In today's AI revolution, Neural Networks (NNs) have taken center stage, driving remarkable progress in artificial intelligence over the past decade. However, existing technologies like FPGA and GPU implementations, though widely used for accelerating Convolutional Neural Networks (CNNs) face significant limitations in terms of speed, energy efficiency, and susceptibility to hardware noise. FPGA and GPU implementations struggle with redundant data queries and communication latencies, resulting in under-utilization of computing resources, hindering advancements in energy-efficient computing.
Researchers at George Washington University have developed a novel photonic CNN accelerator, inspired by the efficient Winograd filtering algorithm. The accelerator optimizes hardware resources by minimizing the number of required multiplications for convolution operations, outperforming traditional FPGA and GPU implementations for 3x3 filters in the VGG16 network. Moreover, the photonic accelerator exhibits remarkable energy efficiency, consuming significantly less power and offering up to three orders of magnitude improvement compared to traditional electronic counterparts. By addressing hardware noise sources, the researchers have enhanced the reliability and accuracy of the photonic accelerator, positioning it as a game-changer for diverse applications, from data centers to IoT devices.
High-level architecture of the proposed Photonic Accelerator for Convolutional Neural Networks
Advantages:
Applications: