NeuroCGRA: A CGRA with support for neural networks

Abstract

<div> Coarse Grained Reconfigurable Architectures</div> <div> (CGRAs) are emerging as enabling platforms to meet the high</div> <div> performance demanded by modern embedded applications. In</div> <div> many application domains (e.g. robotics and cognitive embedded</div> <div> systems), the CGRAs are required to simultaneously host</div> <div> processing (e.g. Audio/video acquisition) and estimation (e.g.</div> <div> audio/video/image recognition) tasks. Recent works have revealed</div> <div> that the efficiency and scalability of the estimation algorithms</div> <div> can be significantly improved by using neural networks.</div> <div> However, existing CGRAs commonly employ homogeneous</div> <div> processing resources for both the tasks. To realize the best of</div> <div> both the worlds (conventional processing and neural networks),</div> <div> we present NeuroCGRA. NeuroCGRA allows the processing</div> <div> elements and the network to dynamically morph into either</div> <div> conventional CGRA or a neural network, depending on the</div> <div> hosted application. We have chosen the DRRA as a vehicle to</div> <div> study the feasibility and overheads of our approach. Simulation</div> <div> using edge detection reveal that the neural networks can</div> <div> successfully process real-time video for up to 1M pixels.</div> <div> Synthesis results reveal that the proposed enhancements incur</div> <div> negligible overheads (4.4% area and 9.1% power) compared to</div> <div> the original DRRA cell.</div>