Suitability of GPUs for real-time control of large astronomical adaptive optics instruments

Adaptive optics (AO) is a technique for correcting aberrations introduced when light propagates through a medium, for example, the light from stars propagating through the turbulent atmosphere. The components of an AO instrument are: (1) a camera to record the aberrations, (2) a corrective mechanism to correct them, (3) a real-time controller (RTC) that processes the camera images and steers the corrective mechanism on milliseconds timescales. We have accelerated the image processing for the AO RTC with the use of graphics processing units (GPUs). It is crucial that the image is processed before the atmospheric turbulence has changed, i.e., in one or two milliseconds. The main task is to transfer the images to the GPU memory with a minimum delay. The key result of this paper is a demonstration that this can be done fast enough using commercial frame grabbers and standard CUDA tools. Our benchmarking image consists of pixels out of which are used in processing. The images are characterized and reduced into a set of 9248 numbers; about one-third of the total processing time is spent on this characterization. This set of numbers is then used to calculate the commands for the corrective system, which takes about two-third of the total time. The processing 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. The variation in processing time (jitter) has a root-mean-square value of 20-30 s and about one outlier in a million cycles.