Huawei: AI system performance is no longer just about the chip

At HUAWEI CONNECT 2026, Huawei argued that AI cluster performance depends on memory, interconnects, storage, software and power, not just NPU counts and FLOPS. In traditional 100,000-NPU clusters, intra-cluster communications account for more than 40% of total training time.

• Intra-cluster communications take over 40% of training time in traditional 100,000-NPU clusters
• Ascend 950 offers 2 TB/s interconnect bandwidth, about 2.5x that of Ascend 910C
• Atlas 950 supports up to 8,192 Ascend 950 processors at 8 EFLOPS FP8 and 16 EFLOPS FP4
• Atlas 960E reaches the same 8 EFLOPS FP8 with 4,096 Ascend 960 chips, half the NPUs