[Paper] astroCAMP: A Community Benchmark and Co-Design Framework for Sustainable SKA-Scale Radio Imaging

Source: arXiv - 2512.13591v1

Overview

The paper presents astroCAMP, a new benchmarking and co‑design framework aimed at making the massive radio‑imaging workloads of the Square Kilometre Array (SKA) both high‑performance and energy‑efficient. By unifying scientific‑quality metrics with hardware‑performance and sustainability indicators, astroCAMP gives engineers a concrete way to balance image fidelity against power, cost, and carbon footprints.

Key Contributions

• Unified metric suite that simultaneously measures scientific fidelity, runtime, energy use, carbon emissions, and total cost of ownership.
• Standardised SKA‑scale datasets and reference outputs, enabling reproducible benchmarking across CPUs, GPUs, and emerging accelerators (e.g., FPGAs).
• Multi‑objective co‑design formulation that links fidelity constraints to hardware‑level decisions, exposing Pareto‑optimal trade‑offs.
• Open‑source reproducibility kit (datasets, scripts, and benchmark results) released to the community.
• Case studies on the WSClean and IDG imaging pipelines using an AMD EPYC 9334 CPU and an NVIDIA H100 GPU, plus a demonstration of heterogeneous CPU‑FPGA design‑space exploration.

Methodology

1. Metric Definition – The authors first identify four pillars:

   • Scientific fidelity (e.g., image residuals, dynamic range)
   • Computational performance (throughput, latency)
   • Sustainability (energy consumption, carbon‑equivalent emissions)
   • Economic (hardware acquisition cost, operational OPEX)

2. Benchmark Dataset Curation – Realistic SKA observation data are distilled into a set of representative visibilities, together with “gold‑standard” reference images produced by a high‑precision pipeline.

3. Reference Implementations – Two widely used imaging codes (WSClean, IDG) are compiled for the target platforms (CPU, GPU).

4. Co‑Design Optimization Loop –

   • Run each implementation on the benchmark data, collecting the full metric suite.
   • Feed the results into a multi‑objective optimizer that searches for configurations (e.g., thread counts, precision levels, accelerator off‑loading) that satisfy user‑defined fidelity thresholds while minimizing energy or cost.

5. Design‑Space Exploration – Extend the optimizer to heterogeneous CPU‑FPGA setups, evaluating how off‑loading specific kernels (e.g., FFT, gridding) changes the trade‑offs.

All steps are scripted and containerised, so other teams can plug in their own pipelines or hardware and obtain comparable results.

Results & Findings

• Utilisation Gap Confirmed – Baseline runs of WSClean and IDG on the AMD EPYC 9334 achieved only 5‑12 % of the processor’s peak FLOPS, mirroring the community’s known inefficiency.
• Energy Savings via Precision Tuning – Switching from double‑precision to mixed‑precision kernels on the NVIDIA H100 cut energy consumption by ~30 % while keeping image residuals within a 2 % error margin.
• Pareto Front Identification – The optimizer highlighted three sweet spots:
  1. CPU‑only – lowest acquisition cost, highest energy per image.
  2. GPU‑accelerated – best runtime, moderate energy use.
  3. CPU‑FPGA hybrid – slightly higher hardware cost but up to 45 % lower carbon‑to‑solution for the same fidelity.
• Reproducibility Kit Validation – Independent researchers reproduced the benchmark numbers within 3 % variance, confirming the robustness of the dataset and metric collection pipeline.

Practical Implications

• For SKA Operations – astroCAMP gives system architects a data‑driven way to meet the SKA’s strict power caps (≈ 10 MW for imaging) without sacrificing scientific output, directly translating into lower OPEX and carbon bills.
• For Developers of Imaging Pipelines – The metric suite highlights which code paths are energy‑hungry, encouraging targeted optimisations (e.g., kernel fusion, precision scaling).
• For Hardware Vendors – The framework provides a concrete benchmark suite that goes beyond FLOPS, allowing GPU, CPU, and accelerator manufacturers to showcase real‑world sustainability performance.
• For Cloud / Edge Providers – The multi‑objective model can be used to price SKA‑scale workloads based on carbon‑aware SLAs, opening the door to greener “science‑as‑a‑service” offerings.

Limitations & Future Work

• Fidelity Metric Scope – The current suite focuses on residual‑based metrics; more nuanced astrophysical quality measures (e.g., source detection completeness) are still needed.
• Hardware Diversity – Benchmarks were limited to one CPU and one GPU model; broader coverage (ARM, newer FPGA families, ASICs) will improve generalisability.
• Dynamic Workloads – The study assumes static datasets; extending astroCAMP to streaming or real‑time imaging pipelines (e.g., transient detection) is an open challenge.
• Automation – While the optimizer is functional, tighter integration with CI pipelines and auto‑tuning frameworks would make continuous co‑design easier for developers.

astroCAMP sets a solid foundation for making the SKA’s petascale imaging both scientifically powerful and environmentally responsible—an essential step as we head toward the next generation of data‑intensive astronomy.

Authors

• Denisa-Andreea Constantinescu
• Rubén Rodríguez Álvarez
• Jacques Morin
• Etienne Orliac
• Mickaël Dardaillon
• Sunrise Wang
• Hugo Miomandre
• Miguel Peón-Quirós
• Jean‑François Nezan
• David Atienza

Paper Information

• arXiv ID: 2512.13591v1
• Categories: cs.DC, astro-ph.IM, cs.PF
• Published: December 15, 2025
• PDF: Download PDF