Performance of Remote FPGA-Based Coprocessors for Image-Processing Applications

This paper describes a performance evaluation of ImageProcessing applications on FPGA-based coprocessors that are part of general-purpose computers. Our experiments show that the maximum speed-up depends on the amount of data processed by the coprocessor. Taking images with 256 x 256 pixels, a moderate FPGA capacity of 10E+5 CLBs provides two orders of magnitude of performance improvement over a Pentium III processor for most of our benchmarks. However, memory organization and host bus degrade these results. Those benchmarks that can exhibit high performance improvement would require about 200 memory banks of 256 bytes and a host bandwidth as high as 30 GB/s. Based on our quantitative approach, it can be explained why some currently available FPGA-based coprocessors do not provide the achievable level of performance for some ImageProcessing applications.