Each individual Atlas-950 SuperPoD has 8192 NPUs & far more integrated than DeepSeek linking together 2048 H800. Get a few Atlas-950, each can train an expert at a time. 160 cabinets - 128 compute & 32 network 16384 800GE optical connection, 4 bw each compute & network rack

8192 / 128 -> each compute rack is 8 sub-racks of 8 NPUs connected by 72 lane 112G thru Electrical backplane (copper) + connected to other compute racks via 2-hops thru network cabinet via 800GE optical lines Sharing single unified address space s/w sees 1 unit, no network pkt

Scale of Atlas vs NVDA - system efficiency: Key is compute efficiency. HW claims losing just 5% performance to communication overhead across 8192 chip SuperPoD. GB200 NVL72 achieves 90% efficiency @ 72 GPUs. Drop-off is larger once you get to other racks due to networking

Atlas-950 network specs & compared to Atlas-960 interconnect in 960 is so much better. Not more heavy Serdes @ transceiver level Hi-ONE directly stacked w/ Ascend-960 die Uses latest OXC (requiring no Serdes) + More NPUs/Cabinet & more cabinets overall.

Atlas-950 is a significant upgrade for DeepSeek, but Atlas-960 is just so much better. Ppl see the per chip figure on Ascend-960 vs 950DT, but the improvement on networking level is very significant. either way, a couple of these should really improve DS training.