{"id":14032,"date":"2025-12-18T16:14:38","date_gmt":"2025-12-18T08:14:38","guid":{"rendered":"https:\/\/www.killtest.com.tw\/?p=14032"},"modified":"2025-12-18T16:14:38","modified_gmt":"2025-12-18T08:14:38","slug":"%e9%9d%a2%e5%90%91ai%e9%ab%98%e6%95%88%e8%83%bd%e7%b6%b2%e8%b7%af%e9%97%9c%e9%8d%b5%e6%8a%80%e8%a1%93%ef%bc%9a%e8%b6%85%e7%af%80%e9%bb%9e%ef%bc%88superpod%ef%bc%89","status":"publish","type":"post","link":"https:\/\/www.killtest.com.tw\/?p=14032","title":{"rendered":"\u9762\u5411AI\u9ad8\u6548\u80fd\u7db2\u8def\u95dc\u9375\u6280\u8853\uff1a\u8d85\u7bc0\u9ede\uff08SuperPod\uff09"},"content":{"rendered":"\n

\u8d85\u7bc0\u9ede(SuperPod)\u662f\u8fd1\u5e74\u4f86\u70ba\u61c9\u5c0dAI\u5927\u6a21\u578b\u8a13\u7df4\u8207\u63a8\u7406\u9700\u6c42\u800c\u767c\u5c55\u8d77\u4f86\u7684\u65b0\u578b\u7b97\u529b\u57fa\u790e\u8a2d\u65bd\u67b6\u69cb\uff0c\u5b83\u900f\u904e\u9ad8\u901f\u4e92\u9023\u6280\u8853\u5c07\u5927\u91cf\u904b\u7b97\u55ae\u5143(\u5982GPU\u3001TPU\u3001NPU\u7b49)\u7dca\u5bc6\u6574\u5408\uff0c\u5efa\u69cb\u4e00\u500b\u9ad8\u983b\u5bec\u57df(HBD)\u3002
\u5177\u9ad4\u4f86\u8aaa\uff0c\u8d85\u7bc0\u9ede\u662f\u6307\u5728\u4e00\u500b\u5be6\u9ad4\u6a5f\u6ac3\u6216\u4e00\u7d44\u7dca\u5bc6\u8026\u5408\u7684\u8a08\u7b97\u55ae\u5143\u5167\u90e8\uff0c\u900f\u904e\u9ad8\u5bc6\u5ea6\u6574\u5408\u8a08\u7b97\u55ae\u5143\u548c\u5c08\u7528\u7684\u9ad8\u901f\u4e92\u806f\u6280\u8853\uff0c\u5be6\u73fe\u8fd1\u4f3c\u55ae\u6a5f\u6027\u80fd\u7684\u8d85\u5927\u898f\u6a21\u5e73\u884c\u8a08\u7b97\u7cfb\u7d71\u3002<\/p>\n\n\n\n

\u5b83\u65e8\u5728\u7a81\u7834\u50b3\u7d71\u4f3a\u670d\u5668\u5167\u90e8\u4ee5\u53ca\u4f3a\u670d\u5668\u4e4b\u9593\u900f\u904ePCle\u6216\u6a19\u6e96\u4e59\u592a\u7db2\u8def\u4e92\u806f\u7684\u983b\u5bec\u548c\u5ef6\u9072\u74f6\u9838\uff0c\u5c07\u6578\u5341\u751a\u81f3\u6578\u767e\u500b\u52a0\u901f\u5668\u7dca\u5bc6\u9023\u63a5\uff0c\u5f62\u6210\u4e00\u500b\u908f\u8f2f\u4e0a\u7684\u8d85\u5927\u4f3a\u670d\u5668\uff0c\u4ee5\u652f\u63f4\u5f35\u91cf\u4e26\u884c\u3001\u5c08\u5bb6\u4e26\u884c\u7b49\u5c0d\u5167\u90e8\u901a\u8a0a\u8981\u6c42\u6975\u9ad8\u7684\u5e73\u884c\u8a08\u7b97\u4efb\u52d9\u3002
\u8d85\u7bc0\u9ede\u904b\u7b97\u67b6\u69cb\u7684\u95dc\u9375\u7279\u5fb5\u5305\u62ec\uff1a
\u25cf\u9ad8\u5bc6\u5ea6\u7b97\u529b\u6574\u5408\uff1a\u5728\u6709\u9650\u7a7a\u9593\u5167\u6574\u5408\u5927\u91cfGPU\u6216\u5176\u4ed6AI\u52a0\u901f\u5668\uff0c\u63d0\u4f9b\u6975\u81f4\u7684\u904b\u7b97\u5bc6\u5ea6\u3002
\u25cf\u9ad8\u901f\u4e92\u806f\uff1a\u63a1\u7528NVLink\u3001InfiniBand\u7b49\u9ad8\u901f\u4e92\u806f\u6280\u8853\uff0c\u5be6\u73feGPU\u4e4b\u9593\u4ee5\u53caGPU\u8207\u7db2\u8def\u4e4b\u9593\u7684\u9ad8\u983b\u5bec\u3001\u4f4e\u5ef6\u9072\u901a\u4fe1\uff0c\u6d88\u9664\u8cc7\u6599\u50b3\u8f38\u74f6\u9838\u3002
\u25cf\u7b97\u529b\u8207\u7db2\u8def\u6df1\u5ea6\u878d\u5408\uff1a\u7db2\u8def\u4e0d\u518d\u53ea\u662f\u8cc7\u6599\u7684\u50b3\u8f38\u901a\u9053\uff0c\u800c\u662f\u8207\u904b\u7b97\u7dca\u5bc6\u7d50\u5408\uff0c\u5be6\u73fe\u7db2\u8def\u611f\u77e5\u904b\u7b97\u3001\u7db2\u8def\u878d\u5408\u904b\u7b97\uff0c\u751a\u81f3\u904b\u7b97\u91cd\u5851\u7db2\u8def\u3002\u4f8b\u5982\uff0c\u5728\u8d85\u7bc0\u9ede\u5167\u90e8\uff0c\u900f\u904e\u5f15\u5165\u7bc0\u9ede\u5167\u4ea4\u63db\u6676\u7247\uff0c\u589e\u5f37\u5361\u9593P2P\u983b\u5bec\uff0c\u6709\u6548\u63d0\u5347\u7bc0\u9ede\u5167\u7db2\u8def\u50b3\u8f38\u6548\u7387\u3002
\u25cf\u7d71\u4e00\u8cc7\u6e90\u7ba1\u7406\u8207\u8abf\u5ea6\uff1a\u5be6\u73fe\u904b\u7b97\u3001\u5132\u5b58\u8207\u7db2\u8def\u8cc7\u6e90\u7684\u7d71\u4e00\u7d0d\u7ba1\u8207\u878d\u5408\u8def\u7531\u8abf\u5ea6\uff0c\u63d0\u5347\u8cc7\u6e90\u5229\u7528\u7387\u8207\u7ba1\u7406\u6548\u7387\u3002
\u76ee\u524d\u696d\u754c\u5178\u578b\u7684\u8d85\u7bc0\u9ede\u65b9\u6848\u5305\u62ec\uff1a
(\u4e00)\u82f1\u5049\u9054<\/a>DGX SuperPOD(\u4ee5NVL72\u70ba\u4f8b)\u82f1\u5049\u9054\u4f5c\u70baAI\u52a0\u901f\u9818\u57df\u7684\u9818\u5c0e\u8005\uff0c\u5176DGX SuperPOD\u7cfb\u5217\u662f\u696d\u754c\u5ee3\u6cdb\u63a1\u7528\u7684AI\u8d85\u7d1a\u8a08\u7b97\u5e73\u53f0\u3002
\u5176\u4e2d\uff0cGB200 NVL72 SuperNode\u662f\u5176\u6700\u65b0\u7684\u4ee3\u8868\u6027\u7522\u54c1\u4e4b\u4e00\u3002 GB200 NVL72 SuperNode\u5c0736\u500bGrace CPU\u548c72\u500bBlackwell GPU\u6574\u5408\u5230\u4e00\u500b\u6db2\u51b7\u6a5f\u6ac3\u4e2d\u3002
\u5b83\u63a1\u7528\u300cGPU-GPU NVLink ScaleUp+Node-Node RDMAScaleOut\u300d\u7684\u4e92\u806f\u65b9\u5f0f\u3002 \u25cf Compute Tray:\u6574\u500b\u7cfb\u7d71\u5305\u62ec18\u500bCompute Tray,\u6bcf\u500bCompute Tray\u5305\u542b2\u500bGB200\u8d85\u7d1a\u6676\u7247\uff0c\u6bcf\u500bGB200\u8d85\u7d1a\u6676\u7247\u53c8\u5305\u542b2\u500bBlackwell B200 GPU\u548c1\u9846Grace CPU,\u6574\u500b\u6a5f\u6ac3\u517172\u500bB200 GPU\u548c36\u500bGrace CPU\u3002
\u900f\u904eNVLink\u8207NVLink-C2C\u6280\u8853\uff0c\u5be6\u73feGPU\u4e4b\u9593\u4ee5\u53caGPU\u8207CPU\u4e4b\u9593\u7684\u9ad8\u901f\u8a18\u61b6\u9ad4\u5171\u4eab\u8207\u8cc7\u6599\u50b3\u8f38\u3002\u55ae\u4e00Compute Tray\u63d0\u4f9b7.2TB\/s(\u55ae\u541128.8Tb\/s)\u983b\u5bec\uff0cNVL72\u6574\u6a5f\u6ac3\u7684Compute Tray\u63d0\u4f9b129.6TB\/S\u7684NVLink\u983b\u5bec\u3002
\u25cf Switch Tray:\u5171\u5305\u542b9\u500bSwitch Tray,\u6bcf\u500bSwitch Tray\u5167\u5efa2\u9846NVSwitch\u6676\u7247\uff0c\u6574\u500b\u6a5f\u6ac3\u63d0\u4f9b18\u500bNVLink Switch\u6676\u7247\u3002\u6574\u6a5f\u6ac3\u5f8c\u90e8\u900f\u904e\u7dda\u7e9c\u5c07ComputeTray\u548cSwitch Tray\u9032\u884c\u4e92\u806f\u3002
\u55ae\u4e00Switch Tray\u63d0\u4f9b14.4TB\/s(\u55ae\u541157.6Tb\/s)\u983b\u5bec\uff0cNVL72\u6574\u6a5f\u6ac3\u7684Switch Tray\u63d0\u4f9b129.6TB\/s\u7684NVLink\u983b\u5bec\u3002\u9019\u6a23\u8d85\u7bc0\u9ede\u6574\u6a5f\u6ac3Compute Tray\u7684GPU\u548cSwitch Tray\u7684\u4ea4\u63db\u6676\u7247\u4e4b\u9593\u5c31\u53ef\u4ee5\u5168\u9023\u63a5\u3002
\u25cf Scale Up:NVL72\u5167\u90e8\u63a1\u7528NVLink5\u548cNVSwitch\u5efa\u69cbScale Up\u7db2\u7d61\uff0c\u63d0\u4f9b\u6975\u9ad8\u7684\u983b\u5bec(\u6bcf\u500bCompute Tray\u900f\u904eNVLink\/NVSwitch\u5177\u67097.2TB\/s\u7684ScaleUp\u9023\u63a5\u983b\u5bec)\u548c\u8d85\u4f4e\u6642\u5ef6(\u9285\u96fb\u7e9c\u9023\u63a5\u7bc0\u7701\u4e86\u5149\u6a21\u7d44\u5f15\u5165\u7684\u5ef6\u9072)\u3002
\u6240\u6709GPU\u53ef\u4ee5\u5b58\u53d6\u6574\u500b\u8d85\u7bc0\u9ede\u5176\u4ed6GPU\u7684HBM\u8a18\u61b6\u9ad4\u548cGrace CPU\u7684DDR\u5167\u5b58\uff0c\u5be6\u73fe\u7d71\u4e00\u8a18\u61b6\u9ad4\u7a7a\u9593\u3002
\u25cf Scale Out:\u900f\u904eCX8800Gbps RNIC\u63a5\u5165InfiniBand RDMA Scale Out\u7db2\u7d61\uff0c\u5be6\u73fe\u591a\u500bNVL72 SuperNode\u7d44\u6210\u66f4\u5927\u898f\u6a21\u7684SuperPOD(\u4f8b\u59828\u500bDGX GB200NVL72\u7d44\u6210\u4e00\u500b\u5305\u542b576\u584aB200 GPU\u7684SuperPOD)\u3002
(\u4e8c)\u83ef\u70baCloudMatrix 384CloudMatrix 384\u662f\u83ef\u70ba\u63a8\u51fa\u7684\u8d85\u5927\u898f\u6a21AI\u8d85\u7bc0\u9ede\u89e3\u6c7a\u65b9\u6848\uff0c\u7531384\u9846\u6607\u9a30910C NPU\u6676\u7247\u900f\u904e\u5168\u9023\u63a5\u62d3\u64b2\u7d50\u69cb\u4e92\u806f\u800c\u6210\u3002\u5b83\u5275\u65b0\u5730\u63d0\u51fa\u4e86\u5c0d\u7b49\u904b\u7b97\u67b6\u69cb\uff0c\u5c07\u532f\u6d41\u6392\u5f9e\u4f3a\u670d\u5668\u5167\u90e8\u64f4\u5c55\u5230\u6574\u6a5f\u6ac3\u751a\u81f3\u8de8\u6a5f\u6ac3\u3002
<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

\u25cf Compute Tray: \u6bcf\u500bCompute Tray\u5305\u542b8\u584a910C NPU,\u4e26\u4e14\u5167\u5efa\u4e867\u500bL1-HCCS-SW\u4ea4\u63db\u6676\u7247(\u4e0a\u806fScale Up)+1\u500bCDR\u4ea4\u63db\u6676\u7247(\u4e0a\u806fScale Out)\u3002
910C NPU\u63a1\u7528Chiplet\u6280\u8853\uff0c\u6574\u5408\u4e862\u9846910B\u548c8\u9846HBM2e\u5167\u5b58\uff0c\u55ae\u5361\u7b97\u529b\u9054\u5230FP16781.25 TFLOPS,\u5167\u5b58\u983b\u5bec3.2TB\/s\u3002
\u6bcf\u5f35NPU\u57fa\u65bcHCCS(High-speedComputing Cluster System)GPU-GPU\u79c1\u6709\u9ad8\u901f\u4e92\u806f\u5354\u5b9a\u4f7f\u75288\u500b\u901a\u8def\u5206\u5225\u9023\u63a5\u5230L1-HCCS-SW\u4ea4\u63db\u6676\u7247\uff0c\u5be6\u73fe\u7121\u963b\u585e\u983b\u5bec\u6536\u6582\u6bd4\u3002 \u25cf Switch Tray: \u6bcf\u500bSwitch Tray(CloudEngine 16800\u4ea4\u63db\u6a5f)\u670916\u500b\u696d\u52d9\u69fd\uff0c\u6bcf\u500b\u696d\u52d9\u69fd\u6700\u5927\u652f\u63f448\u500b400G\u63a5\u53e3\uff0c\u6574\u53f0\u652f\u63f4768\u500b400G\u63a5\u53e3\u3002\u63a1\u7528\u55ae\u5c64\u6241\u5e73\u5316\u62d3\u64b2\uff0c\u5efa\u69cbNPU\u5168\u4e92\u9023(All-to-Al1)\u62d3\u6a38\u7d50\u69cb\uff0c\u6d88\u9664\u50b3\u7d71\u7db2\u8def\u7684\u983b\u5bec\u74f6\u9838\u3002
\u25cf Scale Up:CloudMatrix384\u7684Scale Up\u983b\u5bec\u9ad8\u9054269TB\/s,\u662fNVL72\u76842.1\u500d\u3002\u56e0\u7269\u7406\u8ddd\u96e2\u9650\u5236\uff0c\u63a1\u7528400G\u4f4e\u529f\u8017\u5149\u6a21\u7d44(LPO),\u7701\u7565\u4e86\u50b3\u7d71DSP\u6676\u7247\u4ee5\u964d\u4f4e\u5ef6\u9072\u548c\u529f\u8017\u3002 \u25cf Scale Out:\u63a1\u7528Spine-Leaf 8\u5c0e\u8ecc\u62d3\u64b2\uff0c\u900f\u904e400G\u5149\u6a21\u7d44\u5efa\u69cbScale Out\u7db2\u7d61\uff0c\u5be6\u73fe\u8d85\u7bc0\u9ede\u9593\u7684\u4e92\u806f\uff0c\u7e3d\u983b\u5bec\u662fNVIDIA NVL72\u76845.3\u500d\u3002
(\u4e09)ETH-X\u7531ODCC\u727d\u982d\uff0c\u806f\u5408\u4e2d\u570b\u4fe1\u901a\u9662\u3001\u9a30\u8a0a\u7b49\u55ae\u4f4d\u767c\u8d77\u7684ETH-X\u8a08\u756b\u53ef\u4ee5\u652f\u63f4\u55ae\u4e00\u8d85\u7bc0\u9ede64\u5361\u7684\u904b\u7b97\u80fd\u529b\uff0c\u548c\u82f1\u5049\u9054\u7684\u79c1\u6709NVLink\u65b9\u6848\u4e0d\u540c\uff0cETH-X\u63a1\u7528\u66f4\u958b\u653e\u7684RoCE\u65b9\u6848\u3002
\u25cf Compute Tray:\u6bcf\u500bCompute Tray\u5305\u542b4\u5f35GPU\u548c1\u500bX86 CPU,CPU\u548cGPU\u4e4b\u9593\u900f\u904ePCle Switch\u5c0d\u63a5\u3002\u6574\u500b\u6a5f\u6ac3\u517164\u5f35GPU.\u540c\u6642\u6bcf\u500bCompute Tray\u63d0\u4f9b4\u500bNIC\u7528\u65bcScale Out\u65b9\u5411\u7684\u64f4\u5c55\u3002
\u25cf Switch Tray: \u6bcf\u500bSwitch Tray\u5305\u542b1\u9846\u652f\u63f4RoCE\u7684\u9ad8\u6548\u80fd51.2Tbps\u4e59\u592a\u7db2\u8def\u4ea4\u63db\u6676\u7247\uff0c\u6574\u500b\u6a5f\u6ac3\u63d0\u4f9b8\u500bSwitch\u6676\u7247.GPU\u548cSwitch\u6676\u7247\u652f\u63f4100G serdes\u3002
ETH-X\u6574\u6a5f\u6ac3GPU\u4e92\u806f\u983b\u5bec\u70ba204.8Tbps.8\u500bSwitch Tray\u652f\u63f4409.6Tbps\u7684\u983b\u5bec\uff0c\u4e00\u534a\u7528\u65bc\u8d85\u7bc0\u9ede\u6ac3\u5167\u9023\u63a5GPU,\u53e6\u4e00\u534a\u7684\u983b\u5bec\u7528\u65bc\u80cc\u9760\u80cc\u9023\u63a5\u65c1\u908a\u6a5f\u6ac3\u7684\u8d85\u7bc0\u9ede\u6216\u900f\u904eL2HB Switch\u505a\u66f4\u5927\u7684HBD\u57dfScale Up\u64f4\u5145\u3002
Intel Gaudi3GPU\u63d0\u4f9b4.8Tbps\u7684\u983b\u5bec\uff0c\u6574\u500b\u8d85\u7bc0\u9ede\u6a5f\u6ac3\u9700\u898112\u500bSwitch Tray\u3002 ETH-X\u4e5f\u652f\u63f4Switch Tray\u6c92\u6709\u5916\u90e8Scale Up\u64f4\u5145\u53e3\u7684\u65b9\u6848\uff0c\u6240\u6709serdes\u9023\u63a5\u90fd\u7528\u65bc\u6ac3\u5167\u4e92\u806f\uff0c\u53ea\u9700\u89814\u500b2U\u9ad8\u7684Switch Tray\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

\u8d85\u7bc0\u9ede(SuperPod)\u662f\u8fd1\u5e74\u4f86\u70ba\u61c9\u5c0dAI\u5927\u6a21\u578b\u8a13\u7df4\u8207\u63a8\u7406\u9700\u6c42\u800c\u767c\u5c55\u8d77\u4f86\u7684\u65b0\u578b\u7b97\u529b\u57fa\u790e\u8a2d\u65bd\u67b6\u69cb\uff0c\u5b83\u900f\u904e\u9ad8\u901f\u4e92\u9023\u6280\u8853\u5c07\u5927\u91cf\u904b\u7b97\u55ae\u5143(\u5982GPU\u3001TPU\u3001NPU\u7b49)\u7dca\u5bc6\u6574\u5408\uff0c\u5efa\u69cb\u4e00\u500b\u9ad8\u983b\u5bec\u57df(HBD)\u3002\u5177\u9ad4\u4f86\u8aaa\uff0c\u8d85\u7bc0\u9ede\u662f\u6307\u5728\u4e00\u500b\u5be6\u9ad4\u6a5f\u6ac3\u6216\u4e00\u7d44\u7dca\u5bc6\u8026\u5408\u7684\u8a08\u7b97\u55ae\u5143\u5167\u90e8\uff0c\u900f\u904e\u9ad8\u5bc6\u5ea6\u6574\u5408\u8a08\u7b97\u55ae\u5143\u548c\u5c08\u7528\u7684\u9ad8\u901f\u4e92\u806f\u6280\u8853\uff0c\u5be6\u73fe\u8fd1\u4f3c\u55ae\u6a5f\u6027\u80fd\u7684\u8d85\u5927\u898f\u6a21\u5e73\u884c\u8a08\u7b97\u7cfb\u7d71\u3002 \u5b83\u65e8\u5728\u7a81\u7834\u50b3\u7d71\u4f3a\u670d\u5668\u5167\u90e8\u4ee5\u53ca\u4f3a\u670d\u5668\u4e4b\u9593\u900f\u904ePCle\u6216\u6a19\u6e96\u4e59\u592a\u7db2\u8def\u4e92\u806f\u7684\u983b\u5bec\u548c\u5ef6\u9072\u74f6\u9838\uff0c\u5c07\u6578\u5341\u751a\u81f3\u6578\u767e\u500b\u52a0\u901f\u5668\u7dca\u5bc6\u9023\u63a5\uff0c\u5f62\u6210\u4e00\u500b\u908f\u8f2f\u4e0a\u7684\u8d85\u5927\u4f3a\u670d\u5668\uff0c\u4ee5\u652f\u63f4\u5f35\u91cf\u4e26\u884c\u3001\u5c08\u5bb6\u4e26\u884c\u7b49\u5c0d\u5167\u90e8\u901a\u8a0a\u8981\u6c42\u6975\u9ad8\u7684\u5e73\u884c\u8a08\u7b97\u4efb\u52d9\u3002\u8d85\u7bc0\u9ede\u904b\u7b97\u67b6\u69cb\u7684\u95dc\u9375\u7279\u5fb5\u5305\u62ec\uff1a\u25cf\u9ad8\u5bc6\u5ea6\u7b97\u529b\u6574\u5408\uff1a\u5728\u6709\u9650\u7a7a\u9593\u5167\u6574\u5408\u5927\u91cfGPU\u6216\u5176\u4ed6AI\u52a0\u901f\u5668\uff0c\u63d0\u4f9b\u6975\u81f4\u7684\u904b\u7b97\u5bc6\u5ea6\u3002 \u25cf\u9ad8\u901f\u4e92\u806f\uff1a\u63a1\u7528NVLink\u3001InfiniBand\u7b49\u9ad8\u901f\u4e92\u806f\u6280\u8853\uff0c\u5be6\u73feGPU\u4e4b\u9593\u4ee5\u53caGPU\u8207\u7db2\u8def\u4e4b\u9593\u7684\u9ad8\u983b\u5bec\u3001\u4f4e\u5ef6\u9072\u901a\u4fe1\uff0c\u6d88\u9664\u8cc7\u6599\u50b3\u8f38\u74f6\u9838\u3002 \u25cf\u7b97\u529b\u8207\u7db2\u8def\u6df1\u5ea6\u878d\u5408\uff1a\u7db2\u8def\u4e0d\u518d\u53ea\u662f\u8cc7\u6599\u7684\u50b3\u8f38\u901a\u9053\uff0c\u800c\u662f\u8207\u904b\u7b97\u7dca\u5bc6\u7d50\u5408\uff0c\u5be6\u73fe\u7db2\u8def\u611f\u77e5\u904b\u7b97\u3001\u7db2\u8def\u878d\u5408\u904b\u7b97\uff0c\u751a\u81f3\u904b\u7b97\u91cd\u5851\u7db2\u8def\u3002\u4f8b\u5982\uff0c\u5728\u8d85\u7bc0\u9ede\u5167\u90e8\uff0c\u900f\u904e\u5f15\u5165\u7bc0\u9ede\u5167\u4ea4\u63db\u6676\u7247\uff0c\u589e\u5f37\u5361\u9593P2P\u983b\u5bec\uff0c\u6709\u6548\u63d0\u5347\u7bc0\u9ede\u5167\u7db2\u8def\u50b3\u8f38\u6548\u7387\u3002 \u25cf\u7d71\u4e00\u8cc7\u6e90\u7ba1\u7406\u8207\u8abf\u5ea6\uff1a\u5be6\u73fe\u904b\u7b97\u3001\u5132\u5b58\u8207\u7db2\u8def\u8cc7\u6e90\u7684\u7d71\u4e00\u7d0d\u7ba1\u8207\u878d\u5408\u8def\u7531\u8abf\u5ea6\uff0c\u63d0\u5347\u8cc7\u6e90\u5229\u7528\u7387\u8207\u7ba1\u7406\u6548\u7387\u3002\u76ee\u524d\u696d\u754c\u5178\u578b\u7684\u8d85\u7bc0\u9ede\u65b9\u6848\u5305\u62ec\uff1a(\u4e00)\u82f1\u5049\u9054DGX SuperPOD(\u4ee5NVL72\u70ba\u4f8b)\u82f1\u5049\u9054\u4f5c\u70baAI\u52a0\u901f\u9818\u57df\u7684\u9818\u5c0e\u8005\uff0c\u5176DGX SuperPOD\u7cfb\u5217\u662f\u696d\u754c\u5ee3\u6cdb\u63a1\u7528\u7684AI\u8d85\u7d1a\u8a08\u7b97\u5e73\u53f0\u3002\u5176\u4e2d\uff0cGB200 NVL72 SuperNode\u662f\u5176\u6700\u65b0\u7684\u4ee3\u8868\u6027\u7522\u54c1\u4e4b\u4e00\u3002 GB200 NVL72 SuperNode\u5c0736\u500bGrace CPU\u548c72\u500bBlackwell GPU\u6574\u5408\u5230\u4e00\u500b\u6db2\u51b7\u6a5f\u6ac3\u4e2d\u3002\u5b83\u63a1\u7528\u300cGPU-GPU NVLink ScaleUp+Node-Node RDMAScaleOut\u300d\u7684\u4e92\u806f\u65b9\u5f0f\u3002 \u25cf Compute Tray:\u6574\u500b\u7cfb\u7d71\u5305\u62ec18\u500bCompute Tray,\u6bcf\u500bCompute Tray\u5305\u542b2\u500bGB200\u8d85\u7d1a\u6676\u7247\uff0c\u6bcf\u500bGB200\u8d85\u7d1a\u6676\u7247\u53c8\u5305\u542b2\u500bBlackwell B200 GPU\u548c1\u9846Grace CPU,\u6574\u500b\u6a5f\u6ac3\u517172\u500bB200 GPU\u548c36\u500bGrace CPU\u3002\u900f\u904eNVLink\u8207NVLink-C2C\u6280\u8853\uff0c\u5be6\u73feGPU\u4e4b\u9593\u4ee5\u53caGPU\u8207CPU\u4e4b\u9593\u7684\u9ad8\u901f\u8a18\u61b6\u9ad4\u5171\u4eab\u8207\u8cc7\u6599\u50b3\u8f38\u3002\u55ae\u4e00Compute Tray\u63d0\u4f9b7.2TB\/s(\u55ae\u541128.8Tb\/s)\u983b\u5bec\uff0cNVL72\u6574\u6a5f\u6ac3\u7684Compute Tray\u63d0\u4f9b129.6TB\/S\u7684NVLink\u983b\u5bec\u3002 \u25cf Switch Tray:\u5171\u5305\u542b9\u500bSwitch Tray,\u6bcf\u500bSwitch Tray\u5167\u5efa2\u9846NVSwitch\u6676\u7247\uff0c\u6574\u500b\u6a5f\u6ac3\u63d0\u4f9b18\u500bNVLink Switch\u6676\u7247\u3002\u6574\u6a5f\u6ac3\u5f8c\u90e8\u900f\u904e\u7dda\u7e9c\u5c07ComputeTray\u548cSwitch Tray\u9032\u884c\u4e92\u806f\u3002\u55ae\u4e00Switch Tray\u63d0\u4f9b14.4TB\/s(\u55ae\u541157.6Tb\/s)\u983b\u5bec\uff0cNVL72\u6574\u6a5f\u6ac3\u7684Switch Tray\u63d0\u4f9b129.6TB\/s\u7684NVLink\u983b\u5bec\u3002\u9019\u6a23\u8d85\u7bc0\u9ede\u6574\u6a5f\u6ac3Compute Tray\u7684GPU\u548cSwitch Tray\u7684\u4ea4\u63db\u6676\u7247\u4e4b\u9593\u5c31\u53ef\u4ee5\u5168\u9023\u63a5\u3002 \u25cf Scale Up:NVL72\u5167\u90e8\u63a1\u7528NVLink5\u548cNVSwitch\u5efa\u69cbScale Up\u7db2\u7d61\uff0c\u63d0\u4f9b\u6975\u9ad8\u7684\u983b\u5bec(\u6bcf\u500bCompute Tray\u900f\u904eNVLink\/NVSwitch\u5177\u67097.2TB\/s\u7684ScaleUp\u9023\u63a5\u983b\u5bec)\u548c\u8d85\u4f4e\u6642\u5ef6(\u9285\u96fb\u7e9c\u9023\u63a5\u7bc0\u7701\u4e86\u5149\u6a21\u7d44\u5f15\u5165\u7684\u5ef6\u9072)\u3002\u6240\u6709GPU\u53ef\u4ee5\u5b58\u53d6\u6574\u500b\u8d85\u7bc0\u9ede\u5176\u4ed6GPU\u7684HBM\u8a18\u61b6\u9ad4\u548cGrace CPU\u7684DDR\u5167\u5b58\uff0c\u5be6\u73fe\u7d71\u4e00\u8a18\u61b6\u9ad4\u7a7a\u9593\u3002 \u25cf Scale Out:\u900f\u904eCX8800Gbps RNIC\u63a5\u5165InfiniBand RDMA Scale Out\u7db2\u7d61\uff0c\u5be6\u73fe\u591a\u500bNVL72 SuperNode\u7d44\u6210\u66f4\u5927\u898f\u6a21\u7684SuperPOD(\u4f8b\u59828\u500bDGX GB200NVL72\u7d44\u6210\u4e00\u500b\u5305\u542b576\u584aB200 […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-14032","post","type-post","status-publish","format-standard","hentry","category-killtest"],"_links":{"self":[{"href":"https:\/\/www.killtest.com.tw\/index.php?rest_route=\/wp\/v2\/posts\/14032","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.killtest.com.tw\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.killtest.com.tw\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.killtest.com.tw\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.killtest.com.tw\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=14032"}],"version-history":[{"count":1,"href":"https:\/\/www.killtest.com.tw\/index.php?rest_route=\/wp\/v2\/posts\/14032\/revisions"}],"predecessor-version":[{"id":14034,"href":"https:\/\/www.killtest.com.tw\/index.php?rest_route=\/wp\/v2\/posts\/14032\/revisions\/14034"}],"wp:attachment":[{"href":"https:\/\/www.killtest.com.tw\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=14032"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.killtest.com.tw\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=14032"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.killtest.com.tw\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=14032"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}