FPGA based microserver for high performance real-time computing in Adaptive Optics
FPGA based microserver for high performance real-time computing in Adaptive Optics
复制标题
基于 FPGA 的微服务器，用于自适应光学中的高性能实时计算
DOI：
--
复制
发表时间：
2017
期刊：
影响因子：
--
通讯作者：
D. Gratadour
中科院分区：
文献类型：
--
作者：
C. Patauner;R. Biasi;M. Andrighettoni;G. Angerer;D. Pescoller;F. Porta;D. Gratadour
面向ELT的自适应光学控制系统的实时数据流水线要求非常大的通信带宽、灵活的接口、高计算能力和存储带宽、低延迟和低抖动。Microgate在GreenFlash项目的框架内开发了一种潜在的解决方案，可以满足所有这些具有挑战性的要求，该解决方案基于两个专门开发的FPGA板，具有出色的能效。μXComp板专用于高吞吐量计算任务，特别关注内存带宽，这通常是大量矩阵向量乘法的限制因素，同时仍具有灵活的高带宽接口。第二块板称为μXLink，专用于智能接口，并具有片上系统ARM处理器，可实现独立的微服务器概念。在微服务器中，μXLink板互连并仲裁各种加速和接口板的操作，例如μXComp板或GPU加速器，而不需要主机。这样的微服务器概念可以适于满足不同AO配置的要求，与各种波前相机和可变形反射镜兼容，并且还在AO系统内执行其他专用计算和控制任务。我们报告的设计，性能和测试上的硬件实现到目前为止。
The real-time data pipeline of ELT-oriented adaptive optics control systems requires very large communication bandwidth, flexible interfacing, high computational power and memory bandwidth, low latency and low jitter. A potential solution to all these challenging requirements has been developed by Microgate in the frame of the GreenFlash project and is based on two specifically developed FPGA boards with PCIe backplane interface with excellent energy efficiency. The μXComp board is dedicated to high throughput computational tasks, with particular focus on the memory bandwidth, which is often the limiting factor for heavy matrix-vector multiplication, while still featuring flexible, high bandwidth interfaces. The second board, called μXLink, is dedicated to the smart interfaces and features a system-on-chip ARM processor enabling the implementation of a stand-alone Microserver concept. In the Microserver, the μXLink board interconnects and arbitrates the operation of various acceleration and interface boards, e.g. μXComp boards or GPU accelerators, without requiring a host computer. Such Microserver concept can be adapted to satisfy the requirements of different AO configurations, being compatible with various wavefront cameras and deformable mirrors, and perform also other dedicated computation and control tasks within an AO system. We report the design, the performances and the tests performed on the hardware realized so far.