Atlas 950 SuperPod
A high-performance AI computing cluster platform introduced by Huawei Ascend. It adopts a UB bus interconnect architecture, supports large-scale NPU cluster networking, and is designed for large-model training and inference scenarios.
Timeline (last 90 days)
One PC16 node consists of two 950 CPUs paired with 16 NPUs, with every 8 cards forming one 4-Mesh chassis
Within a 4-Mesh chassis, each NPU is directly connected to the other 7 cards via 7 UB ports with a bidirectional bandwidth of up to 784 GB/s; an additional 8 UB ports are used for cross-chassis communication with a bidirectional bandwidth of 850 GB/s
All Facts (2)
One PC16 node consists of two 950 CPUs paired with 16 NPUs, with every 8 cards forming one 4-Mesh chassis
50%UnverifiedWithin a 4-Mesh chassis, each NPU is directly connected to the other 7 cards via 7 UB ports with a bidirectional bandwidth of up to 784 GB/s; an additional 8 UB ports are used for cross-chassis communication with a bidirectional bandwidth of 850 GB/s
50%