ARMAN: A Reconfigurable Monolithic 3D Accelerator Architecture for Convolutional Neural Networks (2402.04431v1)

Abstract: The Convolutional Neural Network (CNN) has emerged as a powerful and versatile tool for AI applications. Conventional computing architectures face challenges in meeting the demanding processing requirements of compute-intensive CNN applications, as they suffer from limited throughput and low utilization. To this end, specialized accelerators have been developed to speed up CNN computations. However, as we demonstrate in this paper via extensive design space exploration, different neural network models have different characteristics, which calls for different accelerator architectures and configurations to match their computing demand. We show that a one-size-fits-all fixed architecture does not guarantee optimal power/energy/performance trade-off. To overcome this challenge, this paper proposes ARMAN, a novel reconfigurable systolic-array-based accelerator architecture based on Monolithic 3D (M3D) technology for CNN inference. The proposed accelerator offers the flexibility to reconfigure among different scale-up or scale-out arrangements depending on the neural network structure, providing the optimal trade-off across power, energy, and performance for various neural network models. We demonstrate the effectiveness of our approach through evaluations of multiple benchmarks. The results demonstrate that the proposed accelerator exhibits up to 2x, 2.24x, 1.48x, and 2x improvements in terms of execution cycles, power, energy, and EDP respectively, over the non-configurable architecture.