Spectral Method Characterization on FPGA and GPU Accelerators

Abstract

Hybrid core computing, with CPUs augmented with FPGAs and/or GPUs, offers a promising pathway of addressing emerging high-performance computing demands, particularly with respect to performance, power and productivity. This paper compares the sustained performance of a complex, single precision, floating-point, 1D, Fast Fourier Transform (FFT) implementation on state-of-the-art FPGA and GPU accelerators. As results show, FPGA floating-point performance is highly sensitive to the availability of dedicated FPGA resources: DSP48E slices, block RAMs and FPGA I/O banks in particular. Provided results show that for the floating-point FFT benchmark on FPGAs, these resources are the performance limiting factor. For fixed-point FFTs, however, FPGAs exploit a flexible data path width to trade-off circuit cost with speed of computation in applications requiring smaller precision to improve performance, power and device utilization. GPUs cannot fully take advantage of this, having a fixed data-width architecture. Results show a trade-off with respect to performance, memory input/output and available device resources when choosing the right accelerators for a particular application.