This work was presented in ISSSCC 2025. We applied the binary/ternary quantization for large-language-model to solve memory bottlneck. And, we designed energy-efficient hardware architecture for binary/ternary matrix multiplication.
This work was presented in ISSCC 2024. We combine deep-neural-network and neurmorphic-based neural network to increase energy efficiency of language model processing. Also, we proposed efficient weight compression techniques.
This work was presented in ISSCC 2023. We proposed the complementary neural network which can achieve advantages of both deep-neural-network and neuromorphic-based neural network.