The GPU-Resident Astronomy Pipeline
A GPU-resident astronomy pipeline keeps every stage of image analysis — load, align, subtract, detect, review — inside GPU memory instead of round-tripping through the CPU.
Loading post…
NVIDIAQuynh L. NguyenAnnouncement
Faster Scientific Image Analysis with NVIDIA cuPhoton
Summary
NVIDIA cuPhoton is an open-source CUDA-X toolkit of GPU-accelerated building blocks for scientific image analysis, from sensor read through classification, covering spectral and optical astronomy and time-domain laser/X-ray analysis. High-throughput instruments like the Vera C. Rubin Observatory generate up to 20 TB of images per night, but CPU-bound pipelines create delays of hours to months. cuPhoton keeps data on the GPU end-to-end, collapsing nine months of data analytics to four hours on GPU-accelerated Python. The toolkit includes modules for FITS loading, image registration, PSF-matching and subtraction, dipole fitting, and candidate review. Benchmarks on multi-GPU systems show individual operation speedups of up to 14,900x for loading and 14,550x for processing.
Context
Modern scientific instruments like the Vera C. Rubin Observatory generate terabytes of image data per night, but CPU-bound processing pipelines create delays of hours to months. NVIDIA cuPhoton addresses this by providing a GPU-native pathway from sensor read to classification, using modules for loading, alignment, PSF-matching, subtraction, dipole fitting, and candidate review.
Approach / What changed
cuPhoton provides GPU-accelerated modules for FITS loading (xDataReader), image registration (xRep), PSF-matching and subtraction (xPois), dipole fitting (xFit), candidate review (xScan), and X-ray analysis (xRay). Data stays on the GPU from sensor read through classification, eliminating CPU-GPU transfers. The toolkit scales across multi-GPU, multi-node systems using MPI.
Takeaways
cuPhoton accelerates individual operations by up to 14,900x for image loading and 14,550x for signal processing on multi-GPU systems, reducing months of computation to minutes.
The toolkit keeps data on the GPU throughout the pipeline, from FITS loading with xDataReader through to candidate classification with xScan, avoiding costly CPU-GPU transfers.
cuPhoton includes a candidate review interface (xScan review-bokeh) that ranks candidates by model uncertainty or known errors to limit the number requiring human inspection.
FAQ
What is the cuPhoton xDataReader component?
xDataReader accelerates loading FITS image data directly onto the GPU. It splits the workload: CFITSIO plans byte ranges on the CPU, KvikIO reads them using GPUDirect Storage when available, and nvCOMP decompresses GZIP tiles, bypassing CPU-side parsing and PCIe transfers for science kernels.
How does cuPhoton handle observation-to-observation variations?
The xRep module aligns images by resampling each visit onto a shared grid using GPU-accelerated gridding. This accounts for differences in telescope pointing, rotation, and pixel scale, ensuring the same star falls on the same pixel across exposures.
Can cuPhoton scale across multiple GPUs?
Yes, cuPhoton uses MPI to distribute image pairs across GPU workers. After initial registration, it keeps image arrays on the device through subtraction, dipole fitting, and classification. The distributed execution guide supports Slurm, SSH launches, and the Dragon alternative for multi-node systems.
Source and freshness
- Source
- developer.nvidia.com
- Published
- Added to Yomu
Summaries are generated from the original by a language model. Check the original before relying on it.
See the idea
Visual explainers
Keep reading
Related reading
The Good Stuff
Paid picks we use ourselves. Links carry a referral code — a subscription through one earns us a commission at no extra cost to you.
The best engineering reads, once a week
One useful email. Unsubscribe anytime.