Axera’s proprietary NPU is built on deep co-design of algorithms and hardware, adopting a multithreaded, heterogeneous, multi-core architecture. Through tight integration of memory and processing units, a programmable dataflow microarchitecture, and a comprehensive operator and instruction set, it significantly improves computational efficiency and compute density. The architecture natively supports a wide range of models, including Transformer, CNN, and BEV, and, together with a mature software toolchain, enables cost-effective, efficient, and environmentally sustainable model deployment and operations. This fully empowers the large-scale, high-efficiency adoption of edge and terminal AI.