Speaker
Description
The increasing interest of deep-space exploration, commercial use of small satellites, and in-situ information value extraction on space systems (e.g. CubeSat constellations, rovers, earth observation satellites) require more on-board data processing (OBDP). There are several reasons to require increased on-board data processing. Space applications, especially constellations and robotic missions will require a high degree of autonomy, intelligent task planning, data processing (e.g. image processing) and data dissemination. The sensors used will in most cases generate more data than the cross-links or downlink can handle, which drives the need for local reasoning using artificial intelligence and classical analytical models.
In this study, we explored radiation tolerant intelligent on-board processing systems accelerated with peripherals (e.g. GPU, FPGA, DSP) which take an advantage of a new heterogeneous computer architecture, Heterogeneous System Architecture (HSA), in terms of decreasing compute latency and increasing data transfer bandwidth. The study continues prior work which is commercialized by Unibap AB with flight heritage and selected by NASA for on-board processing for the “HyTI” thermal hyperspectral mission. This presentation presents the results of investigations addressing the capabilities of future on-board big data processors. Furthermore, the experimental study covers the performance analysis by using image recognition algorithms, the open standard "OpenVX", and an open source machine learning library ”MIOpen”. Furthermore, we discuss the usability of our method in OBDP regarding heterogeneous computing.
The results show that heterogeneous architectures, especially GPU can make significant improvements in compute efficiency. Heterogeneous GPU accelerated on-board processing achieves 238 times reduced compute time and approximately 13.5 times less energy compared to the traditional CPU centered processing. In addition, the heterogeneous computing method shows 20-70% improvements of the schedulability of the entire application system given different assumptions.
| Paper submission | Yes |
|---|
