Huawei advances Ascend 960 and million-NPU cluster roadmap
Huawei announced the Peerium and UnifiedBus architecture, an Atlas 960E SuperPoD planned to connect up to 4,096 Ascend 960 NPUs, and a SuperCluster design intended to scale to one million NPUs. Ascend 960DT is now targeted for Q1 2027 and 960PR for Q3; OceanStor M900 context-memory storage and an upgraded TaiShan 950 were also unveiled.
Why it matters: The roadmap is a systems-level attempt to offset individual-chip limits with optical interconnect, shared memory and KV-cache infrastructure. If delivered, it would expand the non-NVIDIA path for very large training and agentic inference deployments, but it…