Huawei Unveiled new Atlas 960E SuperPoDs chip (its answer to Nvidia’s latest Rubin systems; Has similar computing power but nine times more memory, and they use optical links instead of copper - HUAWEI CONNECT 2026
- According to Huawei, a single Atlas 960E can scale to 4,096 NPUs and provide 8 EFLOPS of FP8 compute performance with up to 1 petabyte of high-bandwidth memory. Hi-ONE uses NPO, which places optical components closer to the computing or switching silicon, shortening the high-speed electrical path between them. Huawei lists transmission capacity of 7.2 Tbit/s for each Hi-ONE engine. The IEEE Solid-State Circuits Society identifies near-packaged and co-packaged optics as approaches being developed to increase bandwidth density and energy efficiency as AI systems place greater demands on data movement.
- Huawei describes Hi-ONE as the industry’s first NPO product ready for mass production and the first with an integrated light source. Huawei estimates that an Atlas 960E uses about 5,500 Hi-ONE units instead of the roughly 48,000 800G optical modules that would otherwise be required to connect its NPUs. The company estimates that this design reduces power consumption by more than 550 kilowatts and provides 99.8% system availability.
- Huawei also introduced the OceanStor M900, a storage cluster intended to provide multi-tier key-value caching for AI inference workloads. The system uses UnifiedBus to provide direct access to a petabyte-scale KV cache and combines different storage media with a retention algorithm that Huawei says extends SSD read-and-write lifespan 16-fold.
- The TaiShan, Ascend, and storage systems form part of what Huawei calls its agentic SuperCluster architecture. UnifiedBus is used to connect Ascend SuperPoDs, Kunpeng-based SuperPoDs, and KV cache storage while reducing the need for protocol conversion between systems.
- Also plans to release the Ascend 970 in 2028 and the Ascend 980 in 2029. Huawei has not identified all 11 chips individually in the announcement, but described Ascend as the main computing component of the portfolio. UnifiedBus is used to connect the different parts of the system through a common interconnect architecture.
- Provided updates on its software ecosystem: According to Huawei, external developers now account for 61% of all CANN developers, while the community has more than 5,200 monthly active developers. Ascend currently supports more than 90 third-party open-source projects, including PyTorch, Triton, vLLM, and veRL. Jim Zemlin, CEO of the Linux Foundation, also addressed the role of open-source software in AI infrastructure during a keynote at HUAWEI CONNECT 2026. “At every single layer of the agentic AI stack, there is an open-source project powering that technology,” Zemlin said. He cited Linux and Kubernetes at the infrastructure layer, PyTorch for model training, and vLLM and Ray for inference. Zemlin also pointed to emerging open-source protocols such as MCP and A2A at the agent layer.
** Context: Huawei’s Atlas 960E SuperPoD pairs up to 4,096 Ascend 960 NPUs with near-packaged optical interconnects and claims up to 1 petabyte of HBM memory, making memory capacity and system-scale connectivity its principal competitive pitch against Nvidia. The announcement matters because China’s restricted access to leading U.S. AI hardware has elevated domestic compute and memory supply chains; CXMT’s reported HBM3E trial production could help reduce a key bottleneck. The competitive caveat is that CXMT remains behind established HBM leaders, which are already producing HBM4-class products, while Nvidia’s latest outlook assumes no China data-center-compute revenue.
Recent: Huawei moved the Ascend 960DT training-chip launch forward to the first quarter of 2027, while CXMT reportedly began HBM3E trial production earlier this month.
Upcoming: Huawei targets Q1 2027 for Ascend 960DT availability and Q3 2027 for the inference-focused Ascend 960PR.