Npu Cann
The first Guangdong Provincial Artificial Intelligence Application Matchmaking Conference was held on April 27, 2026, in Shenzhen. Huawei Technologies Co., Ltd., introduced Huawei’s new generation supernode product, Atlas 950 SuperPoD, intelligent computing supernode and the TaiShan 950 SuperPoD universal computing supernode. It can support high-speed interconnection of up to 8,192 Ascend cards and can run stably for 30 hours without interruption.
The Atlas 950 SuperPoD supernode can uses 64 cards per cabinet, and support high-speed interconnection of up to 8,192 Ascend NPU cards to form a large-scale cluster. Its hardware architecture is mainly divided into three core modules: computing module, interconnect module, and storage module:
computing module: Ascend 950PR NPU chip, it is the first to support FP4/FP8, with an FP4 computing power of 1.56 PFLOPS. The self-developed memory bandwidth reaches 4TB/s.
Interconnection module: used Unified Bus or Lingqu Interconnection Protocol 2.0, bandwidth reaches 16.3PB/s; reduces RTT (round-trip time) communication latency from 7 microseconds to 3 microseconds. Atlas 950 SuperPoD implements an 1152TB shared memory pool.
Storage module: a hybrid architecture of “distributed storage + local caching,”: supports PB-level capacity expansion and storage bandwidth of over 10TB/s, and can meet the storage and read/write needs of 10 trillion data points. On the other hand, each computing node is equipped with a local high-speed cache, using NVMe SSD storage media, with read/write speeds of over 3000MB/s. Meanwhile, the storage module supports data redundancy backup and automatic fault switching, with data reliability reaching 99.999%.
The software layer is Huawei Compute Architecture for Neural Networks (CANN, it is a open source project, famouse for training DeepSeek-V4. And it supports PyTorch, vLLM, SGLang, xLLM, VeRL, Triton, and TileLang.
The NVDA stock dropped 8% since April 27.