Thirty Meter Telescope narrow-field infrared adaptive optics system real-time controller prototyping results

Abstract

Prototyping and benchmarking was performed for the Real-Time Controller (RTC) of the Narrow Field InfraRed Adaptive Optics System (NFIRAOS). To perform wavefront correction, NFIRAOS utilizes two deformable mirrors (DM) and one tip/tilt stage (TTS). The RTC receives wavefront information from six Laser Guide Star (LGS) ShackHartmann WaveFront Sensors (WFS), one high-order Natural Guide Star Pyramid WaveFront Sensor (PWFS) and multiple low-order instrument detectors. The RTC uses this information to determine the commands to send to the wavefront correctors. NFIRAOS is the first light AO system for the Thirty Meter Telescope (TMT).The prototyping was performed using dual-socket high performance Linux servers with the real-time (PREEMPT_RT) patch and demonstrated the viability of a commercial off-the-shelf (COTS) hardware approach to large scale AO reconstruction. In particular, a large custom matrix vector multiplication (MVM) was benchmarked which met the required latency requirements. In addition all major inter-machine communication was verified to be adequate using 10Gb and 40Gb Ethernet. The results of this prototyping has enabled a CPU-based NFIRAOS RTC design to proceed with confidence and that COTS hardware can be used to meet the demanding performance requirements.

Fast focal plane wavefront sensing on ground-based telescopes

“…Combining the 1024×1024 (rows×columns) inverse covariance matrix, 1024×3610 reference image matrix, and 3610×1 target image vector (the latter two are typically dotted to produce the target image correlation vector) into a single matrix operation produces a (1024 × 3610) · (3610 × 1) matrix multiplication. One Narrow Field InfraRed Adaptive Optics System (NFIRAOS) server processing one laser guide star wavefront sensor (WFS) multiplies a matrix of 7000×5400 by a 5400×1 vector in 500 µs, 15 and so our FAST approach should be less than this amount. Finally, the FFT cannot start before the last pixel is received, so adding a few hundred microseconds to read out the image yields (350 µs) + (less than 500 µs) + (about 200 µs) = (less than about 1.05 ms).…”

Adaptive optics program update at TMT

“…The system architecture is based upon commercial server hardware with a classical Matrix Vector Multiply (MVM) control algorithm. Intensive benchmarking have been conducted to support this architecture and have demonstrated that the performance requirements can be met by using a moderate number of existing off-the-shelf commercial boards, which in addition enable the use of a significantly simpler RTC algorithm (MVM) [14], [15] .…”

Scalable platform for adaptive optics real-time control, part 2: field programmable gate array implementation and performance

“…The most recent hardware benchmarking results point to CPUs and Intel Xeon Phi coprocessors being preferred over GPUs. 13,14 The European Southern Telescope (ESO) has also moved on from a combination of FPGAs and Digital Signal Processors (DSPs) for its AO RTC 15 to a heterogenous architecture involving CPUs and GPUs for mathematical processing and FPGAs for the high-speed interfacing. 16,17 A similar scalable AO kernel involving an FPGA-based Network Interface Card (NIC) for peer-to-peer communication, with GPUs performing the intensive computational part, is explored in Perret et al 18 A more recent approach from ESO (called Green Flash) aims to compare competing technologies of co-processors, GPUs and FPGAs for use as the computing platform for the AO system on the E-ELT.…”