As the compute and memory demands of AI models keep growing, research into AI systems capable of efficient training and inference matters more and more. This laboratory designs high-performance AI systems that handle AI models efficiently on heterogeneous systems mixing CPUs, GPUs and NPUs. On-device LLM inference, neural architecture search, AI accelerator design and multi-dimensional deep learning parallelism for LLMs are its core subjects, and designing algorithm and hardware together is its method.