GPU-Based Convolution Pipeline Processes 1.2 Petabytes of Wide-Field Photometry
What Was Observed (Empirical Data)
Researchers developed GPUPhot, a point-source detection and photometry pipeline that replaces iterative per-source fitting with full-frame convolutions and principal-component Point Spread Function (PSF) eigenmodes. Operating in production at the Teide Observatory since 2023, the software has reduced 3.7 million images (1.2 PB) at rates under 15 seconds per 150-megapixel frame, achieving sub-pixel astrometric accuracy and few-millimagnitude photometric precision for bright sources.
The Reality Check (Anti-Hype Analysis)
This represents a computational engineering optimization for wide-field survey throughput rather than a new physical discovery. The quoted few-millimagnitude precision applies specifically to bright point sources, and the linearized eigenmode decomposition relies on the assumption that spatial PSF variations across a frame can be accurately represented by a small number of fixed principal components.
PRIMARY SOURCE:
arXiv Astrophysics (astro-ph)