Reducing adaptive optics latency using many-core processors
Published Web Locationhttps://doi.org/10.20353/K3T4CP1131546
The high control frequency required of planned E-ELT adaptive optics modules (from several hundreds of Hz tothousands and typically in the region of 500 Hz for first light E-ELT instruments such as HARMONI), along withthe high number of actuators, leads to very demanding computational needs for the real-time AO control (RTC)systems. The number of actuators, proportional the total area of the telescope primary mirror, will range fromseveral thousands to tens of thousands (typically 4-6 thousand for first light E-ELT instruments). TraditionalRTC architectures based on CPU only technologies are typically unable to achieve the required performancein a cost-effective and maintainable manner. Alternative hardware such as many-core hardware acceleratorsneed to be considered to deliver the computational power. These many-core processors offering a highly parallelenvironment, have the potential of coping with the high computational load and of accelerating parts of the AOcontrol loop. AO systems for the E-ELT however, also put heavy constrains on acceptable levels of jitter andlatency.In this paper, we investigate novel hardware that has the potential to accelerate wavefront reconstruction andwavefront pre-processing, respectively the Intel Xeon Phi and the TILERA TILE-Gx processors. We present adetailed performance analysis putting an emphasis on execution time, jitter (i.e. variation in execution time)and outliers (i.e. results significantly apart from mean). Results are explored both for specific first light E-ELTinstruments and, to stay as general as possible and fully appreciate scalability issues, for a much wider range ofAO system sizes. The paper also addresses anticipated near future hardware developments and examines theirsuitability for the E-ELT.