Highly Scalable FFT on GPUs

Introduction

Extended Tarang, a parallel computational fluid dynamics simulator, to enable GPU-based FFTs.

Duration: February 2021 to August 2021

Mentor: Prof. Mahendra Verma, Physics Dept. IIT Kanpur

Details

  • Extended Tarang, a parallel computational fluid dynamics simulator, to enable GPU-based FFTs.
  • Developed MPI and CuFFT-XT-based variable precision 1D and 3D scalable forward and inverse FFTs.
  • Achieved upto 1650× single-node speed-up over FFTW on DGX-A100 and linear strong and weak scaling.
  • Compared DGX-A100’s performance in PARAM Siddhi AI with EPCC Cirrus cluster with 6 V100 GPUs per node.
Avatar
Anish Saxena
Computer Science PhD Student

Computer Architecture, Systems, and Security

Talks

Discussed CUDA programmign concepts as part of guest lecture in course on High Performance Computing