Memory-aware optimization of FPGA-based space systems
Memory-aware optimization of FPGA-based space systems
复制标题
基于 FPGA 的空间系统的内存感知优化
DOI：
10.1109/aero.2015.7119029
复制
发表时间：
2015
期刊：
影响因子：
--
通讯作者：
A. Gordon
中科院分区：
文献类型：
--
作者：
N. Wulf;A. George;A. Gordon
设计基于现场可编程门阵列的空间系统，以满足高性能、低功耗和高可靠性的任务目标是具有挑战性的。我们以前的工作建立了一个框架，帮助基于FPGA的空间系统的设计者考虑广泛的设计，评估这些设计的功率和可靠性，并将巨大的设计搜索空间缩小到显著减少的帕累托最优集合。为了进一步提高和扩展我们的框架评估和优化日益复杂的航天系统的能力，本文详细介绍了我们的框架的内存扩展，它通过对我们框架的原始分析进行细化来实现内存感知分析。内存感知分析通过对三种内存资源进行建模，更准确地预测系统的功率和可靠性：内存容量、内存带宽和外部内存带宽。我们通过研究一个基于高光谱成像卫星任务的增强版本的案例研究，展示了我们框架的内存扩展的重要性。在分析了22个独特的Virtex FPGA器件并对每个器件进行功耗和可靠性优化后，该框架选择了从极低功耗到高可靠性的四种帕累托最优设计。该框架的内存扩展结果表明，内存资源可能会限制基于FPGA的空间系统设计的性能，并对功耗和可靠性结果做出显著贡献。
Designing FPGA-based space systems that meet mission goals of high performance, low power, and high dependability is challenging. Our previous work established a framework to help designers of FPGA-based space systems to consider a wide range of designs, evaluate the power and dependability of those designs, and narrow the large design search space down to a significantly reduced Pareto-optimal set. To further improve and extend our framework's ability to evaluate and optimize increasingly complex aerospace systems, this paper details our framework's memory extension, which enables memory-aware analysis by refinements to our framework's original analysis. The memory-aware analysis more accurately predicts a system's power and dependability by modeling three memory resources: internal-memory capacity, internal-memory bandwidth, and external-memory bandwidth. We demonstrate the importance of our framework's memory extension by investigating a case study based on an enhanced version of a hyperspectral-imaging satellite mission. After analyzing 22 unique Virtex FPGA devices and optimizing each for power and then dependability, the framework selects four Pareto-optimal designs, ranging from very-low power to high dependability. Results of the framework's memory extension show that memory resources may limit the performance of an FPGA-based space-system design and contribute significantly towards power and dependability results.