GAS: A Flexible CNN Accelerator Architecture for Heterogeneous Dataflow Implementation

Convolutional neural networks (CNNs) play an important role in various fields. However, the efficient processing of CNNs presents a significant challenge due to substantial variations within models and even between individual layers. To address these issues, we propose a heterogeneous dataflow accelerator group architecture named GAS that fully exploits the flexibility of dataflow while demonstrating high extensibility. For the architecture, we propose a heterogeneous hierarchical scheduling strategy called H3S. The strategy utilizes a task grouping algorithm to automatically partition and evaluate the matching of tasks and dataflows, considering data dependencies and global load balancing. This approach enhances scheduling efficiency while achieving optimal task performance. Evaluation results demonstrate that, compared to the most advanced scheduling strategy and homogeneous accelerators, the delay-energy product (EDP) is reduced by a minimum 12%, scheduling time is reduced by 50% and latency is reduced by 10%.