Aries: A DNN Inference Scheduling Framework for Multi-core Accelerators
Yunyi Xiang, Zheng Yu Wu, Haidong Yao, Xiankui Xiong, Fan Yang · 2024
To effectively deploy the scaling-up Deep Neural Networks (DNN), the architecture of deep learning accelerators has evolved to multi-core architecture. Deploying these models to multi-core neural processor units (NPU) requires intricate processes such as segmentation, mapping, scheduling, and compiling instructions. Optimizing the entire deployment process represents a sophisticated challenge for the large scheduling search space. To address this challenge, we propose Aries, a DNN scheduling framework for multi-core accelerators. By adopting tensor parallelism and a genetic scheduling algorithm based on an accurate behavior-level architecture model, we have significantly accelerated the process of exploring the scheduling space while ensuring optimization efficiency.