TriGen: NPU Architecture for End-to-End Acceleration of Large Language Models based on SW-HW Co-Design

Researchers announced the development of TriGen, a novel NPU architecture designed to accelerate large language models on resource-constrained devices in a paper published on February 26, 2026, addressing the growing challenge of running increasingly large transformer models with limited parameter reuse capabilities. The research comes at a critical time as transformer-based LLMs have become dominant in AI applications, yet their rapidly growing sizes present significant difficulties for deployment on devices with limited computational resources. Unlike conventional CNNs which benefit from high parameter reuse, these modern language models face extreme challenges when executed end-to-end on resource-limited hardware environments. The TriGen architecture represents a breakthrough approach through its innovative software-hardware co-design methodology specifically tailored to address the unique computational patterns of large transformer models. This co-design strategy optimizes both the processing architecture and the software algorithms to maximize efficiency within the constraints of on-device environments, potentially enabling more powerful AI capabilities directly on smartphones, IoT devices, and other edge computing platforms without relying on cloud-based processing.