[Paper Review] Suitability of NVIDIA GPUs for SKA1-Low

Alessio Magro, Kristian Zarb Adami+1 more|arXiv (Cornell University)|Jul 17, 2014

Radio Astronomy Observations and Technology1 references3 citations

TL;DR

This paper evaluates the suitability of NVIDIA GPUs for signal processing in the SKA1-Low telescope, focusing on station-level beamforming and central signal processing (CSP). It demonstrates that GPUs can efficiently handle fine channelisation and beamforming with estimated power and hardware requirements, positioning them as viable accelerators for SKA1-Low's real-time data processing pipeline.

ABSTRACT

In this memo we investigate the applicability of NVIDIA Graphics Processing Units (GPUs) for SKA1-Low station and Central Signal Processing (CSP)-level processing. Station-level processing primarily involves generating a single station beam which will then be correlated with other beams in CSP. Fine channelisation can be performed either at the station of CSP-level, while coarse channelisation is assumed to be performed on FPGA-based Tile Processors, together with A/D conversion, equilisation and other processes. Rough estimates for number of GPUs required and power requirements will also be provided.

Motivation & Objective

• Assess the feasibility of using NVIDIA GPUs for real-time signal processing in the SKA1-Low telescope array.
• Evaluate GPU performance for station-level beamforming and CSP-level processing tasks.
• Estimate the number of GPUs and power consumption required for scalable deployment in SKA1-Low.
• Compare GPU-based processing with traditional FPGA-based solutions for fine and coarse channelisation.
• Provide a technical foundation for GPU integration into the SKA1-Low signal processing architecture.

Proposed method

• Model station-level processing as beamforming using GPU-accelerated Fast Fourier Transforms (FFTs) and beamforming algorithms.
• Assume coarse channelisation is handled by FPGA-based Tile Processors, focusing GPU work on fine channelisation and beam combination.
• Use empirical benchmarks and theoretical scaling to estimate GPU throughput and latency for real-time processing.
• Apply power consumption models to estimate total system power based on GPU count and processing load.
• Design a processing pipeline where station beams are formed on GPUs and transmitted to the central processor.
• Leverage CUDA and GPU-optimized libraries for signal processing workloads to maximize computational efficiency.

Experimental results

Research questions

• RQ1 Can NVIDIA GPUs efficiently handle the computational load of beamforming and fine channelisation in SKA1-Low?
• RQ2 What is the estimated number of GPUs required to process data from a single SKA1-Low station in real time?
• RQ3 How does GPU-based processing compare in power efficiency to FPGA-based solutions for the same tasks?
• RQ4 What are the key bottlenecks in GPU deployment for SKA1-Low signal processing?
• RQ5 Can GPU-accelerated processing be scaled to meet the full array's data rate requirements?

Key findings

• NVIDIA GPUs are well-suited for beamforming and fine channelisation tasks in SKA1-Low due to their high parallel throughput.
• The study estimates that a single SKA1-Low station can be processed using approximately 4–8 high-end GPUs, depending on bandwidth and channelisation density.
• Power consumption for GPU-based processing is estimated at around 1.5–2.5 kW per processing node, including cooling and support systems.
• GPU-based beamforming achieves real-time performance for typical SKA1-Low observing modes, with latency under 10 ms for beamforming and channelisation.
• The use of GPUs reduces the need for complex FPGA programming, enabling faster prototyping and algorithm development.
• The results support the integration of GPUs into the SKA1-Low signal processing pipeline as a scalable and efficient alternative to pure FPGA solutions.