{"id":669,"date":"2025-07-23T23:50:00","date_gmt":"2025-07-23T15:50:00","guid":{"rendered":"http:\/\/www.applicationsofllm.com\/?p=669"},"modified":"2025-09-16T23:53:44","modified_gmt":"2025-09-16T15:53:44","slug":"%e8%b0%81%e5%9c%a8%e5%ae%9a%e4%b9%89%e6%9c%aa%e6%9d%a5ai%e7%ae%97%e5%8a%9b%ef%bc%9fblackwell%e7%bb%99%e5%87%ba%e5%85%a8%e6%96%b0%e7%ad%94%e6%a1%88","status":"publish","type":"post","link":"http:\/\/www.applicationsofllm.com\/index.php\/2025\/07\/23\/%e8%b0%81%e5%9c%a8%e5%ae%9a%e4%b9%89%e6%9c%aa%e6%9d%a5ai%e7%ae%97%e5%8a%9b%ef%bc%9fblackwell%e7%bb%99%e5%87%ba%e5%85%a8%e6%96%b0%e7%ad%94%e6%a1%88\/","title":{"rendered":"\u8c01\u5728\u5b9a\u4e49\u672a\u6765AI\u7b97\u529b\uff1fBlackwell\u7ed9\u51fa\u5168\u65b0\u7b54\u6848"},"content":{"rendered":"\n

<\/h1>\n\n\n\n

\u5f15\u8a00<\/strong>\u5728\u751f\u6210\u5f0fAI\u548c\u52a0\u901f\u8ba1\u7b97\u8fc8\u5165\u201c\u4e07\u4ebf\u53c2\u6570\u201d\u65f6\u4ee3\u7684\u5173\u952e\u8282\u70b9\uff0cNVIDIA\u63a8\u51faBlackwell\u5e73\u53f0\uff0c\u4ee5\u5168\u6808\u5f0f\u67b6\u6784\u9769\u65b0\u91cd\u5851AI\u57fa\u7840\u8bbe\u65bd\u3002\u8be5\u5e73\u53f0\u4e0d\u518d\u4ec5\u4ec5\u662f\u4e00\u9897GPU\u7684\u5347\u7ea7\uff0c\u800c\u662f\u4ece\u5e95\u5c42\u786c\u4ef6\u3001\u4e92\u8054\u67b6\u6784\u5230\u9ad8\u5c42\u8f6f\u4ef6\u6808\u7684\u7cfb\u7edf\u7ea7\u91cd\u6784\u3002Blackwell\u901a\u8fc7\u5f15\u5165FP4\u4f4e\u7cbe\u5ea6\u8ba1\u7b97\u3001\u7b2c\u4e94\u4ee3NVLink\u4e92\u8054\u3001Quasar\u91cf\u5316\u7cfb\u7edf\u53caGrace-Blackwell\u8d85\u7ea7\u82af\u7247\uff0c\u63d0\u4f9b\u524d\u6240\u672a\u6709\u7684AI\u6027\u80fd\u4e0e\u80fd\u6548\u6bd4\uff0c\u5168\u9762\u652f\u6301\u5927\u6a21\u578b\u8bad\u7ec3\u3001\u63a8\u7406\u53ca\u79d1\u5b66\u8ba1\u7b97\u573a\u666f\uff0c\u6807\u5fd7\u7740\u52a0\u901f\u8ba1\u7b97\u8303\u5f0f\u6b63\u5f0f\u4ece\u82af\u7247\u8d70\u5411\u5e73\u53f0\uff0c\u4ece\u5355\u70b9\u6027\u80fd\u8fc8\u5411\u6570\u636e\u4e2d\u5fc3\u89c4\u6a21\u534f\u540c\u3002<\/p>\n\n\n\n

\u4ee5\u4e0b\u662f\u5bf9\u62a5\u544a\u5185\u5bb9\u7684\u68b3\u7406\u603b\u7ed3\u2014\u2014\u2014\u2014\u2014\u2014\u2014\u2014\u2014\u2014\u2014\u2014\u2014\u2014<\/p>\n\n\n\n

\u4e00\u3001\u5e73\u53f0\u5168\u666f\uff1aBlackwell\u7684\u7cfb\u7edf\u7ea7\u6218\u7565\u8dc3\u8fc1<\/h2>\n\n\n\n



<\/h2>\n\n\n\n

1.1 \u5e73\u53f0\u7ea7\u8bbe\u8ba1\u7406\u5ff5\uff1a\u4ece\u5355\u82af\u7247\u5230\u6570\u636e\u4e2d\u5fc3<\/h3>\n\n\n\n



<\/h3>\n\n\n\n

NVIDIA Blackwell\u5e73\u53f0\u4ee5\u201c\u6570\u636e\u4e2d\u5fc3\u7ea7\u201d\u67b6\u6784\u4e3a\u6838\u5fc3\u7406\u5ff5\uff0c\u6253\u7834\u4f20\u7edf\u5355\u4e00GPU\/\u82af\u7247\u601d\u7ef4\uff0c\u91cd\u5851\u751f\u6210\u5f0fAI\u548c\u52a0\u901f\u8ba1\u7b97\u7684\u57fa\u7840\u8bbe\u65bd\u3002\u5e73\u53f0\u56f4\u7ed5\u5b8c\u6574AI\u6808\u6784\u5efa\uff0c\u6db5\u76d6\u4ece\u6700\u5e95\u5c42\u786c\u4ef6\u5230\u7cfb\u7edf\u8f6f\u4ef6\u3001\u52a0\u901f\u5e93\u4e0e\u5e94\u7528\u6846\u67b6\uff0c\u5b9e\u73b0\u4e86\u5bf9AI\u4e0e\u9ad8\u6027\u80fd\u8ba1\u7b97\u573a\u666f\u7684\u7aef\u5230\u7aef\u8986\u76d6\u3002\u901a\u8fc7\u5e73\u53f0\u5316\u96c6\u6210\uff0cBlackwell\u4e0d\u4ec5\u63d0\u4f9b\u82af\u7247\u7ea7\u6027\u80fd\u8dc3\u5347\uff0c\u66f4\u5c06\u5f02\u6784\u8ba1\u7b97\u80fd\u529b\u5ef6\u5c55\u81f3\u6574\u4e2a\u6570\u636e\u4e2d\u5fc3\uff0c\u5b9e\u73b0\u7b97\u529b\u3001\u901a\u4fe1\u4e0e\u80fd\u6548\u7684\u534f\u540c\u6700\u4f18\u5316\u3002<\/p>\n\n\n\n

1.2 \u6784\u5efa\u5168\u6808\u80fd\u529b\uff1a\u8f6f\u786c\u534f\u540c\u4e0e\u751f\u6001\u6574\u5408<\/h3>\n\n\n\n



<\/h3>\n\n\n\n

Blackwell\u5e73\u53f0\u56f4\u7ed5CUDA-X\u751f\u6001\u6784\u5efa\uff0c\u5185\u542b\u8d85\u8fc7400\u4e2a\u6df1\u5ea6\u4f18\u5316\u7684\u52a0\u901f\u5e93\uff0c\u8986\u76d6\u8bed\u97f3\u3001\u89c6\u89c9\u3001\u63a8\u8350\u7cfb\u7edf\u3001\u79d1\u5b66\u8ba1\u7b97\u3001\u641c\u7d22\u3001\u6570\u5b57\u4eba\u7b49\u5173\u952e\u5e94\u7528\u9886\u57df\u3002\u5176\u8f6f\u4ef6\u5c42\u7531DOCA\u3001Base Command\u3001Magnum IO\u7b49\u7cfb\u7edf\u7ec4\u4ef6\u7ec4\u6210\uff0c\u4e0a\u5c42\u96c6\u6210\u4e86\u5305\u62ecCUDA\u3001RTX\u3001NVIDIA AI\u3001Omniverse\u7b49\u5728\u5185\u7684\u5e94\u7528\u5e73\u53f0\uff0c\u5f62\u6210\u4ece\u5e95\u5c42\u786c\u4ef6\u5230\u9ad8\u5c42\u5e94\u7528\u7684\u7eb5\u5411\u95ed\u73af\u3002\u8f6f\u786c\u534f\u540c\u4e3a\u6a21\u578b\u8bad\u7ec3\u4e0e\u63a8\u7406\u63d0\u4f9b\u5b8c\u6574\u5de5\u5177\u94fe\u652f\u6301\uff0c\u6784\u6210\u53ef\u6301\u7eed\u6269\u5c55\u7684AI\u57fa\u7840\u8bbe\u65bd\u3002<\/p>\n\n\n\n

1.3 \u5e73\u53f0\u5173\u952e\u6784\u6210\uff1aGPU\u3001CPU\u3001NVSwitch\u3001DPU\u3001NIC\u3001\u4ea4\u6362\u7f51\u7edc\u5168\u94fe\u8def\u8986\u76d6<\/h3>\n\n\n\n



<\/h3>\n\n\n\n

Blackwell\u5e73\u53f0\u7684\u6838\u5fc3\u786c\u4ef6\u6784\u6210\u5305\u62ecBlackwell GPU\u3001Grace CPU\u3001\u7b2c\u4e94\u4ee3NVSwitch\u82af\u7247\u3001BlueField-3 DPU\u3001ConnectX-7\/8 NIC\uff0c\u4ee5\u53caSpectrum-X800\u4e0eQuantum-X800\u7b49\u7528\u4e8eAI\u4f18\u5316\u7684\u4ea4\u6362\u8bbe\u5907\u3002\u5e73\u53f0\u4ee5\u9ad8\u5e26\u5bbd\u3001\u4f4e\u5ef6\u8fdf\u3001\u9ad8\u80fd\u6548\u4e3a\u57fa\u7840\uff0c\u901a\u8fc7Compute Tray\u548cNVSwitch Tray\u7b49\u6a21\u5757\u5316\u5355\u5143\u5b9e\u73b0\u89c4\u6a21\u5316\u90e8\u7f72\uff0c\u652f\u6491\u4ece\u5355\u8282\u70b9\u5230\u6570\u5343GPU\u7ea7\u522b\u7684\u8bad\u7ec3\u548c\u63a8\u7406\u5de5\u4f5c\u8d1f\u8f7d\u3002<\/p>\n\n\n\n

1.4 \u5e94\u7528\u5bfc\u5411\uff1a\u751f\u6210\u5f0fAI\u3001\u79d1\u5b66\u8ba1\u7b97\u3001\u6570\u5b57\u4eba\u7b49\u573a\u666f\u5168\u8986\u76d6<\/h3>\n\n\n\n



<\/h3>\n\n\n\n

Blackwell\u5e73\u53f0\u8bbe\u8ba1\u9762\u5411\u4e0b\u4e00\u4ee3\u8d85\u5927\u89c4\u6a21AI\u6a21\u578b\uff0c\u5305\u62ec\u751f\u6210\u5f0f\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u3001\u89c6\u89c9\u8bed\u8a00\u6a21\u578b\uff08VLM\uff09\u3001\u63a8\u8350\u7cfb\u7edf\u3001\u79d1\u5b66\u6a21\u62df\u3001\u6570\u5b57\u4eba\u548c\u7269\u7406AI\u7b49\u590d\u6742\u8d1f\u8f7d\u3002\u5176\u8f6f\u786c\u4ef6\u534f\u540c\u6808\u652f\u6301FP4\/FP6\u7b49\u4f4e\u7cbe\u5ea6\u63a8\u7406\u80fd\u529b\uff0c\u663e\u8457\u63d0\u5347\u5355\u4f4d\u80fd\u6548\u6bd4\u4e0e\u8ba1\u7b97\u541e\u5410\uff0c\u540c\u65f6\u4fdd\u969c\u6a21\u578b\u51c6\u786e\u6027\u4e0e\u7cfb\u7edf\u7a33\u5b9a\u6027\u3002\u5e73\u53f0\u67b6\u6784\u652f\u6301\u4e13\u5bb6\u5e76\u884c\u3001\u5f20\u91cf\u5e76\u884c\u4e0e\u6d41\u6c34\u5e76\u884c\u7b49\u591a\u79cd\u8c03\u5ea6\u6a21\u5f0f\uff0c\u9002\u914d\u591a\u6837\u5316\u7684AI\u63a8\u7406\u548c\u8bad\u7ec3\u573a\u666f\u3002<\/p>\n\n\n\n

\u4e8c\u3001Blackwell GPU\u6df1\u5ea6\u89e3\u6790\uff1a\u6027\u80fd\u4e0e\u67b6\u6784\u7684\u8dc3\u8fc1\u5f0f\u6f14\u8fdb<\/h2>\n\n\n\n



<\/h2>\n\n\n\n

2.1 \u5236\u7a0b\u4e0e\u5c01\u88c5\uff1aTSMC 4NP + \u53ccReticle\u5408\u5e76\u8bbe\u8ba1<\/h3>\n\n\n\n



<\/h3>\n\n\n\n

Blackwell GPU\u91c7\u7528\u53f0\u79ef\u75354NP\u5148\u8fdb\u5236\u7a0b\u5de5\u827a\uff0c\u5728\u7845\u7247\u9762\u79ef\u4e0a\u7a81\u7834\u7269\u7406\u6781\u9650\uff0c\u91c7\u7528\u53ccReticle\u9650\u5236\u7684\u82af\u7247\u5408\u5e76\u4e3a\u5355\u4e00\u5927\u82af\u7ed3\u6784\uff0c\u96c6\u6210\u9ad8\u8fbe2080\u4ebf\u4e2a\u6676\u4f53\u7ba1\u3002\u8fd9\u4e00\u67b6\u6784\u4e0d\u4ec5\u63d0\u5347\u4e86\u82af\u7247\u89c4\u6a21\uff0c\u4e5f\u6781\u5927\u589e\u5f3a\u4e86\u6838\u5fc3\u5185\u90e8\u7684\u901a\u4fe1\u5bc6\u5ea6\u4e0e\u80fd\u6548\u6548\u7387\uff0c\u4e3aAI\u6a21\u578b\u63d0\u4f9b\u66f4\u4f18\u7684\u6267\u884c\u8def\u5f84\u4e0e\u541e\u5410\u6027\u80fd\u3002<\/p>\n\n\n\n

2.2 \u8fd0\u7b97\u80fd\u529b\u7a81\u7834\uff1a20 PetaFLOPS FP4\u30018 TB\/s \u5e26\u5bbd<\/h3>\n\n\n\n



<\/h3>\n\n\n\n

\u5728\u8fd0\u7b97\u80fd\u529b\u4e0a\uff0cBlackwell GPU\u521b\u9020\u4e86\u5355\u82af\u7247AI\u6027\u80fd\u7684\u65b0\u7eaa\u5f55\u3002\u5176\u652f\u6301\u9ad8\u8fbe20 PetaFLOPS\u7684FP4\u7cbe\u5ea6AI\u8ba1\u7b97\u80fd\u529b\uff0c\u5e76\u5177\u59078 TB\/s\u7684\u663e\u5b58\u5e26\u5bbd\uff0c\u642d\u8f7d8\u5806HBM3e\u9ad8\u5e26\u5bbd\u5185\u5b58\u3002\u540c\u65f6\uff0c\u5176\u53cc\u5411NVLink\u5e26\u5bbd\u8fbe\u52301.8 TB\/s\uff0c\u8f85\u4ee5\u4e0eGrace CPU\u4e4b\u95f4\u7684\u9ad8\u901fC2C\u4e92\u8054\uff0c\u663e\u8457\u964d\u4f4e\u5ef6\u8fdf\u74f6\u9888\uff0c\u63d0\u5347\u6574\u4f53\u4efb\u52a1\u6267\u884c\u6548\u7387\u3002<\/p>\n\n\n\n

2.3 \u5185\u90e8\u901a\u4fe1\u4f18\u5316\uff1aNV-HBI + \u9ad8\u5bc6\u5ea6\u4f4e\u5ef6\u8fdf\u67b6\u6784<\/h3>\n\n\n\n



<\/h3>\n\n\n\n

Blackwell\u5f15\u5165NVIDIA\u9ad8\u5e26\u5bbd\u4e92\u8054\u63a5\u53e3\uff08NV-HBI\uff09\uff0c\u5728\u82af\u7247\u8fb9\u7f18\u5b9e\u73b0\u6bcf\u79d210 TB\u7684\u53cc\u5411\u5e26\u5bbd\u4f20\u8f93\uff0c\u80fd\u6548\u4f18\u5316\u663e\u8457\uff0c\u652f\u6301GPU\u95f4\u4e00\u81f4\u6027\u8fde\u63a5\u3002\u5728\u7edf\u4e00GPU\u7ed3\u6784\u4e0b\uff0cBlackwell\u5b9e\u73b0\u5168\u6027\u80fd\u91ca\u653e\uff0c\u5177\u5907\u6781\u9ad8\u7684\u901a\u4fe1\u5bc6\u5ea6\u3001\u6781\u4f4e\u7684\u8bbf\u95ee\u5ef6\u8fdf\u548c\u5353\u8d8a\u7684\u80fd\u6548\u8868\u73b0\uff0c\u4e3a\u6570\u636e\u4e2d\u5fc3\u7ea7\u90e8\u7f72\u5960\u5b9a\u9ad8\u901f\u4e92\u8fde\u57fa\u7840\u3002<\/p>\n\n\n\n

2.4 \u6838\u5fc3\u6280\u672f\u4eae\u70b9\uff1aTransformer Engine\u3001\u5168\u94fe\u8def\u52a0\u5bc6\u3001800GB\/s\u89e3\u538b\u5f15\u64ce<\/h3>\n\n\n\n



<\/h3>\n\n\n\n

Blackwell GPU\u5185\u5d4c\u65b0\u4e00\u4ee3Transformer Engine\uff0c\u4e13\u4e3a\u5927\u6a21\u578b\u8bad\u7ec3\u548c\u63a8\u7406\u8bbe\u8ba1\uff0c\u652f\u6301FP4\/FP6\/FP8\u7b49\u65b0\u7cbe\u5ea6\u683c\u5f0f\u3002\u82af\u7247\u914d\u5907RAS\u5f15\u64ce\uff0c\u5b9e\u73b0100%\u7cfb\u7edf\u5185\u81ea\u68c0\u80fd\u529b\uff0c\u5e76\u96c6\u6210\u5168\u6027\u80fd\u5b89\u5168\u52a0\u5bc6\u673a\u5236\u4e0e\u53d7\u4fe1\u6267\u884c\u73af\u5883\uff08TEE\uff09\uff0c\u4fdd\u969cAI\u7b97\u529b\u5728\u5173\u952e\u4efb\u52a1\u4e2d\u7684\u6570\u636e\u5b89\u5168\u6027\u3002\u6b64\u5916\uff0c\u5185\u7f6e\u89e3\u538b\u5f15\u64ce\u5e26\u5bbd\u9ad8\u8fbe800 GB\/s\uff0c\u4e3a\u5927\u6a21\u578b\u7684\u5feb\u901f\u52a0\u8f7d\u4e0e\u5904\u7406\u63d0\u4f9b\u5f3a\u5927\u652f\u6491\u3002<\/p>\n\n\n\n

2.5 \u4e0e\u524d\u4ee3\u5bf9\u6bd4\uff1a\u4eceVolta\u5230Hopper\u7684\u6f14\u8fdb\u8f68\u8ff9<\/h3>\n\n\n\n



<\/h3>\n\n\n\n

\u5bf9\u6bd4NVIDIA\u8fc7\u5f80\u67b6\u6784\u4ee3\u9645\u6f14\u8fdb\uff0cBlackwell GPU\u5728\u6676\u4f53\u7ba1\u89c4\u6a21\u3001\u8ba1\u7b97\u5bc6\u5ea6\u4e0e\u4e92\u8054\u5e26\u5bbd\u4e0a\u5b9e\u73b0\u91cf\u7ea7\u63d0\u5347\uff1a\u4eceVolta\u7684210\u4ebf\u6676\u4f53\u7ba1\uff08TSMC 12nm\uff09\u5230Hopper\u7684800\u4ebf\uff08TSMC 4N\uff09\uff0c\u518d\u5230Blackwell\u76842080\u4ebf\uff08TSMC 4NP\uff09\u3002\u540c\u65f6\uff0c\u5176\u82af\u7247\u9762\u79ef\u8d85\u8fc71600mm\u00b2\uff0c\u4e3a\u8fc7\u5f80\u7684\u4e24\u500d\uff0c\u5c55\u73b0\u51fa\u9762\u5411\u4e0b\u4e00\u9636\u6bb5AI\u57fa\u7840\u8bbe\u65bd\u7684\u7269\u7406\u6781\u9650\u7a81\u7834\u80fd\u529b\u3002<\/p>\n\n\n\n

\u4e09\u3001Grace Blackwell\u8d85\u7ea7\u82af\u7247\uff1a\u7b97\u529b\u5355\u5143\u7684\u518d\u5b9a\u4e49<\/h2>\n\n\n\n



<\/h2>\n\n\n\n

3.1 \u67b6\u6784\u7ec4\u5408\uff1a1 Grace CPU + 2 Blackwell GPU<\/h3>\n\n\n\n



<\/h3>\n\n\n\n

NVIDIA Grace Blackwell\u8d85\u7ea7\u82af\u7247\uff08GB200\uff09\u5c061\u9897Grace CPU\u4e0e2\u9897Blackwell GPU\u901a\u8fc7NVLink-C2C\u9ad8\u901f\u4e92\u8054\u96c6\u6210\u4e3a\u5355\u4e00\u8ba1\u7b97\u5355\u5143\uff0c\u6784\u5efa\u5f02\u6784\u878d\u5408\u7684AI\u8ba1\u7b97\u6838\u5fc3\u3002\u8be5\u7ec4\u5408\u4e0d\u4ec5\u5f3a\u5316\u4e86\u5904\u7406\u5668\u95f4\u7684\u534f\u540c\u8ba1\u7b97\u80fd\u529b\uff0c\u4e5f\u5728\u5e26\u5bbd\u4e0e\u529f\u8017\u4e4b\u95f4\u5b9e\u73b0\u6700\u4f18\u5e73\u8861\uff0c\u5f62\u6210\u9002\u914d\u5927\u6a21\u578b\u63a8\u7406\u4e0e\u8bad\u7ec3\u7684\u65b0\u4e00\u4ee3\u7b97\u529b\u5f62\u6001\u3002<\/p>\n\n\n\n

3.2 \u6027\u80fd\u53c2\u6570\uff1a40 PetaFLOPS FP4 \/ 20 PetaFLOPS FP8<\/h3>\n\n\n\n



<\/h3>\n\n\n\n

Grace Blackwell\u8d85\u7ea7\u82af\u7247\u5728AI\u7cbe\u5ea6\u8ba1\u7b97\u4e0a\u5b9e\u73b0\u6307\u6570\u7ea7\u7a81\u7834\uff1a\u5355\u4e2a\u82af\u7247\u652f\u630140 PetaFLOPS\u7684FP4\u4e0e20 PetaFLOPS\u7684FP8\u8ba1\u7b97\u6027\u80fd\uff0c\u8fdc\u8d85\u524d\u4ee3\u5e73\u53f0\u3002\u8fd9\u4e00\u7b97\u529b\u8dc3\u8fc1\u4f9d\u6258\u65b0\u4e00\u4ee3Tensor Core\u67b6\u6784\u4e0e\u66f4\u9ad8\u6570\u91cf\u7684\u6d41\u5904\u7406\u5668\uff08SM\uff09\uff0c\u53ef\u5168\u9762\u8986\u76d6\u9ad8\u5e76\u53d1\u3001\u591a\u7ef4\u5ea6\u7684AI\u63a8\u7406\u8d1f\u8f7d\u3002<\/p>\n\n\n\n

3.3 \u6258\u76d8\u8bbe\u8ba1\uff1a\u591a\u82af\u591a\u5361\u534f\u540c\u90e8\u7f72\u7684\u53ef\u6269\u5c55\u6027<\/h3>\n\n\n\n



<\/h3>\n\n\n\n

\u5e73\u53f0\u652f\u6301\u6a21\u5757\u5316\u7684Compute Tray\uff08\u8ba1\u7b97\u6258\u76d8\uff09\u90e8\u7f72\u65b9\u5f0f\uff0c\u6bcf\u4e2a\u6258\u76d8\u96c6\u62102\u9897Grace CPU\u4e0e4\u9897Blackwell GPU\uff0c\u53ef\u7075\u6d3b\u6269\u5c55\u6784\u5efa\u5927\u578b\u96c6\u7fa4\u3002\u6258\u76d8\u95f4\u901a\u8fc7\u7b2c\u4e94\u4ee3NVLink\u5b9e\u73b0\u9ad8\u5e26\u5bbd\u4e92\u8054\uff0c\u7cfb\u7edf\u6574\u4f53\u5177\u5907\u9ad8\u541e\u5410\u3001\u4f4e\u5ef6\u8fdf\u3001\u7ebf\u6027\u6269\u5c55\u7684\u80fd\u529b\uff0c\u4e3a\u4e91\u7aef\u5927\u6a21\u578b\u63a8\u7406\u573a\u666f\u63d0\u4f9b\u7b97\u529b\u57fa\u7840\u3002<\/p>\n\n\n\n

3.4 \u6570\u636e\u5904\u7406\u4f18\u52bf\uff1a\u9ad8\u5e26\u5bbd\u3001\u4f4e\u5ef6\u8fdf\u3001KV\u7f13\u5b58\u4f18\u5316<\/h3>\n\n\n\n



<\/h3>\n\n\n\n

GB200\u67b6\u6784\u9762\u5411AI\u63a8\u7406\u8fc7\u7a0b\u4e2d\u7684\u74f6\u9888\u95ee\u9898\uff0c\u96c6\u6210\u5173\u952e\u503c\uff08Key Value\uff09\u7f13\u5b58\u673a\u5236\uff0c\u63d0\u5347\u5927\u6a21\u578b\u4e2d\u7684\u8bb0\u5fc6\u8bbf\u95ee\u6548\u7387\u3002\u5728\u9ad8\u5e26\u5bbd\u4f4e\u5ef6\u8fdf\u7684\u6570\u636e\u4f20\u8f93\u80fd\u529b\u57fa\u7840\u4e0a\uff0cBlackwell-GPU\u4e0eGrace-CPU\u534f\u540c\u5904\u7406\u80fd\u529b\u663e\u8457\u589e\u5f3a\uff0c\u6709\u6548\u964d\u4f4eToken-to-Token\u63a8\u7406\u5ef6\u8fdf\uff0c\u4e3a\u5b9e\u65f6\u751f\u6210\u5f0fAI\u63d0\u4f9b\u7cfb\u7edf\u7ea7\u652f\u6491\u3002<\/strong>\u56db\u3001Quasar\u91cf\u5316\u7cfb\u7edf\uff1a\u4f4e\u7cbe\u5ea6AI\u7684\u8303\u5f0f\u7a81\u7834<\/strong><\/p>\n\n\n\n

4.1 \u4f4e\u7cbe\u5ea6\u8ba1\u7b97\u7684\u6311\u6218\u4e0e\u58c1\u5792<\/h3>\n\n\n\n



<\/h3>\n\n\n\n

\u5728\u4f4e\u7cbe\u5ea6AI\u63a8\u7406\u9886\u57df\uff0c\u4f20\u7edf\u65b9\u6848\u9762\u4e34\u591a\u4e2a\u6839\u672c\u6027\u6311\u6218\uff1a\u975e\u5747\u5300\u5c42\u7ea7\u654f\u611f\u6027\u3001\u52a8\u6001\u8303\u56f4\u72ed\u7a84\u3001\u91cf\u5316\u566a\u58f0\u5e26\u6765\u7684\u7cbe\u5ea6\u635f\u5931\uff0c\u4ee5\u53ca\u63a8\u7406\u51c6\u786e\u7387\u4e0b\u964d\u95ee\u9898\u3002Blackwell\u5e73\u53f0\u901a\u8fc7\u7cfb\u7edf\u7ea7\u8f6f\u786c\u534f\u540c\u8bbe\u8ba1\uff0c\u529b\u56fe\u6253\u7834\u4f4e\u7cbe\u5ea6\u8ba1\u7b97\u4e0e\u9ad8\u51c6\u786e\u7387\u4e4b\u95f4\u7684\u56fa\u6709\u6743\u8861\uff0c\u63a8\u52a8AI\u80fd\u6548\u6781\u9650\u7a81\u7834\u3002<\/p>\n\n\n\n

4.2 \u5fae\u5f20\u91cf\u7f29\u653e\u673a\u5236\uff1aFP4\/FP6\/FP8\u65b0\u683c\u5f0f\u89e3\u8bfb<\/h3>\n\n\n\n



<\/h3>\n\n\n\n

Blackwell\u5f15\u5165\u7b2c\u4e94\u4ee3Tensor Core\uff0c\u652f\u6301\u5168\u65b0\u7684\u5fae\u5f20\u91cf\u7f29\u653e\u683c\u5f0f\uff08Micro Tensor Scaling\uff09FP4\u3001FP6\u4e0e\u589e\u5f3a\u578bFP8\u3002\u8fd9\u4e9b\u683c\u5f0f\u5728\u6bcf\u4e2a\u56fa\u5b9a\u957f\u5ea6\u5411\u91cf\u4e0a\u5f15\u5165\u72ec\u7acb\u7f29\u653e\u56e0\u5b50\uff0c\u4f7f\u5176\u5177\u5907\u66f4\u5e7f\u7684\u52a8\u6001\u8303\u56f4\u3001\u66f4\u9ad8\u7684\u5e26\u5bbd\u6548\u7387\u4e0e\u66f4\u4f4e\u7684\u529f\u8017\uff0c\u89e3\u51b3\u7cbe\u5ea6\u9000\u5316\u4e0e\u6a21\u578b\u4e0d\u7a33\u5b9a\u95ee\u9898\u3002<\/p>\n\n\n\n

4.3 \u7b97\u6cd5+\u786c\u4ef6\u534f\u540c\uff1aTensor Core + TensorRT + Megatron \u6808\u5f0f\u878d\u5408<\/h3>\n\n\n\n



<\/h3>\n\n\n\n

Quasar\u91cf\u5316\u7cfb\u7edf\u878d\u5408\u4e86\u786c\u4ef6\u5c42\uff08Transformer Engine\u3001Blackwell Tensor Cores\uff09\u3001\u7cfb\u7edf\u5e93\u5c42\uff08TensorRT\u3001cuDNN\u3001Megatron-Core\uff09\u3001\u4ee5\u53ca\u4f18\u5316\u7b97\u6cd5\u5c42\uff08Dynamic Range Regularization\u3001Layer Sensitivity Selection\u7b49\uff09\uff0c\u6784\u6210\u4ece\u6a21\u578b\u5230\u5e95\u5c42\u786c\u4ef6\u7684\u7aef\u5230\u7aef\u91cf\u5316\u6808\uff0c\u652f\u6491\u5927\u6a21\u578b\u7684\u4f4e\u7cbe\u5ea6\u8bad\u7ec3\u4e0e\u63a8\u7406\u6d41\u7a0b\u3002<\/p>\n\n\n\n

4.4 \u7cbe\u5ea6\u4e0e\u80fd\u6548\u5e76\u4e3e\uff1aNemotron-4 340B\u5168\u7cbe\u5ea6\u5bf9\u6bd4\u6570\u636e\u9a8c\u8bc1<\/h3>\n\n\n\n



<\/h3>\n\n\n\n

\u57fa\u4e8eBlackwell\u5b9e\u9645\u7845\u7247\u6d4b\u8bd5\u7ed3\u679c\uff0c\u4f7f\u7528Quasar\u7cfb\u7edf\u5c06Nemotron-4\u6a21\u578b\u91cf\u5316\u4e3aFP4\u540e\uff0c\u6a21\u578b\u5728MMLU\u6d4b\u8bd5\u4e2d\u7684\u51c6\u786e\u7387\u4e0eBF16\u7248\u672c\u4fdd\u6301\u4e00\u81f4\uff0c\u751a\u81f3\u5728\u90e8\u5206\u6a21\u578b\uff08\u5982Nemotron-4 15B\uff09\u4e2d\u7565\u6709\u63d0\u5347\uff0864.2 \u2192 64.5\uff09\u3002\u8fd9\u4e00\u6210\u679c\u9a8c\u8bc1\u4e86\u5176\u5728\u786e\u4fdd\u7cbe\u5ea6\u7684\u524d\u63d0\u4e0b\u5b9e\u73b0\u7b97\u529b\u4e0e\u5e26\u5bbd\u7684\u5927\u5e45\u8282\u7ea6\u3002<\/p>\n\n\n\n

4.5 \u5e94\u7528\u8868\u73b0\uff1a\u56fe\u50cf\u751f\u6210\u3001LLM\u63a8\u7406\u7cbe\u5ea6\u4e0e\u901f\u5ea6\u517c\u5f97<\/h3>\n\n\n\n



<\/h3>\n\n\n\n

\u901a\u8fc7FP4\u683c\u5f0f\u5728Stable Diffusion XL\u7b49\u751f\u6210\u6a21\u578b\u4e0a\u7684\u6d4b\u8bd5\uff0cBlackwell\u5c55\u73b0\u51fa\u5728\u56fe\u50cf\u751f\u6210\u8d28\u91cf\u3001\u8bed\u4e49\u8868\u8fbe\u5b8c\u6574\u6027\u4e0e\u63a8\u7406\u54cd\u5e94\u901f\u5ea6\u4e0a\u7684\u5168\u65b9\u4f4d\u4f18\u52bf\u3002\u5f97\u76ca\u4e8eQuasar\u7684\u91cf\u5316\u673a\u5236\uff0c\u6a21\u578b\u53ef\u5728\u6781\u4f4e\u7cbe\u5ea6\u4e0b\u5b9e\u73b0\u9ad8\u4fdd\u771f\u5ea6\u8f93\u51fa\uff0c\u4f7fBlackwell\u6210\u4e3a\u5927\u89c4\u6a21\u751f\u6210\u5f0fAI\u63a8\u7406\u7684\u7406\u60f3\u5e73\u53f0\u3002<\/p>\n\n\n\n

\u4e94\u3001\u7f51\u7edc\u4e0e\u4e92\u8fde\uff1a\u6784\u5efa\u4e07\u4ebf\u53c2\u6570\u65f6\u4ee3\u7684\u5e95\u5c42\u80fd\u529b<\/h2>\n\n\n\n



<\/h2>\n\n\n\n

5.1 \u6a21\u578b\u89c4\u6a21\u589e\u957f\u8d8b\u52bf\u4e0e\u5e76\u884c\u63a8\u7406\u9700\u6c42<\/h3>\n\n\n\n



<\/h3>\n\n\n\n

\u968f\u7740AI\u6a21\u578b\u53c2\u6570\u89c4\u6a21\u5448\u6307\u6570\u7ea7\u589e\u957f\uff0c\u4eceResNet-50\u7684\u6570\u5343\u4e07\u53c2\u6570\u5230GPT-MoE-1.8T\u7684\u4e07\u4ebf\u7ea7\u6a21\u578b\uff0c\u8fc7\u53bb\u5341\u5e74\u5b9e\u73b0\u4e86\u8d8570,000\u500d\u6269\u5c55\u3002\u63a8\u7406\u9636\u6bb5\u5bf9\u591aGPU\u534f\u540c\u8ba1\u7b97\u7684\u9700\u6c42\u6fc0\u589e\uff0c\u5c24\u5176\u662f\u5728\u751f\u6210\u5f0fAI\u5b9e\u65f6\u54cd\u5e94\u573a\u666f\u4e0b\uff0c\u5bf9\u541e\u5410\u7387\u3001\u901a\u4fe1\u5e26\u5bbd\u4e0e\u7aef\u5230\u7aef\u5ef6\u8fdf\u63d0\u51fa\u4e86\u7a7a\u524d\u8981\u6c42\u3002<\/p>\n\n\n\n

5.2 NVLink 5\u6280\u672f\u6f14\u8fdb\uff1a1800 GB\/s\u4e92\u8054\u5e26\u5bbd<\/h3>\n\n\n\n



<\/h3>\n\n\n\n

Blackwell\u5e73\u53f0\u642d\u8f7d\u7b2c\u4e94\u4ee3NVLink\u4e92\u8fde\u67b6\u6784\uff0c\u6bcf\u9897GPU\u5177\u590718\u4e2aNVLink\u63a5\u53e3\uff0c\u6bcf\u4e2a\u94fe\u8def\u652f\u6301100 GB\/s\u7684\u5e26\u5bbd\uff0c\u603b\u5e26\u5bbd\u9ad8\u8fbe1.8 TB\/s\u3002\u76f8\u6bd4Hopper\uff08900 GB\/s\uff09\u548cAmpere\uff08600 GB\/s\uff09\uff0cNVLink 5\u5b9e\u73b0\u7ffb\u500d\u63d0\u5347\uff0c\u91c7\u7528PAM4\u7f16\u7801\u4e0e\u66f4\u9ad8\u7684\u4fe1\u53f7\u901f\u7387\uff0c\u6ee1\u8db3\u9ad8\u5e26\u5bbd\u3001\u4f4e\u5ef6\u8fdf\u7684\u6a21\u578b\u5206\u5e03\u5f0f\u63a8\u7406\u9700\u6c42\u3002<\/p>\n\n\n\n

5.3 NVLink Switch Tray\u4e0eSwitch Chip\uff1a\u652f\u649172 GPU\u5168\u8fde\u63a5\u67b6\u6784<\/h3>\n\n\n\n



<\/h3>\n\n\n\n

\u7b2c\u56db\u4ee3NVLink Switch\u82af\u7247\u57fa\u4e8eTSMC 4NP\u5de5\u827a\u5236\u9020\uff0c\u9762\u79ef\u8d85\u8fc7800 mm\u00b2\uff0c\u5355\u82af\u652f\u630172\u8defGPU\u5168\u4e92\u8054\uff0c\u5168\u53cc\u5de5\u5e26\u5bbd\u8fbe7.2 TB\/s\u3002\u6bcf\u4e2aSwitch Tray\u96c6\u62102\u9897\u82af\u7247\uff0c\u603b\u5e26\u5bbd\u8fbe14.4 TB\/s\uff0c\u5e76\u5185\u5efaSHARP\uff08Scalable Hierarchical Aggregation and Reduction Protocol\uff09\u8ba1\u7b97\u5f15\u64ce\uff0c\u5177\u59073.6 TFLOPS\u805a\u5408\u6027\u80fd\uff0c\u5b9e\u73b0\u901a\u4fe1\u8ba1\u7b97\u878d\u5408\u3002<\/p>\n\n\n\n

5.4 Spectrum-X\u4e0eRoCE\u4f18\u5316\uff1a\u4e3aAI\u8bbe\u8ba1\u7684\u4ee5\u592a\u7f51\u7ed3\u6784<\/h3>\n\n\n\n



<\/h3>\n\n\n\n

Spectrum-X\u662f\u5168\u7403\u9996\u4e2a\u4e13\u4e3aAI\u8d1f\u8f7d\u8bbe\u8ba1\u7684\u4ee5\u592a\u7f51\u7f51\u7edc\u7ed3\u6784\uff0c\u91c7\u7528Spectrum-4\u4ea4\u6362\u82af\u7247\u4e0eBlueField-3 DPU\uff0c\u5177\u5907128 x 400G\u621664 x 800G\u7aef\u53e3\u300151.2 Tbps\u603b\u5e26\u5bbd\u548c1000\u4ebf\u6676\u4f53\u7ba1\u96c6\u6210\u5ea6\u3002\u7cfb\u7edf\u5185\u5d4cRoCE\uff08RDMA over Converged Ethernet\uff09\u81ea\u9002\u5e94\u8def\u7531\u3001\u62e5\u585e\u63a7\u5236\u4e0e\u591a\u79df\u6237\u9694\u79bb\u673a\u5236\uff0c\u5c06\u6709\u6548\u5e26\u5bbd\u5229\u7528\u7387\u4ece60%\u63d0\u5347\u81f395%\u3002<\/p>\n\n\n\n

5.5 GB200 NVL72\u7cfb\u7edf\u89e3\u6784\uff1a36 Grace CPU + 72 Blackwell GPU\u7684\u8ba1\u7b97\u5f15\u64ce<\/h3>\n\n\n\n



<\/h3>\n\n\n\n

GB200 NVL72\u6784\u6210Blackwell\u5e73\u53f0\u7684\u6838\u5fc3\u8ba1\u7b97\u5355\u5143\uff0c\u75315\u4e2aCompute Tray\u30019\u4e2aNVLink Switch Tray\u7ec4\u6210\uff0c\u96c6\u621036\u9897Grace CPU\u4e0e72\u9897Blackwell GPU\uff0c\u5177\u5907720 PFLOPS\u8bad\u7ec3\u6027\u80fd\u4e0e1440 PFLOPS\u63a8\u7406\u6027\u80fd\uff0c\u5168\u9762\u652f\u649127\u4e07\u4ebf\u53c2\u6570\u89c4\u6a21\u7684\u5927\u6a21\u578b\u8fd0\u884c\u3002\u7cfb\u7edf\u4e92\u8054\u603b\u5e26\u5bbd\u8fbe130 TB\/s\uff0cAll-Reduce\u80fd\u529b\u8fbe260 TB\/s\uff0c\u4e3a\u8d85\u5927\u6a21\u578b\u63d0\u4f9b\u7ebf\u6027\u53ef\u6269\u5c55\u67b6\u6784\u57fa\u7840\u3002<\/p>\n\n\n\n

5.6 \u67b6\u6784\u7ea7\u63a8\u7406\u4f18\u52bf\uff1a30x\u541e\u5410\u91cf\u300125x\u80fd\u6548\u63d0\u5347\u3001MoE\u4e13\u5bb6\u5e76\u884c\u8c03\u5ea6<\/h3>\n\n\n\n



<\/h3>\n\n\n\n

\u901a\u8fc7\u91c7\u7528\u4e13\u5bb6\u5e76\u884c\uff08Expert Parallel\uff09\u6280\u672f\uff0cBlackwell\u5e73\u53f0\u53ef\u5c06\u4e00\u4e2a\u5b8c\u6574\u7684\u4e13\u5bb6\u6a21\u578b\u90e8\u7f72\u81f3\u5355\u9897GPU\u6267\u884c\uff0c\u7ed3\u5408\u9ad8\u5e26\u5bbdNVLink\u4e92\u8054\u5b9e\u73b0\u4f4e\u5ef6\u8fdf\u5206\u5e03\u5f0f\u8c03\u5ea6\u3002\u5bf9\u6bd4H100\u5e73\u53f0\uff0cGB200 NVL72\u57281.8T\u53c2\u6570GPT-MoE\u6a21\u578b\u4e0a\u7684\u63a8\u7406\u541e\u5410\u63d0\u534730\u500d\uff0c\u80fd\u6548\u63d0\u534725\u500d\uff0cTCO\u663e\u8457\u964d\u4f4e\uff0c\u6ee1\u8db3\u5b9e\u65f6\u54cd\u5e94\u4e0e\u6210\u672c\u63a7\u5236\u7684\u53cc\u91cd\u76ee\u6807\u3002<\/strong>\u516d\u3001\u67b6\u6784\u5e76\u884c\u7b56\u7565\u4e0e\u6a21\u578b\u8c03\u5ea6\u673a\u5236<\/strong><\/p>\n\n\n\n



<\/h2>\n\n\n\n

6.1 \u591a\u7ef4\u5e76\u884c\u7b56\u7565\uff1aTensor \/ Pipeline \/ Expert \/ Data<\/h3>\n\n\n\n



<\/h3>\n\n\n\n

\u9762\u5bf9\u8d85\u5927\u89c4\u6a21\u6a21\u578b\u7684\u8bad\u7ec3\u4e0e\u63a8\u7406\u9700\u6c42\uff0cBlackwell\u5e73\u53f0\u63d0\u4f9b\u591a\u7ef4\u5e76\u884c\u8c03\u5ea6\u7b56\u7565\uff0c\u5305\u62ec\u5f20\u91cf\u5e76\u884c\uff08Tensor Parallel\uff09\u3001\u6d41\u6c34\u5e76\u884c\uff08Pipeline Parallel\uff09\u3001\u4e13\u5bb6\u5e76\u884c\uff08Expert Parallel\uff09\u4e0e\u6570\u636e\u5e76\u884c\uff08Data Parallel\uff09\uff0c\u4ee5\u7075\u6d3b\u9002\u914d\u4e0d\u540c\u8ba1\u7b97\u5bc6\u5ea6\u4e0e\u5ef6\u8fdf\u654f\u611f\u6027\u7684\u5e94\u7528\u8d1f\u8f7d\u3002\u5176\u4e2d\uff0c\u591a\u7b56\u7565\u7ec4\u5408\uff08\u5982TP4.EP16\u3001TP2.EP8.DP4\uff09\u5728\u541e\u5410\u6548\u7387\u4e0e\u4ea4\u4e92\u6027\u80fd\u4e4b\u95f4\u5b9e\u73b0\u52a8\u6001\u6743\u8861\u3002<\/p>\n\n\n\n

6.2 Expert Parallel\u4e3a\u4e3b\u7684\u8c03\u5ea6\u8def\u5f84\uff1aBlackwell\u4e0b\u7684\u63a8\u7406\u9996\u9009<\/h3>\n\n\n\n



<\/h3>\n\n\n\n

Blackwell\u901a\u8fc7\u5f3a\u5927\u7684\u7247\u4e0a\u8ba1\u7b97\u80fd\u529b\u4e0e\u9ad8\u6548NVLink\u5e26\u5bbd\uff0c\u4f7f\u5f97\u6bcf\u9897GPU\u53ef\u4ee5\u5bb9\u7eb3\u5b8c\u6574\u201c\u4e13\u5bb6\u6a21\u5757\u201d\uff08Expert\uff09\uff0c\u65e0\u9700\u8de8GPU\u5207\u5206\u6267\u884c\uff0c\u663e\u8457\u51cf\u5c11\u901a\u4fe1\u5ef6\u8fdf\u548c\u6fc0\u6d3b\u91cd\u7ec4\u6210\u672c\u3002\u76f8\u6bd4Hopper\u4ecd\u9700\u4f9d\u8d56Tensor Parallel\u6280\u672f\uff0cBlackwell\u4ee5\u4e13\u5bb6\u5e76\u884c\u4e3a\u4e3b\u7b56\u7565\uff0c\u5927\u5e45\u63d0\u5347\u4e86\u5927\u6a21\u578b\u5b9e\u65f6\u63a8\u7406\u6027\u80fd\uff0c\u5c24\u5176\u9002\u7528\u4e8eMixture-of-Experts\u7ed3\u6784\u5982GPT-MoE-1.8T\u3002<\/p>\n\n\n\n

6.3 \u63a8\u7406\u5ef6\u8fdf\u63a7\u5236\uff1aTTL\u4e0eFTL\u6027\u80fd\u6307\u6807<\/h3>\n\n\n\n



<\/h3>\n\n\n\n

Blackwell\u5e73\u53f0\u5728\u7aef\u5230\u7aef\u63a8\u7406\u5ef6\u8fdf\uff08Token-to-Token Latency, TTL\uff09\u63a7\u5236\u65b9\u9762\u53d6\u5f97\u7a81\u7834\uff1a\u5728GPT-3\uff08175B\uff09\u6a21\u578b\u4e0a\uff0c\u5c06\u9996\u6b21Token\u5ef6\u8fdf\uff08First Token Latency, FTL\uff09\u4ece2\u79d2\u964d\u81f350\u6beb\u79d2\u4ee5\u5185\uff1b\u57281.8T\u6a21\u578b\u4e0a\uff0c\u8f93\u5165\u5e8f\u5217\u957f\u5ea6\u8fbe32,768\u3001\u8f93\u51fa\u5e8f\u52171,024\u7684\u524d\u63d0\u4e0b\uff0c\u4e5f\u5c06FTL\u4ece5\u79d2\u663e\u8457\u7f29\u77ed\u3002\u9ad8\u9891\u4f4e\u5ef6\u8fdf\u54cd\u5e94\u80fd\u529b\u4e3aLLM\u5728\u5b9e\u65f6\u5bf9\u8bdd\u3001\u4ee3\u7801\u751f\u6210\u7b49\u573a\u666f\u63d0\u4f9b\u7cfb\u7edf\u7ea7\u4fdd\u969c\u3002<\/p>\n\n\n\n

6.4 \u7b97\u6cd5\u8c03\u5ea6\u5bf9\u541e\u5410\u4e0e\u6210\u672c\u7684\u975e\u7ebf\u6027\u63d0\u5347<\/h3>\n\n\n\n



<\/h3>\n\n\n\n

\u901a\u8fc7\u7075\u6d3b\u9009\u62e9\u5e76\u884c\u7b56\u7565\u7ec4\u5408\uff0cBlackwell\u5e73\u53f0\u5728\u6bcf\u7528\u6237\u541e\u5410\u91cf\uff08Tokens per Second per User\uff09\u7ef4\u5ea6\u5c55\u73b0\u663e\u8457\u4f18\u52bf\uff1aGB200 FP4\u914d\u7f6e\u5728GPT-MoE-1.8T\u63a8\u7406\u4efb\u52a1\u4e2d\u8fbe\u5230\u8d85\u8fc7150 Tokens\/s\uff0c\u76f8\u6bd4H200\u6216B200 FP8\u63d0\u5347\u8fd130\u500d\u3002\u540c\u65f6\uff0c\u5f97\u76ca\u4e8eFP4\u4f4e\u7cbe\u5ea6\u8ba1\u7b97\u3001\u4e13\u5bb6\u7ea7\u8c03\u5ea6\u4e0e\u5168NVLink\u4e92\u8054\uff0c\u6574\u4f53\u7cfb\u7edf\u80fd\u6548\u63d0\u534725\u500d\uff0c\u663e\u8457\u964d\u4f4e\u8bad\u7ec3\u4e0e\u63a8\u7406TCO\u3002<\/p>\n\n\n\n

\u4e03\u3001\u8def\u7ebf\u56fe\u4e0e\u4ea7\u4e1a\u5f71\u54cd\uff1aBlackwell\u4e4b\u540e\u7684\u672a\u6765<\/h2>\n\n\n\n



<\/h2>\n\n\n\n

7.1 GPU\u6f14\u8fdb\u8def\u5f84\uff1a\u4eceHopper\u5230Blackwell\u518d\u5230Rubin<\/h3>\n\n\n\n



<\/h3>\n\n\n\n

NVIDIA GPU\u8def\u7ebf\u56fe\u663e\u793a\uff0cBlackwell\u662f\u7ee7Hopper\u4e4b\u540e\u7684\u4e0b\u4e00\u4ee3\u6838\u5fc3\u5e73\u53f0\uff0c\u57fa\u4e8eTSMC 4NP\u5236\u7a0b\uff0c\u91c7\u75288\u5806HBM3e\u9ad8\u5e26\u5bbd\u5185\u5b58\uff08HBM\uff09\u5c01\u88c5\uff0c\u5e26\u6765\u6570\u91cf\u7ea7\u7684AI FLOPS\u63d0\u5347\u3002\u4e0b\u4e00\u4ee3Rubin\u67b6\u6784\u5c06\u8fdb\u4e00\u6b65\u96c6\u621012\u5806HBM4\uff0c\u5e76\u5f15\u5165Rubin Ultra GPU\uff0c\u9884\u8ba1\u4f7f\u7528\u66f4\u9ad8\u9636\u7684\u5236\u9020\u5de5\u827a\u4e0e\u4e92\u8fde\u65b9\u6848\uff0c\u5ef6\u7eedNVIDIA\u5728AI\u57fa\u7840\u786c\u4ef6\u9886\u57df\u7684\u6280\u672f\u9886\u5148\u3002<\/p>\n\n\n\n

7.2 \u7f51\u7edc\u4e92\u8054\u6f14\u8fdb\uff1aNVLink 5\u21926\uff0cSpectrum X800\u2192Ultra<\/h3>\n\n\n\n



<\/h3>\n\n\n\n

\u7f51\u7edc\u4e0e\u4e92\u8054\u65b9\u9762\uff0cNVIDIA\u6b63\u6301\u7eed\u63a8\u52a8\u5e26\u5bbd\u63d0\u5347\u4e0e\u534f\u8bae\u4f18\u5316\u3002Blackwell\u6240\u91c7\u7528\u7684NVLink 5\u63d0\u4f9b\u9ad8\u8fbe1800 GB\/s\u5e26\u5bbd\uff0c\u4e0b\u4e00\u4ee3NVLink 6\u5c06\u5b9e\u73b03600 GB\/s\u4e92\u8fde\u901f\u7387\uff1bSuperNIC\u4eceCX7\u5347\u7ea7\u81f3CX8\uff0c\u5e76\u89c4\u5212\u5411CX9\u8fc7\u6e21\uff0c\u541e\u5410\u7387\u5c06\u8fbe\u52301600 Gb\/s\u3002\u4ee5Spectrum-X800\u4e3a\u4ee3\u8868\u7684\u4ee5\u592a\u7f51\u7ed3\u6784\u4e5f\u5c06\u5728\u4e0b\u4e00\u4ee3Ultra\u67b6\u6784\u4e2d\u5b9e\u73b0\u66f4\u9ad8\u96c6\u6210\u5ea6\u3001\u66f4\u5e7f\u8fde\u63a5\u8303\u56f4\u3002<\/p>\n\n\n\n

7.3 \u5e73\u53f0\u5316\u8d8b\u52bf\uff1a\u4e00\u4f53\u5316\u7b97\u529b\u5e73\u53f0\u6784\u5efa\u884c\u4e1a\u95e8\u69db<\/h3>\n\n\n\n



<\/h3>\n\n\n\n

Blackwell\u4ee3\u8868\u7684\u4e0d\u4ec5\u662fGPU\u7684\u8fdb\u5316\uff0c\u66f4\u662f\u201c\u6570\u636e\u4e2d\u5fc3\u7ea7\u5e73\u53f0\u201d\u7684\u843d\u5730\u3002\u5176\u5c06GPU\u3001CPU\u3001DPU\u3001NIC\u3001\u4ea4\u6362\u82af\u7247\u4e0eAI\u4f18\u5316\u8f6f\u4ef6\u6808\u96c6\u6210\u4e3a\u7edf\u4e00\u5e73\u53f0\uff0c\u6253\u9020\u51fa\u9ad8\u5ea6\u4e00\u4f53\u5316\u3001\u5f02\u6784\u53ef\u6269\u5c55\u7684AI\u57fa\u7840\u8bbe\u65bd\u3002\u8fd9\u79cd\u5e73\u53f0\u5316\u67b6\u6784\u5bf9\u901a\u7528\u8ba1\u7b97\u5e73\u53f0\u63d0\u51fa\u66f4\u9ad8\u51c6\u5165\u95e8\u69db\uff0c\u6b63\u5728\u91cd\u5851\u4e91\u670d\u52a1\u5546\u3001AI\u6a21\u578b\u5f00\u53d1\u5546\u4e0e\u7cfb\u7edf\u96c6\u6210\u5546\u7684\u7ade\u4e89\u683c\u5c40\u3002<\/p>\n\n\n\n

7.4 \u5bf9\u4e91\u670d\u52a1\u5546\u3001\u6a21\u578b\u5f00\u53d1\u8005\u548c\u7cfb\u7edf\u96c6\u6210\u5546\u7684\u7cfb\u7edf\u5f71\u54cd<\/h3>\n\n\n\n



<\/h3>\n\n\n\n

Blackwell\u5e73\u53f0\u7684\u80fd\u529b\u91ca\u653e\u5c06\u6df1\u523b\u5f71\u54cd\u4e09\u7c7b\u6838\u5fc3\u73a9\u5bb6\uff1a<\/p>\n\n\n\n

    \n
  • \u4e91\u670d\u52a1\u5546<\/strong>\uff1a\u53ef\u57fa\u4e8eGB200 NVL72\u7b49\u7cfb\u7edf\u6784\u5efa\u201c\u5373\u63d2\u5373\u7528\u201d\u5f0f\u63a8\u7406\u4e0e\u8bad\u7ec3\u96c6\u7fa4\uff0c\u5927\u5e45\u63d0\u9ad8\u6570\u636e\u4e2d\u5fc3\u5355\u4f4d\u9762\u79efAI\u7b97\u529b\uff1b<\/li>\n\n\n\n
  • \u6a21\u578b\u5f00\u53d1\u8005<\/strong>\uff1a\u5f97\u76ca\u4e8eFP4\/FP6\u7b49\u7cbe\u5ea6\u683c\u5f0f\u4e0eQuasar\u91cf\u5316\u7cfb\u7edf\uff0c\u5f00\u53d1\u8005\u53ef\u5728\u4fdd\u7cbe\u5ea6\u7684\u540c\u65f6\u5b9e\u73b0\u66f4\u9ad8\u6548\u7387\u7684\u8bad\u7ec3\u4e0e\u63a8\u7406\u90e8\u7f72\uff1b<\/li>\n\n\n\n
  • \u7cfb\u7edf\u96c6\u6210\u5546<\/strong>\uff1aBlackwell\u5e73\u53f0\u63d0\u4f9b\u6807\u51c6\u5316Compute Tray\u4e0eSwitch Tray\u6a21\u5757\uff0c\u652f\u6301\u89c4\u6a21\u7ea7\u7075\u6d3b\u90e8\u7f72\uff0c\u6210\u4e3a\u6784\u5efa\u4e0b\u4e00\u4ee3AI\u8d85\u7ea7\u8ba1\u7b97\u5e73\u53f0\u7684\u6838\u5fc3\u7ec4\u4ef6\u3002<\/li>\n<\/ul>\n\n\n\n

    Blackwell\u4e0d\u4ec5\u4ee3\u8868\u5f53\u4ee3\u6700\u5f3aAI\u7b97\u529b\u96c6\u7fa4\uff0c\u5176\u6240\u5f15\u9886\u7684\u5e73\u53f0\u5316\u8303\u5f0f\u5c06\u6210\u4e3a\u672a\u67655\u5e74\u5185\u6574\u4e2a\u52a0\u901f\u8ba1\u7b97\u4ea7\u4e1a\u7684\u5e95\u5c42\u7ed3\u6784\u53d8\u9769\u6838\u5fc3\u3002<\/p>\n\n\n\n

    <\/p>\n\n\n\n

    \u516b\u3001\u7ed3\u8bed\uff1aBlackwell\u7684\u4ea7\u4e1a\u7ea7\u610f\u4e49\u4e0e\u6218\u7565\u5c55\u671b<\/strong><\/p>\n\n\n\n



    <\/h2>\n\n\n\n

    Blackwell\u5e73\u53f0\u4ee3\u8868\u4e86NVIDIA\u5728\u751f\u6210\u5f0fAI\u548c\u52a0\u901f\u8ba1\u7b97\u9886\u57df\u7684\u4e00\u6b21\u5168\u6808\u5f0f\u8dc3\u8fc1\u3002\u4ece\u5355\u82af\u7247\u8ba1\u7b97\u67b6\u6784\u8fc8\u5411\u6570\u636e\u4e2d\u5fc3\u7ea7\u7cfb\u7edf\u5e73\u53f0\uff0cBlackwell\u6574\u5408\u4e86GPU\u3001CPU\u3001DPU\u3001NIC\u3001\u4ea4\u6362\u82af\u7247\u4e0eAI\u8f6f\u4ef6\u6808\uff0c\u6784\u5efa\u8d77\u4e00\u4e2a\u5b8c\u6574\u7684\u5f02\u6784\u8ba1\u7b97\u751f\u6001\u3002\u8fd9\u79cd\u67b6\u6784\u4e0d\u4ec5\u5728\u6027\u80fd\u4e0e\u80fd\u6548\u5c42\u9762\u5b9e\u73b0\u8de8\u4ee3\u5f0f\u7a81\u7834\uff0c\u66f4\u5728\u5e73\u53f0\u53ef\u6269\u5c55\u6027\u4e0e\u7cfb\u7edf\u6574\u5408\u80fd\u529b\u4e0a\u6811\u7acb\u884c\u4e1a\u65b0\u57fa\u51c6\u3002<\/p>\n\n\n\n

    Quasar\u91cf\u5316\u7cfb\u7edf\u901a\u8fc7\u8f6f\u786c\u4ef6\u534f\u540c\u8bbe\u8ba1\uff0c\u5c06FP4\/FP6\u7b49\u4f4e\u7cbe\u5ea6\u8ba1\u7b97\u65b9\u5f0f\u63a8\u5411\u5b9e\u7528\u5316\uff0c\u5728\u4fdd\u969c\u6a21\u578b\u7cbe\u5ea6\u7684\u540c\u65f6\u663e\u8457\u964d\u4f4e\u80fd\u8017\u4e0e\u63a8\u7406\u5ef6\u8fdf\uff0c\u4e3aAI\u5927\u6a21\u578b\u90e8\u7f72\u63d0\u4f9b\u65b0\u8def\u5f84\u3002\u4ee5GB200 NVL72\u4e3a\u4ee3\u8868\u7684\u9ad8\u5bc6\u5ea6\u7cfb\u7edf\uff0c\u53ef\u652f\u6301\u4e07\u4ebf\u53c2\u6570\u7ea7\u6a21\u578b\u7684\u5b9e\u65f6\u63a8\u7406\uff0c\u5b9e\u73b0\u541e\u5410\u91cf\u63d0\u534730\u500d\u3001\u80fd\u6548\u63d0\u534725\u500d\uff0c\u4e3a\u4e91\u8ba1\u7b97\u3001\u79d1\u7814\u3001\u81ea\u52a8\u9a7e\u9a76\u3001\u6570\u5b57\u4eba\u7b49\u9886\u57df\u63d0\u4f9b\u7b97\u529b\u57fa\u7840\u3002<\/p>\n\n\n\n

    \u4ece\u8def\u7ebf\u56fe\u89c6\u89d2\u770b\uff0cBlackwell\u662fNVIDIA\u9762\u5411Rubin\u65f6\u4ee3\u7684\u8fc7\u6e21\u8282\u70b9\uff0c\u4e5f\u662f\u5176\u5e73\u53f0\u5316\u6218\u7565\u7684\u96c6\u5927\u6210\u8005\u3002\u4e00\u4f53\u5316\u3001\u7cfb\u7edf\u7ea7\u3001\u4f4e\u7cbe\u5ea6\u3001\u53ef\u6269\u5c55\u7684Blackwell\u67b6\u6784\uff0c\u4e0d\u4ec5\u652f\u6491\u5f53\u4e0b\u5927\u6a21\u578b\u63a8\u7406\u4e3b\u6d41\u9700\u6c42\uff0c\u66f4\u4e3a\u672a\u6765\u6570\u5e74AI\u57fa\u7840\u8bbe\u65bd\u7684\u6f14\u5316\u5960\u5b9a\u4e86\u6280\u672f\u8303\u5f0f\u3002<\/p>\n\n\n\n

    \u8fd9\u4e0d\u662f\u4e00\u9897GPU\u7684\u80dc\u5229\uff0c\u800c\u662f\u4e00\u4e2a\u8ba1\u7b97\u5e73\u53f0\u7684\u91cd\u6784\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

    \u5f15\u8a00\u5728\u751f\u6210\u5f0fAI\u548c\u52a0\u901f\u8ba1\u7b97\u8fc8\u5165\u201c\u4e07\u4ebf\u53c2\u6570\u201d\u65f6\u4ee3\u7684\u5173\u952e\u8282\u70b9\uff0cNVIDIA\u63a8\u51faBlackwell\u5e73\u53f0\uff0c\u4ee5\u5168\u6808\u5f0f\u67b6\u6784 […]<\/p>\n","protected":false},"author":1,"featured_media":676,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[7],"tags":[],"class_list":["post-669","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-7"],"_links":{"self":[{"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/posts\/669","targetHints":{"allow":["GET"]}}],"collection":[{"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/comments?post=669"}],"version-history":[{"count":1,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/posts\/669\/revisions"}],"predecessor-version":[{"id":678,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/posts\/669\/revisions\/678"}],"wp:featuredmedia":[{"embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/media\/676"}],"wp:attachment":[{"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/media?parent=669"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/categories?post=669"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/tags?post=669"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}