Table 3. GPU processing times on NVIDIA QUADRO P6000 of the proposed online algorithms, as shown using a NVIDIA profiler.

They are compatible with the available 1 ms time slot of the foreseen SSX experiment. For comparison with real-world use cases, we provide numbers for the yet to be published peak_finder algorithm from the GPU pyFAI suite. This algorithm implements the peakfinder8 algorithm from CrystFEL on GPUs.

	Kernel execution time (in nanoseconds for one image of the given size)
Data batch (image + pedestal)	JUNGFRAU raw-data pre-processing†	Pre-processing + ‘simplistic’ Bragg’s peaks count†	Pre-processing + CSR matrix compression‡	Peak_finder (pyFAI)§
No. of image × size
10 × 500K pixels	28	31	147	N/A
10 × 4M pixels	219	240	469	2167
10 × 2070 × 2167 (JUNGFRAU)	212	240	492	N/A

^†Performed on a stack of images.

^‡Performed sequentially.

^§OpenCL implementation, for reference.