Huawei Unveils Kirin 9050 with New DaVinci NPU for On-Device 30B MoE Models
Huawei officially announced the Kirin 9050 series chip featuring a new generation DaVinci architecture NPU, enabling the industry's first execution of a 30B MoE large language model directly on-device. The new NPU delivers a 140% performance increase and a 66% power reduction compared to previous iterations. This breakthrough allows complex AI models to run locally without internet connectivity or cloud dependency, significantly enhancing edge AI responsiveness and user data privacy. It marks a major milestone in mobile hardware capabilities, bridging the gap between heavy cloud-based computation and everyday edge devices. The Kirin 9050 chip incorporates 'logical folding' technology and supports Transformer engines, ultra-low power perception engines, and the CANN heterogeneous computing architecture. These capabilities enable advanced on-device features such as AI screen reading, Xmage Smart Photography 2.0, and offline map navigation.
## BACKGROUND
The DaVinci architecture is Huawei's proprietary AI computing architecture featuring high performance and flexible scalability, designed around 3D Cube matrix multiplication units. Mixture-of-Experts (MoE) is a machine learning technique where only a subset of network parameters or 'experts' are activated for a given input, reducing computational costs while maintaining model capacity.