Perryman, Noah
(2024)
Strategies to Enable Reliable Next-Generation Applications on Embedded Space Platforms.
Doctoral Dissertation, University of Pittsburgh.
(Unpublished)
Abstract
Space computing has considerable performance restrictions that are imposed by the limited onboard-processing capabilities provided by heritage single-board computers. Due to these limitations, current state-of-the-art devices for space edge computing are unable to meet the resource and performance requirements for next-generation communication, navigation, and artificial intelligence (AI) applications planned for future science and defense missions. To address these issues, domain-specific architectures with specialized acceleration hardware, such as the Xilinx Versal Adaptive System-on-Chip architecture, have been developed. This heterogeneous platform provides significant energy-efficient compute capabilities, but it is susceptible to radiation-induced single-event effects and therefore the dependability of the device must be characterized prior to inclusion on future space-computing platforms.
Conversely, there is a direct need for high-visibility NASA missions that provide significant scientific impact or have a high mission class using completely radiation-hardened electronics solutions, to enable AI applications in harsh environments despite severe size, weight, and power constraints. For these missions, where current state-of-the-art solutions such as the Versal are too power-demanding or are incapable of surviving the intended radiation environment, an alternative radiation-hardened processing architecture that can leverage the control-flow capabilities of scalar processors while also incorporating the hardware-acceleration capabilities of an FPGA is of significant value.
In this research, performance, energy-efficiency, and resource-utilization tradeoffs for the Versal were evaluated by benchmarking a suite of representative next-generation AI and communication applications for space. These experiments included four image-classification models based on convolutional neural network architectures and the multitaper spectral estimation algorithm, a relevant communication algorithm based on the fast Fourier transform. Next, a methodology for evaluating and increasing the dependability of semantic segmentation deep learning models on heterogeneous systems featuring FPGAs as well as other compute elements is proposed. To demonstrate this methodology, three semantic segmentation DL models accelerated on the AMD-Xilinx Deep Learning Processing Unit for the Versal are evaluated. Lastly, an alternative architecture to the Versal for ultra-low power, high-reliability spaceflight applications is investigated. This investigation resulted in the design of an architecture with the highly reliable, radiation-hardened GR740 scalar processor paired with the low-power, radiation-tolerant CertusPro-NX-RT FPGA for increased performance.
Share
| Citation/Export: |
|
| Social Networking: |
|
Details
| Item Type: |
University of Pittsburgh ETD
|
| Status: |
Unpublished |
| Creators/Authors: |
|
| ETD Committee: |
|
| Date: |
6 September 2024 |
| Date Type: |
Publication |
| Defense Date: |
23 July 2024 |
| Approval Date: |
6 September 2024 |
| Submission Date: |
20 July 2024 |
| Access Restriction: |
No restriction; Release the ETD for access worldwide immediately. |
| Number of Pages: |
139 |
| Institution: |
University of Pittsburgh |
| Schools and Programs: |
Swanson School of Engineering > Electrical and Computer Engineering |
| Degree: |
PhD - Doctor of Philosophy |
| Thesis Type: |
Doctoral Dissertation |
| Refereed: |
Yes |
| Uncontrolled Keywords: |
Versal, FPGA, artificial intelligence, AI Engine, deep learning, semantic segmentation, earth observation, single-event effects, fault injection, space computing |
| Date Deposited: |
06 Sep 2024 20:04 |
| Last Modified: |
06 Sep 2024 20:04 |
| URI: |
http://d-scholarship.pitt.edu/id/eprint/46708 |
Metrics
Monthly Views for the past 3 years
Plum Analytics
Actions (login required)
 |
View Item |
![[feed]](/images/feed-icon-32x32.png){kind=icon}