Results of the NFIRAOS RTC trade study

Abstract

With two large deformable mirrors with a total of more than 7000 actuators that need to be driven from the measurements of six 60x60 LGS WFSs (total 1.23Mpixels) at 800Hz with a latency of less than one frame, NFIRAOS presents an interesting real-time computing challenge. This paper reports on a recent trade study to evaluate which current technology could meet this challenge, with the plan to select a baseline architecture by the beginning of NFIRAOS construction in 2014. We have evaluated a number of architectures, ranging from very specialized layouts with custom boards to more generic architectures made from commercial off-the-shelf units (CPUs with or without accelerator boards). For each architecture, we have found the most suitable algorithm, mapped it onto the hardware and evaluated the performance through benchmarking whenever possible. We have evaluated a large number of criteria, including cost, power consumption, reliability and flexibility, and proceeded with scoring each architecture based on these criteria. We have found that, with today's technology, the NFIRAOS requirements are well within reach of off-theshelf commercial hardware running a parallel implementation of the straightforward matrix-vector multiply (MVM) algorithm for wave-front reconstruction. Even accelerators such as GPUs and Xeon Phis are no longer necessary. Indeed, we have found that the entire NFIRAOS RTC can be handled by seven 2U high-end PC-servers using 10GbE connectivity. Accelerators are only required for the off-line process of updating the matrix control matrix every ~10s, as observing conditions change.

The Real-Time controller (RTC) for the Narrow Field Infrared Adaptive Optics System (NFIRAOS) for TMT final design

“…Throughout the development of the RTC, several different hardware platforms have been investigated, including field-programmable gate array (FPGA) based systems and commercial off the shelf (COTS) CPU-based servers, the latter both with and without graphics processing units (GPU). A previous trade study ("Results of the NFIRAOS RTC trade study" [3]), with updated analysis also presented at the December 2016 Paris RTC4AO4 workshop concluded that a purely CPU-based solution was within the capabilities of existing servers at the time, and would be the least complex, and result in reduced risks (e.g. schedule and obsolescence), though with an increase in hardware cost and power consumption over the other approaches.…”

“…Building upon the RTC Preliminary Design that was described in [1] [2] [3], this paper summarizes updated architectural details and prototyping results that were presented as part of the RTC's Final Design, which successfully passed its external review in December 2017. This involved external reviewers from key international organizations.…”

Spatio-angular minimum-variance tomographic controller for multi-object adaptive-optics systems

“…We have only considered the MMSE implemented as a full VMM since there is growing evidence that this option maps well into GPUs and should therefore be preferred to iterative implementations [30]. For the time being we exclude from this analysis other considerations such as memory bandwidth and pipelinability.…”

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

“…GPUs excel in cost, floating-point processing and development costs while FPGAs are known for their predictable timing latency, low power consumption and high-speed interfacing options. 7,8 A recent stagnation in compelling FPGA hardware, the high initial programming overhead for FPGAs and the rapid evolution of GPU and Central Processing Unit (CPU) performance has resulted in GPUs and CPUs dominating the choice of computing hardware for NFIRAOS 3,9,10 and E-ELT 11,12 AO over the years. The cost of hardware and development makes GPUs a compelling hardware choice for AO.…”