An FPGA Based Computing Platform for Adaptive Optics Control

Abstract

We outline a concept design for an FPGA based real-time adaptive optics controller that meets the requirements of the Thirty Meter Telescope (TMT) first light Narrow Field Infrared Adaptive Optics system (NFIRAOS). The data and processing rates for NFIRAOS are very demanding. It has pixel input rates of about 5 GB/s and processing requirements of order several hundred GMACs. The hardware proposed is programmable and flexible, and can be used in other AO systems. It is based on a custom FPGA compute blade housed with COTS interface and computer boards in an ATCA chassis. The system accepts data from eight wave front sensors at 10 Gbps each via thirty-two serial front panel data ports (sFPDPs). Nine FPGA blades provide about 2.7 TMAC/s of processing power and are used to perform pixel processing, tomography and deformable mirror fitting. Two general-purpose computer boards are used to perform background processing and control, each with three Intel L7400 dual core processors. The system is also capable of storing 90 TB of telemetry data at rate of 3.5 GB/s. Turbulent atmosphere distorts the wavefront and limits the resolution of large ground-based optical telescopes, such as the proposed Thirty Metre Telescope (TMT). Such distortions can largely be com- pensated by adaptive optic controllers that re-construct the distortions along the light path and adjust deformable mirrors to improve the resolution of the telescope. Re-construction is an inverse problem that by nature is compute intensive. The problem is exacerbated by a large aperture and the time vari- ability of the atmosphere. In the case of the TMT, megabytes of data need to be processed within a few hundred microseconds with low jitter and latency. Such a computing load is of order 10 11 op- erations per second and cannot be handled practically using general purpose computing approaches. Such loads, however, are commonly handled by FPGA (Field Programmable Gate Array) based dig- ital signal processing systems. What is atypical is using them to solve inverse type problems. In this paper we present our investigations on the feasibility of such an approach. We start by reviewing the requirements of such a system and compare the trade-o s of various alternative approaches. We then outline our concept design and the results from our benchmark measurements. In the final section we provide conclusions and a technology outlook for future real-time controllers.