Huawei Semiconductor Chief Liao Heng Shares AI Chip Insights and '18-Layer Pagoda' Theory
Huawei's Chief Semiconductor Scientist Liao Heng gave a rare interview, introducing his '18-layer pagoda' theory of the semiconductor stack and expressing concern over a critical shortage of hardware talent. He also praised DeepSeek founder Liang Wenfeng for adopting a sparse activation architecture rather than blindly chasing the Scaling Law. As geopolitical restrictions limit access to advanced manufacturing nodes, Huawei's strategy relies on cross-layer synergy and system-level clusters to compensate for single-chip hardware constraints. Liao's insights highlight the shifting focus of the AI chip industry from raw hardware scaling to full-stack co-design and architectural innovation. Liao outlined that the semiconductor stack spans 18 levels from raw materials to applications, where the shortest board determines the overall system limit. He also noted that Moore's Law has economically stagnated at the 16nm and 7nm nodes, making cross-disciplinary talent who understand both hardware and software constraints highly critical.
## BACKGROUND
Huawei's Ascend series represents its proprietary line of AI processors designed for deep learning, supported by CANN (Compute Architecture for Neural Networks), a heterogeneous computing architecture. As physical scaling limits and geopolitical restrictions impact chip manufacturing, hardware developers are increasingly focusing on software-hardware co-design to optimize AI workloads.