F3D: Accelerating 3D Convolutional Neural Networks in Frequency Space Using ReRAM
Hosted in Virtual Platform
Description3D frequency-domain convolutional neural networks (CNNs) are gaining popularity in video analysis for their superior capability of replacing convolutions with simpler element-wise multiplications. Conventional dedicated accelerators employ memory hierarchy organization for high throughput but at the expensive costs of significant amount of data movements and energy consumptions. This paper presents F3D, a processing-in-memory frequency-domain accelerator using resistive random-access memory (ReRAM). We alleviate the overheads of redundant data movements in ReRAM-based computing by the data reuse from multiple adjacent frames and the inherent symmetry of inputs in frequency space. Evaluation results demonstrate that F3D outperforms state-of-the-art accelerators with lower energy consumption.