Characterize GPU performance for real-time PyWFS denoising

Characterize the GPU performance of the FFT-based block-matching and global 3D wavelet-thresholding implementation for deterministic, low-latency pyramid wavefront sensor denoising.

Background

The proposed full-frame denoising algorithm requires approximately 0.79 seconds per frame on CPU hardware, which is incompatible with the kilohertz adaptive-optics control-loop rates targeted by the application. The paper identifies GPU execution as a promising route because block matching and hard thresholding are amenable to parallel, pipelined computation. However, the actual GPU execution time and its suitability for deterministic real-time operation are not established in the paper.

References

Characterisation of GPU performance for this application is identified as future work.

— Full frame denoising for pyramid wavefront sensors  (2608.19934 - Schwartz et al., 20 Aug 2026) in Section 2.2, “Hardware scaling, and execution time”