3D recursive Gaussian IIR on GPU and FPGAs — A case for accelerating bandwidth-bounded applications

GPU device typically has a higher off-chip bandwidth than FPGA-based systems. Thus typically GPU should perform better for bandwidth-bounded massive parallel applications. In this paper, we present our implementations of a 3D recursive Gaussian IIR on multi-core CPU, many-core GPU and multi-FPGA platforms. Our baseline implementation on the CPU features the smallest arithmetic computation (2 MADDs per dimension). While this application is clearly bandwidth bounded, the difference on the memory subsystems translates to different bandwidth optimization techniques. Our implementations on the GPU and FPGA platforms show 26X and 33X speedup respectively over optimized single-thread code on CPU.