Huawei
2026-08-13
Huawei Ascend 910B and 8192-Chip Supernode: In-Memory Computing Breakthrough
Summary
Key Takeaways
Huawei is ramping up testing of its latest AI chip, the Ascend 910B, to challenge NVIDIA's dominance amid U.S. sanctions. It showcased the Atlas 950 SuperPoD, an 8192-chip supernode delivering 524 EFLOPS FP8, planned for Q4 2026. Additionally, Huawei outlined the Atlas 950 SuperCluster with over 520,000 chips providing 1 ZFLOPS FP4. At ISSCC 2026, Huawei unveiled a 55nm in-memory computing AI chip based on the Tao Law architecture, achieving 3.2x energy efficiency over 7nm von Neumann, earning dual recognition. These developments demonstrate Huawei's innovation from chip to system level. The Ascend 910B is expected to rival NVIDIA H100. The 8192-chip supernode highlights Huawei's system integration capability for large-scale AI training. The in-memory computing chip targets inference, reducing power by minimizing data movement. This launch is strategic for China's AI infrastructure amid export restrictions.
Why It Matters
Huawei's announcement is a defensive move to secure its share in China's AI chip market against NVIDIA and encircle domestic startups. The Atlas 950 SuperPoD and CANN software stack create a closed ecosystem, locking users into proprietary interconnects and tools. Huawei downplays physical constraints: the 8192-chip supernode likely requires extensive liquid cooling, increasing TCO. The FP4 precision poses accuracy risks, requiring additional calibration. The 55nm in-memory chip, while energy-efficient, has low compute density and limited application scope. By shifting the control plane from standard Ethernet to proprietary chip interconnects, Huawei reduces network flexibility, similar to NVIDIA's NVLink but more closed, leading to long-term asset lock-in.
PRO Decision
[Vendors] NVIDIA and AMD should leverage their open ecosystems. NVIDIA should promote NVLink open standard and CUDA maturity, while AMD highlights ROCm and Ethernet flexibility. Both should attack Huawei's proprietary interconnects and CANN lock-in, questioning FP4 precision in training and 55nm in-memory chip generality.
[Enterprises] CIOs must audit Atlas 950 SuperPoD power, cooling, and interconnect specs. Test FP4 accuracy impact and CANN compatibility. Demand standard PCIe/Ethernet options to ensure portability. Include liquid cooling costs in TCO.
[Investors] Look beyond the hype: Ascend 910B lags due to process constraints; in-memory chip has limited market; SuperPoD faces competition from NVIDIA's next-gen. Focus on ecosystem adoption rates. Huawei has policy support in China but limited global reach.
Get 3-5 key AI infrastructure signals weekly →
💬 Comments (0)