Huawei's LogicFolding: 3D Stacking Rewrites AI Chip Rules
Summary
Key Takeaways
Huawei's rotating chairman Xu Zhijun publicly thanked US sanctions for forcing China's semiconductor acceleration. At IEEE ISCAS, chip division head He Tingbo introduced Tau Scaling Law, shifting focus from transistor shrinkage to signal propagation time reduction. The LogicFolding architecture stacks logic cells vertically, claiming 55% density improvement and 41% power efficiency gain, targeting 1.4nm class by 2031—an effective density via packaging, not lithography. Huawei also launched Ascend 920 (6nm, >900 TFLOPS, 4 TB/s HBM3) aimed at the restricted Nvidia H20. Jensen Huang confirmed Nvidia's China AI accelerator share dropped to zero. Huawei's HarmonyOS now runs on over 1 billion devices, forming a full-stack ecosystem. A Huawei-linked research group used >1,000 Ascend 910C for post-training DeepSeek's 1.6 trillion-parameter V4-Pro model, proving real-world AI workload capability, though full pretraining remains unproven.
Why It Matters
Huawei's move is a strategic ecosystem lock-in disguised as a breakthrough. LogicFolding relies on advanced packaging (HBM3) and 3D stacking, with unaddressed tail latency and thermal issues. The 1.4nm class claim obscures the true lithography gap with TSMC. Through CANN and MindSpore, Huawei traps customers into its HarmonyOS stack, killing cross-platform portability. For AI training, Huawei lacks an InfiniBand-class network; its RoCEv2 suffers PFC/ECN bottlenecks and higher tail latency at scale. Software maturity (CUDA replacement) is years behind, and supply chain reliance on non-Chinese HBM/equipment persists.
PRO Decision
【Vendors】Nvidia, AMD, Intel should exploit Huawei's weaknesses: highlight interconnect bottlenecks and immature software for large-scale pretraining, promote open alternatives (ROCm, oneAPI), and publish independent benchmarks comparing Nvidia InfiniBand vs. Huawei RoCEv2 on tail latency and cluster efficiency. 【Enterprises】CIOs must demand third-party tests for Ascend 920/910C tail latency, power, and training throughput at scale. Avoid toolchain lock-in via CANN/MindSpore; insist on cross-cloud portability with PyTorch/TensorFlow. Diversify suppliers to mitigate concentration risk. 【Investors】See through the hype: Huawei's architecture is real but still lags in lithography and depends on non-Chinese HBM supply chain. Its closed ecosystem limits global reach. Monitor Nvidia/AMD 3D stacking roadmaps and ASML High-NA EUV progress as counterweights.
Get 3-5 key AI infrastructure signals weekly →
💬 Comments (0)