OpenCL Programmed FPGAs Claim a 3X Performance-to-Power Advantage at Microsoft

The Microsoft white paper, “Accelerating Deep Convolutional Neural Networks Using Specialized Hardware” describes an OpenCL programmed implementation of Convolutional Neural Networks (CNNs) that touts a conservative estimate of 3x the performance-to-power advantage over NVIDIA GPUs when running on new FPGA hardware.

 (Image courtesy Microsoft)

Doug Berger posted on the Inside Microsoft Research Blog that Microsoft expects “a throughput increase of 70% with comparable energy used” when using the new Arria 10 FPGAs from Altera. CNN based techniques are widely used to push the boundaries of computer vision processing, such as image recognition, and are used by the Bing search engine.

Doug Berger, Director Client & Cloud Apps