GPUs for adaptive optics: simulations and real-time control

With the emergence of General Purpose computations on Graphic Processing Units (GP-GPUs) this architecture has become amazingly attractive for large scale applications such as numerical simulations of complex systems. While the number of degrees of freedom of an adaptive optics (AO) system scales with the square of the telescope diameter, the system model exhibits a rather high level of parallelism especially when simulating Shack-Hartmann (SH) wavefront sensors (WFS). The use of massively parallel devices such as GPUs to simulate next generation AO systems for the European Extremely Large Telescope (E-ELT) thus makes a lot of sense. Our team has developed such simulation tools and first results show that speeds of about a thousand of iterations per second were achievable on a single high-end GPU for an eXtreme (X)AO system such as SPHERE including a single layer turbulence model generated on-the-fly. These numerical models include all the operations executed by the real-time controller (RTC) of a real system. The achieved simulation speeds show that a single high-end GPU could drive a XAO system on the VLT and, depending on the centroiding algorithm and the control scheme chosen, could even drive a classical AO system on the E-ELT. While the main challenge resides in the data transfer speed to and from the GPU, developing and testing AO control algorithms for the simulation code on the same hardware as the system RTC would bring a lot of benefits. In this paper we present the simulation results as well as strategies to build GPU-powered AO systems.