TianU: A Multimodal Accelerator with Fine-Grained Processing Unit Scheduling for Efficient Inference

Efficient execution of multimodal tasks on edge platforms remains challenging due to the heterogeneity of data types, model structures, and computational demands. This paper proposes a multimodal accelerator architecture featuring a fine-grained PU scheduling strategy. The proposed method dynami-cally partitions a shared 14×14 PU array at the layer level across different tasks, enabling concurrent processing while maintaining high hardware utilization. A reconfigurable adder tree ensures flexible accumulation under varying task patterns, while a configurable tiny value skipping mechanism further enhances energy efficiency. The system supports both high-performance and low-power modes to adapt to diverse application scenarios. Experimental results show that the proposed scheduling strategy achieves an average PU utilization of over 94% under both modes, significantly outperforming static scheduling baselines. The hardware overhead of the scheduler is minimal, making it well-suited for deployment in intelligent wearables, health monitoring devices, and other resource-constrained multimodal edge systems.