Demonstration of the suitability of GPUs for AO real-time control at ELT scales

We have implemented the full AO data-processing pipeline on Graphics Processing Units (GPUs), within the framework of Durham AO Real-time Controller (DARC). The wavefront sensor images are copied from the CPU memory to the GPU memory. The GPU processes the data and the DM commands are copied back to the CPU. For a SCAO system of 80x80 subapertures, the rate achieved on a single GPU is about 700 frames per second (fps). This increases to 1100 fps (1565 fps) if we use two (four) GPUs. Jitter exhibits a distribution with the root-mean-square value of 20 mu s - 30 mu s and a negligible number of outliers. The increase in latency due to the pixel data copying from the CPU to the GPU has been reduced to the minimum by copying the data in parallel to processing them. An alternative solution in which the data would be moved from the camera directly to the GPU, without CPU involvement, could be about 10%-20% faster. We have also implemented the correlation centroiding algorithm, which - when used - reduces the frame rate by about a factor of 2-3.