[Paper Review] Cutting the cost of pulsar astronomy: Saving time and energy when searching for binary pulsars using NVIDIA GPUs

Jack White, Karel Adámek+1 more|arXiv (Cornell University)|Nov 24, 2022

Radio Astronomy Observations and Technology4 citations

TL;DR

This paper proposes a hybrid optimization strategy combining mixed-precision computing (bfloat16) and dynamic frequency scaling on NVIDIA A100 GPUs to reduce energy consumption in Fourier Domain Acceleration Search (FDAS) for binary pulsar detection. By lowering GPU clock frequencies and using lower-precision arithmetic, the method achieves a 58% reduction in energy usage with less than 3% loss in numerical sensitivity, significantly improving energy efficiency without sacrificing performance.

ABSTRACT

Using the Fourier Domain Acceleration Search (FDAS) method to search for binary pulsars is a computationally costly process. Next generation radio telescopes will have to perform FDAS in real time, as data volumes are too large to store. FDAS is a matched filtering approach for searching time-domain radio astronomy datasets for the signatures of binary pulsars with approximately linear acceleration. In this paper we will explore how we have reduced the energy cost of an SKA-like implementation of FDAS in AstroAccelerate, utilising a combination of mixed-precision computing and dynamic frequency scaling on NVIDIA GPUs. Combining the two approaches, we have managed to save 58% of the overall energy cost of FDAS with a (<3%) sacrifice in numerical sensitivity.

Motivation & Objective

• To reduce the energy consumption of FDAS, a computationally intensive method for detecting binary pulsars in time-domain radio astronomy datasets.
• To address the growing energy demands of next-generation radio telescopes like the SKA, which require real-time processing of massive data volumes.
• To explore energy-efficient alternatives to traditional high-performance computing by optimizing GPU workloads using precision reduction and clock scaling.
• To maintain numerical sensitivity while minimizing power draw, thereby reducing CO2 emissions from datacenter-scale astronomical computing facilities.

Proposed method

• Employed mixed-precision computing by converting the most computationally intensive convolution stage of FDAS to bfloat16, reducing memory footprint and data movement.
• Applied dynamic frequency scaling to lower the GPU core clock frequency during FDAS execution, reducing power draw despite increased execution time.
• Used an SKA-like configuration (600 s observation time, 64 μs time resolution, 5925 DM trials, 193 acceleration trials) for realistic performance evaluation.
• Measured energy and time trade-offs across different GPU frequencies and precision levels using NVIDIA A100 and RTX 4090 GPUs.
• Evaluated the impact of precision reduction on numerical sensitivity, finding a maximum 2.7% error in detection strength magnitude.
• Combined both techniques to assess synergistic effects on energy efficiency and performance.

Experimental results

Research questions

• RQ1 Can mixed-precision computing with bfloat16 reduce energy consumption in FDAS without significantly degrading detection sensitivity?
• RQ2 Does dynamic frequency scaling reduce GPU power draw more effectively than performance gains, leading to net energy savings in bandwidth-bound workloads?
• RQ3 What is the optimal balance between execution time and energy usage when combining mixed-precision and dynamic frequency scaling in FDAS?
• RQ4 How do newer GPU architectures (e.g., A100, RTX 4090) compare in energy efficiency for FDAS under these optimizations?
• RQ5 To what extent can energy savings be achieved in SKA-scale FDAS workloads while maintaining acceptable sensitivity and performance?

Key findings

• A 58% reduction in energy usage was achieved by combining mixed-precision computing (bfloat16) and dynamic frequency scaling on an NVIDIA A100 GPU.
• The maximum numerical error introduced by bfloat16 conversion in the convolution stage was 2.7% in detection strength magnitude, resulting in less than 3% sensitivity loss.
• Energy savings were maximized at lower GPU clock frequencies, even though execution time increased, demonstrating that bandwidth-bound workloads benefit from reduced clock speeds.
• The trade-off between speed and energy efficiency showed diminishing returns at higher frequencies, confirming the bandwidth-bound nature of FDAS.
• Dynamic frequency scaling alone reduced power draw significantly, with the most efficient configuration occurring at lower frequencies despite longer execution times.
• The combination of both techniques yielded synergistic energy savings, outperforming either method applied in isolation.