\u6b3e\u5ba2\u88fd\u5316 AMD Instinct MI400 \u52a0\u901f\u5668\u4e2d\u9078\u64c7\uff0c\u4ee5\u5339\u914d\u7368\u7279\u7684 AI \u5de5\u4f5c\u8ca0\u8f09\u9700\u6c42\u3002\u57fa\u65bc MI450 \u7684\u65b9\u6848\u5c24\u5176\u7a81\u51fa\uff0c\u9762\u5411\u7279\u5b9a\u4f01\u696d\u8207\u7814\u7a76\u61c9\u7528\u3002\u4f8b\u5982\uff0cMI450X \u5c08\u6ce8\u5927\u898f\u6a21 AI \u5de5\u4f5c\u8ca0\u8f09\uff0c\u70ba\u63a8\u7406\u4efb\u52d9\u63d0\u4f9b\u6700\u4f73\u5316\u6548\u80fd\uff1bMI430X \u5247\u805a\u7126\u9ad8\u7cbe\u5ea6 HPC \u8207\u4e3b\u6b0a AI \u5de5\u4f5c\u8ca0\u8f09\uff0c\u63a1\u7528\u5f37\u5927\u7684 FP64 \u8a08\u7b97\u80fd\u529b\u548c CPU+GPU \u6df7\u5408\u8a2d\u8a08\u3002<\/p>\n \u4e0b\u8868\u662f\u9019\u4e9b\u5ba2\u88fd\u578b\u865f\u7684\u7c21\u8981\u6bd4\u8f03\uff1a<\/p>\n \n\n\n\n\n<\/colgroup>\n\n\n\n\n \u578b\u865f<\/th>\n \u76ee\u6a19\u61c9\u7528<\/th>\n \u6548\u80fd\u7279\u6027<\/th>\n<\/tr>\n MI450X<\/td>\n \u5927\u898f\u6a21 AI \u5de5\u4f5c\u8ca0\u8f09<\/td>\n \u91dd\u5c0d AI \u63a8\u7406\u6700\u4f73\u5316\uff0c\u652f\u63f4 Ultra Ethernet<\/td>\n<\/tr>\n MI430X<\/td>\n HPC \u8207\u4e3b\u6b0a AI \u5de5\u4f5c\u8ca0\u8f09<\/td>\n \u786c\u9ad4\u7d1a FP64\uff0cCPU+GPU \u6df7\u5408\u904b\u7b97\u67b6\u69cb<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n \u9019\u4e9b\u5ba2\u88fd\u65b9\u6848\u662f\u8207\u5408\u4f5c\u5925\u4f34\uff08\u5982 Meta\uff09\u5171\u540c\u8a2d\u8a08\uff0c\u56e0\u6b64\u80fd\u66f4\u597d\u5730\u76f8\u5bb9\u4e3b\u6d41\u6a21\u578b\u67b6\u69cb\u3002\u4f60\u53ef\u4ee5\u7372\u5f97\u4f4e\u5ef6\u9072\u8207\u9ad8\u541e\u5410\u8868\u73fe\uff0c\u9019\u5c0d\u65bc\u9762\u5411\u7d42\u7aef\u4f7f\u7528\u8005\u7684 AI \u61c9\u7528\u5c24\u70ba\u95dc\u9375\u3002<\/p>\n \u63d0\u793a\uff1a<\/strong> \u5ba2\u88fd\u5316\u89e3\u6c7a\u65b9\u6848\u53ef\u4ee5\u5e6b\u52a9\u4f60\u5728\u4f01\u696d\u74b0\u5883\u4e2d\u6eff\u8db3\u56b4\u82db\u7684\u6548\u80fd\u8207\u76f8\u5bb9\u6027\u8981\u6c42\u3002<\/p><\/blockquote>\n \u8a18\u61b6\u9ad4\u6548\u7387\u8207\u5de5\u4f5c\u8ca0\u8f09\u6700\u4f73\u5316<\/h3>\n \u5728\u8655\u7406\u5927\u578b AI \u6a21\u578b\u8207\u8cc7\u6599\u6642\uff0c\u4f60\u9700\u8981\u9ad8\u6548\u7387\u7684\u8a18\u61b6\u9ad4\u7cfb\u7d71\u3002\u5ba2\u88fd\u5316 AMD Instinct MI400 \u52a0\u901f\u5668\u76f8\u8f03\u6a19\u6e96\u578b\u865f\u63d0\u4f9b\u66f4\u9ad8\u7684\u8a18\u61b6\u9ad4\u5bb9\u91cf\u8207\u983b\u5bec\u3002\u4f8b\u5982\uff0c\u8207\u6a19\u6e96\u578b\u865f\u7684 288GB\u30018 TB\/s \u76f8\u6bd4\uff0c\u4f60\u53ef\u4ee5\u7372\u5f97\u9ad8\u9054 432GB \u8a18\u61b6\u9ad4\u8207 19.6 TB\/s \u983b\u5bec\u3002\u6a6b\u5411\u64f4\u5c55\u983b\u5bec\u4e5f\u53ef\u9054\u5230 300 GB\/s\uff0c\u80fd\u66f4\u5feb\u5730\u5728\u52a0\u901f\u5668\u4e4b\u9593\u50b3\u8f38\u8cc7\u6599\u3002<\/p>\n \n\n\n\n\n<\/colgroup>\n\n\n\n\n\n \u7279\u6027<\/th>\n MI400 \u7cfb\u5217<\/th>\n \u6a19\u6e96\u578b\u865f<\/th>\n<\/tr>\n \u8a18\u61b6\u9ad4\u5bb9\u91cf<\/td>\n 432 GB<\/td>\n 288 GB<\/td>\n<\/tr>\n \u8a18\u61b6\u9ad4\u983b\u5bec<\/td>\n 19.6 TB\/s<\/td>\n 8 TB\/s<\/td>\n<\/tr>\n \u6a6b\u5411\u64f4\u5c55\u983b\u5bec<\/td>\n 300 GB\/s<\/td>\n N\/A<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n \u900f\u904e\u904b\u7528\u9019\u4e9b\u984d\u5916\u7684\u8a18\u61b6\u9ad4\u8207\u983b\u5bec\uff0c\u4f60\u53ef\u4ee5\u5c0d AI \u5de5\u4f5c\u8ca0\u8f09\u9032\u884c\u6df1\u5ea6\u6700\u4f73\u5316\uff0c\u5f9e\u800c\u8a13\u7df4\u66f4\u5927\u6a21\u578b\u3001\u8655\u7406\u66f4\u591a\u8cc7\u6599\u4e26\u6e1b\u5c11\u74f6\u9838\u3002Instinct \u67b6\u69cb\u5c0d\u9019\u4e9b\u63d0\u5347\u63d0\u4f9b\u539f\u751f\u652f\u63f4\uff0c\u56e0\u6b64\u7121\u8ad6\u8a13\u7df4\u9084\u662f\u63a8\u7406\uff0c\u4f60\u90fd\u80fd\u7372\u5f97\u66f4\u4f73\u6210\u6548\u3002<\/p>\n \u6ce8\u610f\uff1a<\/strong> \u5f97\u76ca\u65bc AMD \u5c0d\u8a18\u61b6\u9ad4\u6548\u7387\u7684\u91cd\u8996\uff0c\u4f60\u53ef\u4ee5\u5728\u4e0d\u53d7\u786c\u9ad4\u9650\u5236\u675f\u7e1b\u7684\u524d\u63d0\u4e0b\uff0c\u6301\u7e8c\u64f4\u5c55 AI \u5c08\u6848\u898f\u6a21\u3002<\/p><\/blockquote>\n AMD \u8207 NVIDIA\uff1a\u5e02\u5834\u6bd4\u8f03<\/strong><\/h2>\n \u6548\u80fd\u8207\u80fd\u6548<\/h3>\n \u4f60\u4e5f\u95dc\u5fc3 AMD Instinct MI400 \u7cfb\u5217\u8207 NVIDIA GPU \u7684\u6bd4\u8f03\u3002MI400 \u7cfb\u5217\u5728 AI \u5de5\u4f5c\u8ca0\u8f09\uff0c\u5c24\u5176\u662f\u5927\u898f\u6a21\u8a13\u7df4\u8207\u63a8\u7406\u4e2d\uff0c\u5c55\u73fe\u51fa\u5f37\u52c1\u6548\u80fd\uff0c\u517c\u5177\u9ad8\u904b\u7b97\u80fd\u529b\u8207\u9ad8\u8a18\u61b6\u9ad4\u983b\u5bec\u3002\u5728\u80fd\u6548\u65b9\u9762\uff0c\u4f60\u6703\u770b\u5230\u5728\u529f\u8017\u8207\u55ae\u4f4d\u529f\u8017\u6548\u80fd\uff08Performance per Watt\uff09\u4e0a\u7684\u5dee\u7570\u3002<\/p>\n \n\n\n\n\n\n<\/colgroup>\n\n\n\n\n GPU \u578b\u865f<\/th>\n \u7e3d\u7e6a\u5716\u529f\u8017 (TGP)<\/th>\n \u5cf0\u503c\u6548\u80fd (FLOPS)<\/th>\n \u55ae\u4f4d\u529f\u8017\u6548\u80fd (GFLOPS\/W)<\/th>\n<\/tr>\n Nvidia B200<\/td>\n 1000 W<\/td>\n \u7d04 9 PFLOPS<\/td>\n \u7d04 9 GFLOPS\/W<\/td>\n<\/tr>\n AMD Instinct MI355X<\/td>\n 1400 W<\/td>\n \u7d04 10 PFLOPS<\/td>\n \u7d04 7 GFLOPS\/W<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n \u4f60\u6703\u767c\u73fe\uff0cAMD \u5728\u5cf0\u503c\u6548\u80fd\u4e0a\u5360\u512a\uff0c\u800c NVIDIA \u5728\u55ae\u4f4d\u529f\u8017\u6548\u7387\u4e0a\u66f4\u5177\u512a\u52e2\u3002\u9078\u64c7\u6642\uff0c\u4f60\u61c9\u7d9c\u5408\u8003\u91cf\u6a5f\u623f\u7684\u96fb\u529b\u8207\u6563\u71b1\u689d\u4ef6\u3002<\/p>\n \u6210\u672c\u8207\u50f9\u503c<\/h3>\n \u5728\u6295\u8cc7 AI \u786c\u9ad4\u6642\uff0c\u4f60\u81ea\u7136\u6703\u95dc\u6ce8\u6574\u9ad4\u50f9\u503c\u3002AMD \u900f\u904e\u5177\u7af6\u722d\u529b\u7684\u5b9a\u50f9\u8207\u5ba2\u88fd\u5316\u65b9\u6848\u6311\u6230 NVIDIA \u7684\u5e02\u5834\u5730\u4f4d\u3002\u4f60\u53ef\u4ee5\u9078\u64c7\u6700\u5951\u5408\u81ea\u8eab\u5de5\u4f5c\u8ca0\u8f09\u8207\u9810\u7b97\u7684 MI400 \u7cfb\u5217\u52a0\u901f\u5361\u3002AMD \u7684\u958b\u653e\u8edf\u9ad4\u57fa\u790e\u907f\u514d\u4e86\u904e\u5ea6\u7d81\u5b9a\u55ae\u4e00\u5ee0\u5546\uff0c\u5e6b\u52a9\u4f60\u5728\u4e0d\u70ba\u5c01\u9589\u6280\u8853\u4ed8\u51fa\u6ea2\u50f9\u7684\u524d\u63d0\u4e0b\uff0c\u4eab\u53d7\u5148\u9032\u529f\u80fd\u3002<\/p>\n \ud83d\udca1 \u63d0\u793a\uff1a\u5c07 AMD Instinct GPU \u8207 AMD EPYC \u8655\u7406\u5668\u642d\u914d\uff0c\u4e26\u63a1\u7528\u958b\u653e\u7522\u696d\u6a19\u6e96\uff0c\u53ef\u4ee5\u9032\u4e00\u6b65\u63d0\u5347\u6574\u9ad4\u6027\u50f9\u6bd4\u3002<\/p><\/blockquote>\n \u751f\u614b\u7cfb\u7d71\u8207\u76f8\u5bb9\u6027<\/h3>\n \u4f60\u9700\u8981\u4e00\u500b\u53ef\u9760\u7684\u751f\u614b\u7cfb\u7d71\u4f86\u652f\u6490 AI \u5c08\u6848\u3002AMD \u900f\u904e ROCm.ai \u6253\u9020\u4e86\u5805\u5be6\u7684\u8edf\u9ad4\u57fa\u790e\uff0c\u91cd\u9ede\u5f37\u5316\u958b\u767c\u8005\u5de5\u5177\u8207\u7b2c\u4e09\u65b9\u652f\u63f4\u3002\u4f60\u5c07\u5f9e\u8207 OpenAI\u3001Anthropic\u3001Meta \u548c Microsoft \u7b49\u5408\u4f5c\u5925\u4f34\u7684\u5354\u4f5c\u4e2d\u53d7\u76ca\u3002ROCm.ai \u81f4\u529b\u65bc\u63d0\u4f9b\u539f\u751f AI \u958b\u767c\u9ad4\u9a57\uff0c\u4ee5\u53ca\u7d71\u4e00\u7684\u8edf\u9ad4\u5e73\u53f0\u4f86\u652f\u6490\u667a\u6167\u90e8\u7f72\u3002<\/p>\n \n MI400 \u7cfb\u5217\u4f9d\u8a17 AMD ROCm \u8edf\u9ad4\u5806\u758a\u670d\u52d9 AI \u8207 HPC \u60c5\u5883\u3002<\/li>\n \u4f60\u53ef\u4ee5\u5728\u4e0d\u540c\u786c\u9ad4\u8207\u8edf\u9ad4\u7cfb\u7d71\u9593\u7372\u5f97\u826f\u597d\u7684\u4e92\u901a\u6027\u8207\u53ef\u651c\u6027\u3002<\/li>\n AMD Helios \u8a2d\u8a08\u5c07 Instinct GPU \u8207 EPYC CPU \u53ca Pensando \u7db2\u8def\u6df1\u5ea6\u6574\u5408\u3002<\/li>\n \u4f60\u53ef\u4ee5\u78ba\u4fdd\u8207\u73fe\u6709\u8cc7\u6599\u4e2d\u5fc3\u67b6\u69cb\u53ca\u958b\u653e\u7522\u696d\u6a19\u6e96\u7684\u76f8\u5bb9\u6027\u3002<\/li>\n<\/ul>\n \u4f60\u53ef\u4ee5\u5728\u4fdd\u6301\u5f48\u6027\u64f4\u5c55\u7684\u540c\u6642\uff0c\u9069\u61c9\u5feb\u901f\u6f14\u9032\u7684 AI \u9700\u6c42\u3002\u96a8\u8457 AMD \u4e0d\u65b7\u52a0\u5927\u5728 AI \u9818\u57df\u7684\u6295\u5165\uff0c\u5176\u751f\u614b\u7cfb\u7d71\u4e5f\u5c07\u6301\u7e8c\u58ef\u5927\uff0c\u9032\u4e00\u6b65\u7c21\u5316\u5148\u9032 AI \u5de5\u4f5c\u8ca0\u8f09\u7684\u90e8\u7f72\u8207\u71df\u904b\u3002<\/p>\n MI400 \u7684\u4f7f\u7528\u60c5\u5883\u8207\u90e8\u7f72<\/strong><\/h2>\n \u7522\u696d\u61c9\u7528<\/h3>\n \u4f60\u6703\u770b\u5230 MI400 \u7cfb\u5217\u5728\u773e\u591a\u7522\u696d\u9818\u5148\u5c08\u6848\u4e2d\u767c\u63ee\u91cd\u8981\u4f5c\u7528\u3002\u8a31\u591a\u4f01\u696d\u5229\u7528\u9019\u4e9b GPU \u89e3\u6c7a\u8907\u96dc\u554f\u984c\u4e26\u63d0\u5347\u6548\u7387\u3002\u4f8b\u5982\uff0cAT&T \u4f7f\u7528 MI355X GPU \u9032\u884c\u96fb\u4fe1 AI \u8a13\u7df4\uff0c\u900f\u904e\u5728\u55ae\u5361\u4e0a\u904b\u884c\u5b8c\u6574\u6a21\u578b\uff0c\u5be6\u73fe\u4e86 94% \u7684\u6548\u7387\uff1b\u7d22\u8afe\u62c9\u5927\u5b78\u7684 Yuca \u8d85\u7d1a\u96fb\u8166\u662f\u58a8\u897f\u54e5\u9802\u5c16\u7814\u7a76\u53e2\u96c6\uff0c\u7531 MI400 \u7cfb\u5217\u9a45\u52d5\uff1bTensorWave \u6b63\u5728\u5efa\u69cb\u57fa\u65bc MI400 \u7684 GPU \u96f2\uff0c\u53ef\u63d0\u4f9b\u9ad8\u9054\u5169\u500d\u6548\u80fd\u4e26\u964d\u4f4e\u6210\u672c\uff1bMaincode \u5728\u6fb3\u6d32\u6295\u8cc7 3000 \u842c\u7f8e\u5143\u5efa\u7f6e AI \u5de5\u5ee0\uff0c\u63a1\u7528 MI355X GPU \u6253\u9020\u9ad8\u6027\u50f9\u6bd4\u7cfb\u7d71\u3002<\/p>\n \n\n\n\n\n<\/colgroup>\n\n\n\n\n\n\n \u516c\u53f8<\/th>\n \u61c9\u7528\u985e\u578b<\/th>\n \u8aaa\u660e<\/th>\n<\/tr>\n AT&T<\/td>\n AI \u8a13\u7df4<\/td>\n \u4f7f\u7528 MI355X GPU \u9032\u884c\u96fb\u4fe1 AI \u8a13\u7df4\uff0c\u5728\u55ae\u5361\u4e0a\u5be6\u73fe 94% \u6548\u7387\u3002<\/td>\n<\/tr>\n UNISON<\/td>\n \u8d85\u7d1a\u904b\u7b97<\/td>\n Yuca \u8d85\u7d1a\u96fb\u8166\u662f\u58a8\u897f\u54e5\u9802\u7d1a\u7814\u7a76\u53e2\u96c6\uff0c\u7531 AMD \u63d0\u4f9b\u904b\u7b97\u652f\u63f4\u3002<\/td>\n<\/tr>\n TensorWave<\/td>\n AI \u96f2<\/td>\n \u57fa\u65bc MI400 GPU \u63d0\u4f9b\u6700\u591a 2 \u500d\u6548\u80fd\u63d0\u5347\u8207\u6210\u672c\u7bc0\u7701\u3002<\/td>\n<\/tr>\n Maincode<\/td>\n AI \u5de5\u5ee0<\/td>\n \u5728\u6fb3\u6d32\u5efa\u7f6e 3000 \u842c\u7f8e\u5143\u898f\u6a21 AI \u5de5\u5ee0\uff0c\u63a1\u7528 MI355X GPU\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n \u5be6\u969b\u90e8\u7f72\u60c5\u5883<\/h3>\n \u4f60\u53ef\u4ee5\u5728\u591a\u7a2e\u4f01\u696d\u8207\u7814\u7a76\u60c5\u5883\u4e2d\u90e8\u7f72 MI400 \u7cfb\u5217 GPU\u3002\u5728\u672c\u5730\u4f01\u696d\u7d1a AI \u90e8\u7f72\u4e2d\uff0c\u53ef\u4f7f\u7528 MI440X \u578b\u865f\u4f86\u652f\u6490\u53ef\u64f4\u5c55\u7684\u8a13\u7df4\u3001\u5fae\u8abf\u8207\u63a8\u7406\u5de5\u4f5c\u8ca0\u8f09\uff1bAI \u5de5\u5ee0\u7d1a\u8d85\u7d1a\u96fb\u8166\u5247\u5f80\u5f80\u63a1\u7528 MI430X\uff0c\u7528\u65bc\u9ad8\u6548\u80fd\u904b\u7b97\u8207\u79d1\u5b78\u61c9\u7528\u3002<\/p>\n \n\n\n\n\n<\/colgroup>\n\n\n\n\n GPU \u578b\u865f<\/th>\n \u90e8\u7f72\u60c5\u5883<\/th>\n \u61c9\u7528\u9818\u57df<\/th>\n<\/tr>\n MI440X<\/td>\n \u672c\u5730\u4f01\u696d\u7d1a AI<\/td>\n \u53ef\u64f4\u5c55\u8a13\u7df4\u3001\u5fae\u8abf\u8207\u63a8\u7406<\/td>\n<\/tr>\n MI430X<\/td>\n AI \u5de5\u5ee0\u7d1a\u8d85\u7d1a\u96fb\u8166<\/td>\n HPC \u8207\u79d1\u5b78\u904b\u7b97\u61c9\u7528<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n \ud83d\udca1 \u63d0\u793a\uff1a\u85c9\u52a9 MI400 \u7cfb\u5217\uff0c\u4f60\u53ef\u4ee5\u6309\u9700\u64f4\u5c55\u57fa\u790e\u8a2d\u65bd\uff0c\u4ee5\u6eff\u8db3\u7814\u7a76\u8207\u5546\u696d\u9818\u57df\u65e5\u76ca\u6210\u9577\u7684 AI \u9700\u6c42\u3002<\/p><\/blockquote>\n \u4e3b\u6b0a AI \u8207\u5927\u578b\u8a9e\u8a00\u6a21\u578b<\/h3>\n \u900f\u904e MI400 \u7cfb\u5217\uff0c\u4f60\u53ef\u4ee5\u66f4\u597d\u5730\u652f\u63f4\u4e3b\u6b0a AI \u8a08\u756b\u8207\u5927\u578b\u8a9e\u8a00\u6a21\u578b\u90e8\u7f72\u3002\u9ad8\u8a18\u61b6\u9ad4\u5bb9\u91cf\u8207\u9ad8\u901f\u904b\u7b97\u80fd\u529b\u53ef\u4ee5\u9ad8\u6548\u652f\u6490 LLM \u7684\u8a13\u7df4\u548c\u63a8\u7406\u3002\u4f60\u53ef\u4ee5\u70ba\u570b\u5bb6\u7d1a\u8207\u4f01\u696d\u7d1a\u60c5\u5883\u5efa\u69cb\u5b89\u5168\u7684\u4e3b\u6b0a AI \u7cfb\u7d71\uff0c\u5728\u8655\u7406\u654f\u611f\u8cc7\u6599\u548c\u8907\u96dc\u6a21\u578b\u6642\u7dad\u6301\u9ad8\u6548\u80fd\u8207\u9ad8\u53ef\u9760\u6027\u3002<\/p>\n \u6ce8\u610f\u4e8b\u9805\u8207\u9650\u5236<\/strong><\/h2>\n \u76f8\u5bb9\u6027\u8207\u751f\u614b\u7cfb<\/h3>\n \u5728\u70ba\u5c08\u6848\u5c0e\u5165\u65b0\u786c\u9ad4\u6642\uff0c\u4f60\u53ef\u80fd\u6703\u9762\u81e8\u4e00\u4e9b\u6311\u6230\u3002\u8f49\u5411 AMD Instinct MI400 \u7cfb\u5217\u6642\uff0c\u5c24\u5176\u662f\u5718\u968a\u9577\u671f\u4f7f\u7528\u5176\u4ed6\u5e73\u53f0\u7684\u60c5\u6cc1\u4e0b\uff0c\u76f8\u5bb9\u6027\u554f\u984c\u9700\u8981\u4e8b\u5148\u8003\u91cf\u3002<\/p>\n \n \u5982\u679c\u4f60\u5f9e NVIDIA CUDA \u9077\u79fb\u81f3 AMD ROCm\uff0c\u8edf\u9ad4\u6574\u5408\u904e\u7a0b\u53ef\u80fd\u6703\u66f4\u8907\u96dc\u3002<\/li>\n \u4f60\u7684\u958b\u767c\u4eba\u54e1\u9700\u8981\u6642\u9593\u719f\u6089 ROCm\uff0c\u9019\u5728\u77ed\u671f\u5167\u53ef\u80fd\u653e\u7de9\u6574\u9ad4\u958b\u767c\u7bc0\u594f\u3002<\/li>\n \u5c0d\u65bc\u8ffd\u6c42\u9ad8\u5ea6\u7a69\u5b9a\u8207\u53ef\u9810\u671f\u6027\u7684\u4f01\u696d\u5718\u968a\u800c\u8a00\uff0c\u788e\u7247\u5316\u7684\u8edf\u9ad4\u5806\u758a\u53ef\u80fd\u5e36\u4f86\u984d\u5916\u7ba1\u7406\u8ca0\u64d4\u3002<\/li>\n<\/ul>\n \u4f60\u61c9\u4e8b\u5148\u898f\u5283\u8a13\u7df4\u8207\u6280\u8853\u652f\u63f4\uff0c\u4ee5\u5354\u52a9\u5718\u968a\u9806\u5229\u904e\u6e21\uff0c\u5f9e\u800c\u6700\u5927\u5316\u65b0\u786c\u9ad4\u5e36\u4f86\u7684\u6548\u76ca\u3002<\/p>\n \u6311\u6230\u8207\u4e0d\u8db3<\/h3>\n \u4f60\u5e0c\u671b\u7cfb\u7d71\u9577\u671f\u4fdd\u6301\u7a69\u5b9a\u9ad8\u6548\uff0c\u4f46\u5728\u4f7f\u7528 MI400 \u7cfb\u5217\u6642\u4ecd\u53ef\u80fd\u9047\u5230\u4e00\u4e9b\u4e0d\u8db3\u4e4b\u8655\u3002\u96d6\u7136 ROCm \u751f\u614b\u7cfb\u767c\u5c55\u8fc5\u901f\uff0c\u4f46\u76f8\u8f03\u65bc\u66f4\u65e9\u671f\u7684\u5e73\u53f0\u4ecd\u7565\u986f\u5e74\u8f15\u3002\u4f60\u53ef\u80fd\u6703\u767c\u73fe\u7b2c\u4e09\u65b9\u5de5\u5177\u6578\u91cf\u8f03\u5c11\u3001\u793e\u7fa4\u652f\u63f4\u76f8\u5c0d\u6709\u9650\uff0c\u9019\u6703\u589e\u52a0\u7591\u96e3\u6392\u89e3\u96e3\u5ea6\u3002\u67d0\u4e9b\u65b0\u8208 AI \u6846\u67b6\u5728\u6700\u4f73\u5316\u6216\u76f8\u5bb9\u4e0a\u7684\u66f4\u65b0\u7bc0\u594f\u4e5f\u53ef\u80fd\u7565\u986f\u6eef\u5f8c\u3002\u5982\u679c\u4f60\u7684\u5de5\u4f5c\u6d41\u7a0b\u9ad8\u5ea6\u4f9d\u8cf4\u7279\u5b9a\u8edf\u9ad4\uff0c\u5efa\u8b70\u5728\u5347\u7d1a\u524d\u9810\u5148\u78ba\u8a8d\u76f8\u5bb9\u6027\u3002<\/p>\n \u63d0\u793a\uff1a<\/strong> \u5728\u5168\u9762\u90e8\u7f72\u524d\uff0c\u5148\u5728 AMD \u786c\u9ad4\u4e0a\u5c0d\u95dc\u9375\u5de5\u4f5c\u8ca0\u8f09\u9032\u884c\u6e2c\u8a66\uff0c\u53ef\u4ee5\u6709\u6548\u964d\u4f4e\u98a8\u96aa\u3002<\/p><\/blockquote>\n \u672a\u4f86\u5c55\u671b<\/h3>\n \u4f60\u53ef\u4ee5\u671f\u5f85\u4e0b\u4e00\u4ee3 AMD Instinct GPU \u5e36\u4f86\u91cd\u5927\u98db\u8e8d\u3002\u5f9e\u7522\u54c1\u85cd\u5716\u4f86\u770b\uff0c\u5728\u6548\u80fd\u8207\u8a18\u61b6\u9ad4\u65b9\u9762\u90fd\u5c07\u6709\u986f\u8457\u63d0\u5347\u3002\u672a\u4f86\u53ef\u80fd\u5be6\u73fe\u7684\u76ee\u6a19\u5305\u62ec\uff1a<\/p>\n \n\n\n\n<\/colgroup>\n\n\n\n\n\n\n\n\n\n\n\n\n \u7279\u6027<\/th>\n \u898f\u683c<\/th>\n<\/tr>\n \u76ee\u6a19\u767c\u4f48\u6642\u9593<\/td>\n 2026 \u5e74<\/td>\n<\/tr>\n GPU \u6548\u80fd (FP4)<\/td>\n 40 PetaFLOPS<\/td>\n<\/tr>\n GPU \u6548\u80fd (FP8)<\/td>\n 20 PetaFLOPS<\/td>\n<\/tr>\n HBM \u8a18\u61b6\u9ad4<\/td>\n 432 GB<\/td>\n<\/tr>\n \u8a18\u61b6\u9ad4\u983b\u5bec<\/td>\n 20 TB\/s<\/td>\n<\/tr>\n \u6a6b\u5411\u64f4\u5c55\u983b\u5bec<\/td>\n 300 Gigabits per Second<\/td>\n<\/tr>\n Helios \u6548\u80fd\u5c55\u671b<\/td>\n \u76f8\u8f03 MI355\uff0cAI \u6548\u80fd\u6700\u9ad8\u53ef\u63d0\u5347\u81f3 10 \u500d<\/td>\n<\/tr>\n HBM \u5bb9\u91cf\u6bd4\u8f03<\/td>\n \u6bd4 MI355 \u63d0\u5347 50%<\/td>\n<\/tr>\n HBM \u8a18\u61b6\u9ad4\u983b\u5bec\u6bd4\u8f03<\/td>\n \u6bd4 MI355 \u63d0\u5347 50%<\/td>\n<\/tr>\n \u6a6b\u5411\u64f4\u5c55\u983b\u5bec\u6bd4\u8f03<\/td>\n \u6bd4 MI355 \u6700\u9ad8\u63d0\u5347 50%<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n \u672a\u4f86\u4f60\u5c07\u5728 AI \u5de5\u4f5c\u8ca0\u8f09\u4e2d\u7372\u5f97\u66f4\u5f37\u904b\u7b97\u80fd\u529b\u8207\u66f4\u9ad8\u901f\u5ea6\uff0c\u9019\u4e9b\u5347\u7d1a\u6709\u52a9\u65bc\u4f60\u61c9\u5c0d\u66f4\u5927\u898f\u6a21\u7684\u6a21\u578b\u8207\u66f4\u8907\u96dc\u7684\u4efb\u52d9\u3002\u96a8\u8457 AMD \u6301\u7e8c\u6295\u5165\uff0c\u4f60\u4e5f\u53ef\u4ee5\u671f\u5f85\u66f4\u6210\u719f\u7684\u751f\u614b\u7cfb\u8207\u66f4\u5b8c\u5584\u7684\u652f\u63f4\u9ad4\u7cfb\u3002<\/p>\n \n <\/div>\n<\/div>\n \u5728\u61c9\u5c0d\u9ad8\u5f37\u5ea6 AI \u6216 HPC \u5de5\u4f5c\u8ca0\u8f09\u6642\uff0cAMD Instinct MI400 \u7cfb\u5217\u70ba\u4f60\u63d0\u4f9b\u4e86\u6975\u5177\u50f9\u503c\u7684\u9078\u64c7\u3002\u9019\u4e9b GPU \u5177\u5099\u9ad8\u8a18\u61b6\u9ad4\u5bb9\u91cf\u3001\u5148\u9032\u67b6\u69cb\u4ee5\u53ca\u9748\u6d3b\u7684\u90e8\u7f72\u6a21\u5f0f\u3002\u4f46\u5728\u505a\u51fa\u6c7a\u7b56\u524d\uff0c\u4f60\u4ecd\u61c9\u7d9c\u5408\u8003\u91cf\u8edf\u9ad4\u9700\u6c42\u8207\u751f\u614b\u6210\u719f\u5ea6\u3002\u5ba2\u88fd\u5316 MI400 \u52a0\u901f\u5668\u53ef\u4ee5\u5e6b\u52a9\u4f60\u5728\u7279\u5b9a\u60c5\u5883\u4e0b\u5be6\u73fe\u76ee\u6a19\u6548\u80fd\uff0c\u7279\u5225\u9069\u5408\u5177\u6709\u5c08\u9580\u5316\u9700\u6c42\u7684\u74b0\u5883\u3002<\/p>\n \u5e38\u898b\u554f\u984c FAQ<\/strong><\/h2>\n AMD Instinct MI400 GPU \u652f\u63f4\u54ea\u4e9b AI \u6846\u67b6\uff1f<\/h3>\n \u4f60\u53ef\u4ee5\u4f7f\u7528 PyTorch\u3001TensorFlow \u548c JAX \u7b49\u4e3b\u6d41\u6846\u67b6\u3002ROCm \u5e73\u53f0\u5c0d\u9019\u4e9b\u5de5\u5177\u63d0\u4f9b\u652f\u63f4\uff0c\u56e0\u6b64\u7121\u9808\u4fee\u6539\u7a0b\u5f0f\u78bc\u5373\u53ef\u9077\u79fb AI \u5c08\u6848\uff0c\u5728\u5207\u63db\u5230 AMD \u786c\u9ad4\u6642\u7372\u5f97\u9806\u66a2\u9ad4\u9a57\u3002<\/p>\n \u5982\u4f55\u70ba\u5927\u578b AI \u6a21\u578b\u64f4\u5c55 AMD Instinct MI400 GPU\uff1f<\/h3>\n \u4f60\u53ef\u4ee5\u5c07\u591a\u584a MI400 GPU \u5efa\u69cb\u70ba\u53e2\u96c6\u3002\u5176\u67b6\u69cb\u652f\u63f4\u9ad8\u8a18\u61b6\u9ad4\u5bb9\u91cf\u8207\u9ad8\u901f\u4e92\u9023\uff0c\u5e6b\u52a9\u4f60\u5728\u8a13\u7df4\u5927\u578b\u8a9e\u8a00\u6a21\u578b\u53ca\u8655\u7406\u6d77\u91cf\u8cc7\u6599\u96c6\u6642\u907f\u514d\u74f6\u9838\u3002<\/p>\n AMD Instinct MI400 GPU \u662f\u5426\u76f8\u5bb9\u73fe\u6709\u8cc7\u6599\u4e2d\u5fc3\uff1f<\/h3>\n \u662f\u7684\uff0c\u4f60\u53ef\u4ee5\u5728\u6a19\u6e96\u8cc7\u6599\u4e2d\u5fc3\u6a5f\u67b6\u4e2d\u90e8\u7f72 MI400 GPU\u3002Helios \u6a5f\u67b6\u7d1a\u7cfb\u7d71\u63a1\u7528\u958b\u653e\u7522\u696d\u6a19\u6e96\uff0c\u4f60\u53ef\u4ee5\u5c07\u9019\u4e9b GPU \u8207 AMD EPYC \u8655\u7406\u5668\u53ca Pensando \u7db2\u8def\u89e3\u6c7a\u65b9\u6848\u6574\u5408\uff0c\u5efa\u69cb\u9748\u6d3b\u7684 AI \u57fa\u790e\u8a2d\u65bd\u3002<\/p>\n AMD Instinct MI400 \u7cfb\u5217\u8207 NVIDIA GPU \u6709\u4f55\u4e0d\u540c\uff1f<\/h3>\n AMD \u5728\u9ad8\u8a18\u61b6\u9ad4\u5bb9\u91cf\u8207\u958b\u653e\u8edf\u9ad4\u751f\u614b\u65b9\u9762\u5177\u6709\u512a\u52e2\u3002MI400 \u7cfb\u5217\u63d0\u4f9b\u5ba2\u88fd\u5316\u65b9\u6848\u8207\u5177\u7af6\u722d\u529b\u7684\u50f9\u683c\uff0c\u4f60\u53ef\u4ee5\u6839\u64da\u5be6\u969b AI \u5de5\u4f5c\u8ca0\u8f09\u81ea\u7531\u9078\u64c7\u786c\u9ad4\uff0c\u6e1b\u5c11\u88ab\u55ae\u4e00\u5ee0\u5546\u7d81\u5b9a\u7684\u98a8\u96aa\u3002<\/p>\n AMD Instinct MI400 GPU \u662f\u5426\u9069\u7528\u65bc\u8a13\u7df4\u548c\u63a8\u7406\uff1f<\/h3>\n \u662f\u7684\uff0c\u4f60\u53ef\u4ee5\u540c\u6642\u5c07 MI400 GPU \u7528\u65bc AI \u8a13\u7df4\u8207\u63a8\u7406\u3002\u5f37\u5927\u7684\u904b\u7b97\u6548\u80fd\u8207\u8a18\u61b6\u9ad4\u983b\u5bec\u53ef\u4ee5\u5e6b\u52a9\u4f60\u8a13\u7df4\u5927\u898f\u6a21\u6a21\u578b\uff0c\u4e26\u90e8\u7f72\u56de\u61c9\u5feb\u901f\u7684 AI \u61c9\u7528\u3002<\/p>\n","protected":false},"excerpt":{"rendered":" \u7576\u4f60\u5728\u9ad8\u8981\u6c42\u7684 \u7f8e\u570b\u4f3a\u670d\u5668\u79df\u7528 \u74b0\u5883\u4e0b\u904b\u884c AI \u548c HPC \u5de5\u4f5c\u8ca0\u8f09 \u6642\uff0c\u4f60\u9700\u8981\u517c\u5177\u9ad8\u6548\u80fd\u8207\u9ad8\u6548\u7387\u7684\u65b9\u6848\u3002A […]<\/p>\n Read More…<\/a><\/p>\n","protected":false},"author":2,"featured_media":33708,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[4815],"tags":[12496,12497,12498,1729,525],"class_list":["post-33711","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-america-dedicated-server-tc","tag-amd-instinct-mi400","tag-ai-gpus","tag-data-center-ai-workload","tag-high-performance-computing-tc","tag-dedicated-server-hosting-tc"],"acf":[],"yoast_head":"\n