We present the design and implementation of a custom GPU-based compute cluster that provides the correlation X-engine of the CHIME Pathfinder radio telescope. It is among the largest such systems in operation, correlating 32,896 baselines (256 inputs) over 400MHz of radio bandwidth. Making heavy use of consumer-grade parts and a custom software stack, the system was developed at a small fraction of the cost of comparable installations. Unlike existing GPU backends, this system is built around OpenCL kernels running on consumer-level AMD GPUs, taking advantage of low-cost hardware and leveraging packed integer operations to double algorithmic efficiency. The system achieves the required 105TOPS in a 10kW power envelope, making it among the most power-efficient X-engines in use today.
@article{arxiv.1503.06202,
title = {A GPU-based Correlator X-engine Implemented on the CHIME Pathfinder},
author = {Nolan Denman and Mandana Amiri and Kevin Bandura and Jean-François Cliche and Liam Connor and Matt Dobbs and Mateus Fandino and Mark Halpern and Adam Hincks and Gary Hinshaw and Carolin Höfer and Peter Klages and Kiyoshi Masui and Juan Mena Parra and Laura Newburgh and Andre Recnik and J. Richard Shaw and Kris Sigurdson and Kendrick Smith and Keith Vanderlinde},
journal= {arXiv preprint arXiv:1503.06202},
year = {2016}
}