{"id":1825,"date":"2023-10-08T10:09:46","date_gmt":"2023-10-08T02:09:46","guid":{"rendered":"http:\/\/shlj-tec.com\/?p=1825"},"modified":"2023-10-08T10:09:46","modified_gmt":"2023-10-08T02:09:46","slug":"nvidia-h800-pcie-gpu","status":"publish","type":"post","link":"http:\/\/shlj-tec.com\/index.php\/2023\/10\/08\/nvidia-h800-pcie-gpu\/","title":{"rendered":"NVIDIA H800 PCIe GPU"},"content":{"rendered":"

 <\/p>\n

\"\"<\/p>\n

\u901a\u8fc7 NVIDIA H800 Tensor Core GPU\uff0c\u5728\u6bcf\u4e2a\u5de5\u4f5c\u8d1f\u8f7d\u4e2d\u5b9e\u73b0\u51fa\u8272\u6027\u80fd\u3001\u53ef\u6269\u5c55\u6027\u548c\u5b89\u5168\u6027\u3002\u4f7f\u7528 NVIDIA\u00ae NVLink\u00ae Switch \u7cfb\u7edf\uff0c\u53ef\u8fde\u63a5\u591a\u8fbe 256 \u4e2a H800 \u6765\u52a0\u901f\u767e\u4ebf\u4ebf\u7ea7 (Exascale) \u5de5\u4f5c\u8d1f\u8f7d\uff0c\u53e6\u5916\u53ef\u901a\u8fc7\u4e13\u7528\u7684 Transformer \u5f15\u64ce\u6765\u5904\u7406\u4e07\u4ebf\u53c2\u6570\u8bed\u8a00\u6a21\u578b\u3002\u4e0e\u4e0a\u4e00\u4ee3\u4ea7\u54c1\u76f8\u6bd4\uff0cH800 \u7684\u7efc\u5408\u6280\u672f\u521b\u65b0\u53ef\u4ee5\u5c06\u5927\u578b\u8bed\u8a00\u6a21\u578b\u7684\u901f\u5ea6\u63d0\u9ad8 30 \u500d\uff0c\u4ece\u800c\u63d0\u4f9b\u4e1a\u754c\u9886\u5148\u7684\u5bf9\u8bdd\u5f0f AI\u3002<\/p>\n\n\n\n\n\n\n\n\n
\u578b\u53f7\u540d<\/td>\nNVIDIA H800 80G PCIe<\/td>\nGPU\u663e\u5b58<\/td>\n80GB HBM2e<\/td>\n<\/tr>\n
\u663e\u5b58\u5e26\u5bbd<\/td>\n2TB\/s<\/td>\nCUDA\u6838\u5fc3<\/td>\n14592<\/td>\n<\/tr>\n
\u786c\u4ef6\u63a5\u53e3<\/td>\nPCIe 5.0 x16<\/td>\n\u5149\u7ebf\u8ffd\u8e2a<\/td>\n\u5426<\/td>\n<\/tr>\n
\u82af\u7247\u67b6\u6784<\/td>\nHopper<\/td>\nNVLINK\u652f\u6301<\/td>\n\u662f\uff0c\u4f7f\u7528\u6865\u63a5\u5668<\/td>\n<\/tr>\n
vGPU\u652f\u6301<\/td>\n\u662f<\/td>\nFP16 Tensor Core<\/td>\n756.5 TFLOPS<\/td>\n<\/tr>\n
\u591a\u5b9e\u4f8bGPU\uff08MIG\uff09<\/td>\n\u6700\u591a7\u4e2aMIG\uff0c\u6bcf\u4e2a10GB<\/td>\n\u529f\u8017<\/td>\n300-350w<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

\u4ea7\u54c1\u53c2\u6570<\/b><\/strong><\/p>\n

\u4ea7\u54c1\u8be6\u60c5<\/b><\/strong><\/p>\n

\u901a\u8fc7 NVIDIA H800 Tensor Core GPU\uff0c\u5728\u6bcf\u4e2a\u5de5\u4f5c\u8d1f\u8f7d\u4e2d\u5b9e\u73b0\u51fa\u8272\u6027\u80fd\u3001\u53ef\u6269\u5c55\u6027\u548c\u5b89\u5168\u6027\u3002\u4f7f\u7528 NVIDIA\u00ae NVLink\u00ae Switch \u7cfb\u7edf\uff0c\u53ef\u8fde\u63a5\u591a\u8fbe 256 \u4e2a H800 \u6765\u52a0\u901f\u767e\u4ebf\u4ebf\u7ea7 (Exascale) \u5de5\u4f5c\u8d1f\u8f7d\uff0c\u53e6\u5916\u53ef\u901a\u8fc7\u4e13\u7528\u7684 Transformer \u5f15\u64ce\u6765\u5904\u7406\u4e07\u4ebf\u53c2\u6570\u8bed\u8a00\u6a21\u578b\u3002\u4e0e\u4e0a\u4e00\u4ee3\u4ea7\u54c1\u76f8\u6bd4\uff0cH800 \u7684\u7efc\u5408\u6280\u672f\u521b\u65b0\u53ef\u4ee5\u5c06\u5927\u578b\u8bed\u8a00\u6a21\u578b\u7684\u901f\u5ea6\u63d0\u9ad8 30 \u500d\uff0c\u4ece\u800c\u63d0\u4f9b\u4e1a\u754c\u9886\u5148\u7684\u5bf9\u8bdd\u5f0f AI\u3002<\/p>\n

 <\/p>\n

\u53d8\u9769 AI \u8bad\u7ec3<\/strong><\/p>\n

H800 \u914d\u5907\u7b2c\u56db\u4ee3 Tensor Core \u548c Transformer \u5f15\u64ce\uff08FP8 \u7cbe\u5ea6\uff09\uff0c\u4e0e\u4e0a\u4e00\u4ee3\u4ea7\u54c1\u76f8\u6bd4\uff0c\u53ef\u4e3a\u591a\u4e13\u5bb6 (MoE) \u6a21\u578b\u63d0\u4f9b\u9ad8 9 \u500d\u7684\u8bad\u7ec3\u901f\u5ea6\u3002\u901a\u8fc7\u7ed3\u5408\u53ef\u63d0\u4f9b 400 GB\/s GPU \u95f4\u4e92\u8fde\u7684\u7b2c\u56db\u4ee3 NVlink\u3001\u53ef\u8de8\u8282\u70b9\u52a0\u901f\u6bcf\u4e2a GPU \u901a\u4fe1\u7684 NVLINK Switch \u7cfb\u7edf\u3001PCIe 5.0 \u4ee5\u53ca NVIDIA Magnum IO\u2122 \u8f6f\u4ef6\uff0c\u4e3a\u5c0f\u578b\u4f01\u4e1a\u5230\u5927\u89c4\u6a21\u7edf\u4e00 GPU \u96c6\u7fa4\u63d0\u4f9b\u9ad8\u6548\u7684\u53ef\u6269\u5c55\u6027\u3002<\/p>\n

\u5728\u6570\u636e\u4e2d\u5fc3\u7ea7\u90e8\u7f72 H800 GPU \u53ef\u63d0\u4f9b\u51fa\u8272\u7684\u6027\u80fd\uff0c\u5e76\u4f7f\u6240\u6709\u7814\u7a76\u4eba\u5458\u5747\u80fd\u8f7b\u677e\u4f7f\u7528\u65b0\u4e00\u4ee3\u767e\u4ebf\u4ebf\u6b21\u7ea7 (Exascale) \u9ad8\u6027\u80fd\u8ba1\u7b97 (HPC) \u548c\u4e07\u4ebf\u53c2\u6570\u7684 AI\u3002<\/p>\n

 <\/p>\n

\u5b9e\u65f6\u6df1\u5ea6\u5b66\u4e60\u63a8\u7406<\/strong><\/p>\n

AI \u6b63\u5728\u5229\u7528\u4e00\u7cfb\u5217\u5e7f\u6cdb\u7684\u795e\u7ecf\u7f51\u7edc\u89e3\u51b3\u8303\u56f4\u540c\u6837\u5e7f\u6cdb\u7684\u4e00\u7cfb\u5217\u5546\u4e1a\u6311\u6218\u3002\u51fa\u8272\u7684 AI \u63a8\u7406\u52a0\u901f\u5668\u4e0d\u4ec5\u8981\u63d0\u4f9b\u975e\u51e1\u6027\u80fd\uff0c\u8fd8\u8981\u5229\u7528\u901a\u7528\u6027\u52a0\u901f\u8fd9\u4e9b\u795e\u7ecf\u7f51\u7edc\u3002<\/p>\n

H800 \u8fdb\u4e00\u6b65\u6269\u5c55\u4e86 NVIDIA \u5728\u63a8\u7406\u9886\u57df\u7684\u5e02\u573a\u9886\u5148\u5730\u4f4d\uff0c\u5176\u591a\u9879\u5148\u8fdb\u6280\u672f\u53ef\u5c06\u63a8\u7406\u901f\u5ea6\u63d0\u9ad8 30 \u500d\uff0c\u5e76\u63d0\u4f9b\u8d85\u4f4e\u7684\u5ef6\u8fdf\u3002\u7b2c\u56db\u4ee3 Tensor Core \u53ef\u52a0\u901f\u6240\u6709\u7cbe\u5ea6\uff08\u5305\u62ec FP64\u3001TF32\u3001FP32\u3001FP16 \u548c INT8\uff09\u3002Transformer \u5f15\u64ce\u53ef\u7ed3\u5408\u4f7f\u7528 FP8 \u548c FP16 \u7cbe\u5ea6\uff0c\u51cf\u5c11\u5185\u5b58\u5360\u7528\u5e76\u63d0\u9ad8\u6027\u80fd\uff0c\u540c\u65f6\u4ecd\u80fd\u4fdd\u6301\u5927\u578b\u8bed\u8a00\u6a21\u578b\u7684\u51c6\u786e\u6027\u3002<\/p>\n

 <\/p>\n

\u767e\u4ebf\u4ebf\u6b21\u7ea7\u9ad8\u6027\u80fd\u8ba1\u7b97<\/strong><\/p>\n

NVIDIA \u6570\u636e\u4e2d\u5fc3\u5e73\u53f0\u6027\u80fd\u6301\u7eed\u63d0\u5347\uff0c\u8d85\u8d8a\u6469\u5c14\u5b9a\u5f8b\u3002H800\u00a0\u7684\u5168\u65b0\u7a81\u7834\u6027 AI \u6027\u80fd\u8fdb\u4e00\u6b65\u52a0\u5f3a\u4e86 HPC+AI \u7684\u529b\u91cf\uff0c\u52a0\u901f\u79d1\u5b66\u5bb6\u548c\u7814\u7a76\u4eba\u5458\u7684\u63a2\u7d22\uff0c\u8ba9\u4ed6\u4eec\u5168\u8eab\u5fc3\u6295\u5165\u5de5\u4f5c\uff0c\u89e3\u51b3\u4e16\u754c\u9762\u4e34\u7684\u91cd\u5927\u6311\u6218\u3002<\/p>\n

H800\u00a0\u8fd8\u91c7\u7528 DPX \u6307\u4ee4\uff0c\u5176\u6027\u80fd\u6bd4 NVIDIA A800 Tensor Core GPU \u9ad8 7 \u500d\uff0c\u5728\u52a8\u6001\u7f16\u7a0b\u7b97\u6cd5\uff08\u4f8b\u5982\uff0c\u7528\u4e8e DNA \u5e8f\u5217\u6bd4\u5bf9 Smith-Waterman\uff09\u4e0a\u6bd4\u4ec5\u4f7f\u7528\u4f20\u7edf\u53cc\u8def CPU \u7684\u670d\u52a1\u5668\u5feb 40 \u500d\u3002<\/p>\n

 <\/p>\n

\u52a0\u901f\u6570\u636e\u5206\u6790<\/strong><\/p>\n

\u5728 AI \u5e94\u7528\u5f00\u53d1\u8fc7\u7a0b\u4e2d\uff0c\u6570\u636e\u5206\u6790\u901a\u5e38\u4f1a\u6d88\u8017\u5927\u90e8\u5206\u65f6\u95f4\u3002\u539f\u56e0\u5728\u4e8e\uff0c\u5927\u578b\u6570\u636e\u96c6\u5206\u6563\u5728\u591a\u53f0\u670d\u52a1\u5668\u4e0a\uff0c\u7531\u4ec5\u914d\u5907\u5546\u7528 CPU \u670d\u52a1\u5668\u7ec4\u6210\u6a2a\u5411\u6269\u5c55\u5f0f\u7684\u89e3\u51b3\u65b9\u6848\u7f3a\u4e4f\u53ef\u6269\u5c55\u7684\u8ba1\u7b97\u6027\u80fd\uff0c\u4ece\u800c\u9677\u5165\u56f0\u5883\u3002<\/p>\n

\u642d\u8f7d H800 \u7684\u52a0\u901f\u670d\u52a1\u5668\u53ef\u4ee5\u63d0\u4f9b\u76f8\u5e94\u7684\u8ba1\u7b97\u80fd\u529b\uff0c\u5e76\u5229\u7528 NVLink \u548c NVSwitch \u6bcf\u4e2a GPU 3 TB\/s \u7684\u663e\u5b58\u5e26\u5bbd\u548c\u53ef\u6269\u5c55\u6027\uff0c\u51ed\u501f\u9ad8\u6027\u80fd\u5e94\u5bf9\u6570\u636e\u5206\u6790\u4ee5\u53ca\u901a\u8fc7\u6269\u5c55\u652f\u6301\u5e9e\u5927\u7684\u6570\u636e\u96c6\u3002\u901a\u8fc7\u7ed3\u5408\u4f7f\u7528 NVIDIA Quantum-2 InfiniBand\u3001Magnum IO \u8f6f\u4ef6\u3001GPU \u52a0\u901f\u7684 Spark 3.0 \u548c NVIDIA RAPIDS\u2122\uff0cNVIDIA \u6570\u636e\u4e2d\u5fc3\u5e73\u53f0\u80fd\u591f\u4ee5\u51fa\u8272\u7684\u6027\u80fd\u548c\u6548\u7387\u52a0\u901f\u8fd9\u4e9b\u5927\u578b\u5de5\u4f5c\u8d1f\u8f7d<\/p>\n

 <\/p>\n

\u5185\u7f6e\u673a\u5bc6\u8ba1\u7b97<\/strong><\/p>\n

\u5f53\u4eca\u7684\u673a\u5bc6\u8ba1\u7b97\u89e3\u51b3\u65b9\u6848\u57fa\u4e8e CPU\uff0c\u5bf9\u4e8e AI \u548c HPC \u7b49\u8ba1\u7b97\u5bc6\u96c6\u578b\u5de5\u4f5c\u8d1f\u8f7d\u6765\u8bf4\uff0c\u8fd9\u4e9b\u89e3\u51b3\u65b9\u6848\u8fdc\u8fdc\u65e0\u6cd5\u6ee1\u8db3\u9700\u6c42\u3002NVIDIA \u673a\u5bc6\u8ba1\u7b97\u662f NVIDIA Hopper\u2122 \u67b6\u6784\u7684\u5185\u7f6e\u5b89\u5168\u529f\u80fd\uff0c\u8be5\u529f\u80fd\u4f7f H800 \u6210\u4e3a NVIDIA \u7387\u5148\u63a8\u51fa\u7684\u5177\u6709\u673a\u5bc6\u8ba1\u7b97\u529f\u80fd\u7684\u52a0\u901f\u5668\u3002\u7528\u6237\u53ef\u4ee5\u5728\u83b7\u53d6 H800 GPU \u51fa\u8272\u52a0\u901f\u529f\u80fd\u7684\u540c\u65f6\uff0c\u4fdd\u62a4\u4f7f\u7528\u4e2d\u7684\u6570\u636e\u548c\u5e94\u7528\u7684\u673a\u5bc6\u6027\u548c\u5b8c\u6574\u6027\u3002\u5b83\u521b\u5efa\u4e86\u57fa\u4e8e\u786c\u4ef6\u7684\u53ef\u4fe1\u6267\u884c\u73af\u5883 (TEE)\uff0c\u7528\u4e8e\u4fdd\u62a4\u5e76\u9694\u79bb\u5728\u5355\u4e2a H800 GPU\u3001\u8282\u70b9\u5185\u591a\u4e2a H800 GPU \u6216\u5355\u4e2a MIG \u5b9e\u4f8b\u4e0a\u8fd0\u884c\u7684\u6574\u4e2a\u5de5\u4f5c\u8d1f\u8f7d\u3002\u5728 TEE \u5185\uff0cGPU \u52a0\u901f\u5e94\u7528\u7684\u8fd0\u884c\u53ef\u4ee5\u4fdd\u6301\u4e0d\u53d8\uff0c\u4e14\u4e0d\u5fc5\u5bf9\u5176\u8fdb\u884c\u5206\u533a\u3002\u7528\u6237\u53ef\u4ee5\u5c06\u9002\u7528\u4e8e AI \u548c HPC \u7684 NVIDIA \u8f6f\u4ef6\u7684\u5f3a\u5927\u529f\u80fd\u4e0e NVIDIA \u673a\u5bc6\u8ba1\u7b97\u63d0\u4f9b\u7684\u786c\u4ef6\u4fe1\u4efb\u6839\u7684\u5b89\u5168\u6027\u76f8\u7ed3\u5408\u3002<\/p>\n

 <\/p>\n

\u4e3a\u5927\u89c4\u6a21 AI \u548c\u9ad8\u6027\u80fd\u8ba1\u7b97\u63d0\u4f9b\u51fa\u8272\u7684\u6027\u80fd<\/strong><\/p>\n

Hopper Tensor Core GPU \u5c06\u4e3a NVIDIA Grace Hopper CPU+GPU \u67b6\u6784\u63d0\u4f9b\u652f\u6301\uff0c\u8be5\u67b6\u6784\u4e13\u4e3a TB \u7ea7\u52a0\u901f\u8ba1\u7b97\u800c\u6784\u5efa\uff0c\u53ef\u4e3a\u5927\u578b AI \u548c HPC \u63d0\u4f9b 10 \u500d\u7684\u6027\u80fd\u3002NVIDIA Grace CPU \u5229\u7528 Arm\u00ae \u67b6\u6784\u7684\u7075\u6d3b\u6027\u6765\u521b\u5efa CPU \u548c\u670d\u52a1\u5668\u67b6\u6784\uff0c\u8be5\u67b6\u6784\u662f\u4e13\u95e8\u9488\u5bf9\u52a0\u901f\u8ba1\u7b97\u800c\u4ece\u5934\u5f00\u59cb\u8bbe\u8ba1\u7684\u3002Hopper GPU \u4e0e Grace CPU \u642d\u914d\uff0c\u4f7f\u7528 NVIDIA \u8d85\u5feb\u901f\u7684\u82af\u7247\u95f4\u4e92\u8fde\u6280\u672f\uff0c\u53ef\u63d0\u4f9b 900GB\/s \u7684\u5e26\u5bbd\uff0c\u6bd4 PCIe 5.0 \u5feb 7 \u500d\u3002\u4e0e\u5f53\u4eca\u8fd0\u884c\u6700\u5feb\u7684\u670d\u52a1\u5668\u76f8\u6bd4\uff0c\u8fd9\u79cd\u521b\u65b0\u8bbe\u8ba1\u5c06 GPU \u7684\u805a\u5408\u7cfb\u7edf\u663e\u5b58\u5e26\u5bbd\u63d0\u9ad8 30 \u500d\uff0c\u5e76\u4e14\u4f1a\u5c06\u8fd0\u884c\u6570\u4e07\u4ebf\u5b57\u8282\u6570\u636e\u7684\u5e94\u7528\u6027\u80fd\u63d0\u9ad8 10 \u500d\u3002<\/p>\n

 <\/p>\n

\u6ce8\uff1a\u4e0e NVIDIA \u4ea7\u54c1\u76f8\u5173\u7684\u56fe\u7247\u6216\u89c6\u9891\uff08\u5b8c\u6574\u6216\u90e8\u5206\uff09\u7684\u7248\u6743\u5747\u5f52 NVIDIA Corporation \u6240\u6709\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

  \u901a\u8fc7 NVIDIA H800 Tensor Core GPU\uff0c\u5728\u6bcf\u4e2a\u5de5\u4f5c\u8d1f\u8f7d\u4e2d\u5b9e\u73b0\u51fa\u8272\u6027\u80fd\u3001\u53ef\u6269 […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[74],"tags":[],"_links":{"self":[{"href":"http:\/\/shlj-tec.com\/index.php\/wp-json\/wp\/v2\/posts\/1825"}],"collection":[{"href":"http:\/\/shlj-tec.com\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/shlj-tec.com\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/shlj-tec.com\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/shlj-tec.com\/index.php\/wp-json\/wp\/v2\/comments?post=1825"}],"version-history":[{"count":1,"href":"http:\/\/shlj-tec.com\/index.php\/wp-json\/wp\/v2\/posts\/1825\/revisions"}],"predecessor-version":[{"id":1827,"href":"http:\/\/shlj-tec.com\/index.php\/wp-json\/wp\/v2\/posts\/1825\/revisions\/1827"}],"wp:attachment":[{"href":"http:\/\/shlj-tec.com\/index.php\/wp-json\/wp\/v2\/media?parent=1825"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/shlj-tec.com\/index.php\/wp-json\/wp\/v2\/categories?post=1825"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/shlj-tec.com\/index.php\/wp-json\/wp\/v2\/tags?post=1825"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}