H: Huawei will announce a 'LogicFolding' optimization for the Ascend 910C that reduces memory bandwidth
Huawei will announce a 'LogicFolding' optimization for the Ascend 910C that reduces memory bandwidth requirements by 40% for MoE models, specifically targeting DeepSeek's architecture, within 120 days.
Huawei's relationship graph shows it 'uses' Mixture of Experts (MoE) technology. LogicFolding is a novel architecture for Kirin chips. The Tau Scaling Law suggests a focus on compute efficiency. DeepSeek is a prominent Chinese AI lab with a MoE architecture. Combining these signals, Huawei is likely developing a co-optimized hardware+software solution for the most popular domestic MoE model.
Look for a technical whitepaper, press release, or conference presentation (e.g., at a Chinese AI conference like WAIC or CCF-GAIR) from Huawei detailing a hardware-software co-optimization for MoE models on Ascend, specifically mentioning DeepSeek.