An implementation of the BFDMT in CUDA

CI: – Keith Bannister

Aim: This project aims to produce an optimised CUDA/GPU implementation of the Baseline Fast Dispersion Measure Transform (BFDMT).

Significance: If the desired performance of this implementation can be achieved, it will form an integral part of a new coherent fast radio burst (FRB) detection system for the Australian Square Kilometre Array Pathfinder (ASKAP). This new system will boost the detection rate of arcsecond-localised FRBs to roughly 1 per day. This implementation is a key part of the pipeline. Without an optimised implementation of this algorithm, the upgrade will not be feasible.

Expected outcomes: We hope the ADACS project can produce the GPU-optimised code, which we will integrate into a real-time FRB detection pipeline. Computational resources required: The development and benchmarking of this code only requires access to a science-class GPU (P100 or V100). The real-time processing cluster will be funded separately.

Highlights of the approach: To our knowledge, no CUDA-optimised implementation of this code exists. The whole pipeline will produce search roughly 10 Terapixels per second – equivalent to 1 million people watching Youtube.