FFT on a Heterogeneous System with a General-Purpose Map-Scan Accelerator

Abstract

This paper presents a use-case and evaluation for a custom general-purpose Map-Scan Accelerator using multiple Fast Fourier Transform (FFT) implementations. FFT computing is part of the "Spectral Methods" class of problems and qualifies to be accelerated because it is computationally intensive, of average complexity and widely used as one of the backbone algorithms in modern computing. In order to achieve the desired throughput or latency, different parameters are used for testing, regarding both the accelerator hardware and the FFT. Evaluation is done in simulation, targeting FPGA implementation. FFT is shown to work efficiently on our architecture, either latency or throughput (and in some circumstances, both by a smaller amount) can be improved linearly or even supralinearly compared to a single-core solution. Accelerations obtained vary ∼ 0.5 × p and ∼ 1.5 × p where p the number of accelerator cores.