Have GPUs made FPGAs redundant in the field of video processing?

Since the 1990s FPGAs have been popular for accelerating video processing applications. This paper presents GPUs (Graphics Processing Units) as another viable solution, comparing their throughput with that of FPGAs. Previous work on using FPGAs and GPUs for video processing is analysed, grounds for comparison formulated and then exemplified through case studies of primary colour correction and 2D convolution. GPUs are seen to be advantageous in applications utilising their optimised instruction set and which have a low number of memory accesses. We found that for 2D convolution, the throughput of the GPUs exceeds that of FPGAs if the mask dimension is 2 x 2. For mask sizes greater than 4 X 4 and for primary colour correction FPGAs have higher throughput. The GPU implementation of primary colour correction is shown to be capable of a throughput of 63 MP/s (million pixels per second) enough for high definition video at 30 f/s (frames per second). For 2D convolution GPUs achieved the target throughput rate of 8MP/s (512 x 512 frames at 30f/s) up to size 7 x 7. The Spartan 3 FPGA was capable of over 60 MP/s throughput and the Virtex 11 Pro over 110MP/s at this size of convolution.