Collaborative Acceleration for FFT on Commercial Processing-In-Memory Architectures

TL;DR

This paper investigates the potential of commercial processing-in-memory (PIM) architectures to accelerate FFT computations, finding that collaborative use with GPUs and optimized mappings can significantly improve performance and data movement efficiency.

Contribution

It introduces Pimacolaba, an optimized PIM FFT mapping that, combined with GPU collaboration, enhances FFT acceleration and reduces data movement.

Findings

PIM alone is not effective for FFT acceleration.

Collaborative PIM-GPU approach improves performance.

Pimacolaba reduces data movement by up to 2.76×.

Abstract

This paper evaluates the efficacy of recent commercial processing-in-memory (PIM) solutions to accelerate fast Fourier transform (FFT), an important primitive across several domains. Specifically, we observe that efficient implementations of FFT on modern GPUs are memory bandwidth bound. As such, the memory bandwidth boost availed by commercial PIM solutions makes a case for PIM to accelerate FFT. To this end, we first deduce a mapping of FFT computation to a strawman PIM architecture representative of recent commercial designs. We observe that even with careful data mapping, PIM is not effective in accelerating FFT. To address this, we make a case for collaborative acceleration of FFT with PIM and GPU. Further, we propose software and hardware innovations which lower PIM operations necessary for a given FFT. Overall, our optimized PIM FFT mapping, termed Pimacolaba, delivers performance and data movement savings of up to 1.38$\times$ and 2.76$\times$, respectively, over a range of FFT sizes.