fpgaConvNet: Mapping Convolutional Neural Networks to FPGAs


Christos Bouganis, Imperial College London, UK  -- 14-10-2016


In recent years, Convolutional Neural Networks (ConvNets) have managed to become the state-of-the-art in several AI tasks. In this context, FPGAs can provide a potential platform that can map such networks achieving high performance at low power consumption. The talk will focus on our recent tool, fpgaConvNet, that automates the design space exploration process and maps efficiently the given ConvNet to an FPGA device. Our framework yields designs that improve the performance density and the performance efficiency by up to 6× and 4.49× respectively over existing highly-optimised FPGA, DSP and embedded GPU work, providing practitioners an automated tool-flow for mapping ConvNets to FPGAs.


Christos-Savvas Bouganis is a Senior Lecturer with the Electrical and Electronic Engineering Department, Imperial College London, London, U.K. He has published over 60 research papers in peer-referred journals and international conferences, and he has contributed three book chapters on digital system design. His current research interests include the theory and practice of reconfigurable computing and design automation, mainly targeting digital signal processing algorithms.




CE Tweets