Huawei Launches AI Clusters to Challenge Nvidia

🔑 Enhanced Key Takeaways

• •Atlas 950 SuperPoD delivers 8 EFLOPS in FP8 and 16 EFLOPS in FP4, with 16 PB/s interconnect bandwidth and over 1 PB of memory.[1][2][3]
• •Atlas 960 SuperPoD doubles the performance with 30 EFLOPS FP8, 60 EFLOPS FP4, 4,460 TB memory, and 34 PB/s bandwidth.[1][2]
• •Atlas 950 SuperCluster scales to 64 supernodes with over 520,000 Ascend 950DT chips, providing 524 EFLOPS FP8, launching Q4 2026.[2][5]
• •Products announced earlier at Huawei Connect 2025, featuring innovations like all-optical interconnect and cableless electrical interconnection.[1][4]

🛠️ Technical Deep Dive

• •Ascend 950DT chip: 1 PFLOPS FP8, 2 PFLOPS FP4, 2 TB/s interconnect bandwidth (2.5x higher than Ascend 910C).[1]
• •Atlas 950 SuperPoD: 160 cabinets (128 compute, 32 communications) in 1,000 m², all-optical interconnect, 16 PB/s bandwidth.[1]
• •UnifiedBus interconnect enables ultra-high bandwidth, ultra-low latency, unified memory addressing as single logical machine.[4][6]
• •Orthogonal architecture with floating blind-mate liquid-cooling connector for zero leaks and double reliability.[4]
• •Compared to prior Atlas 900, training performance improved 17x to 4.91M TPS, inference up to 26.5x to 19.6M TPS with FP4.[2]

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline