Evaluation of Xilinx Versal Architecture for Next-Gen Edge Computing in Space

Space edge computing has unique considerations (e.g., size, power, space radiation, etc.) that limit the performance capabilities of achievable onboard processing. Due to these restrictions, current state-of-the-art devices for space edge computing are unable to meet the resource and performance requirements for next-gen communication, navigation, and artificial intelligence (AI) applications planned for future science and defense missions. To address these issues, space designers are considering domain-specific architectures (DSAs) with specialized acceleration hardware, such as the Xilinx Versal Adaptive Compute Acceleration Platform (ACAP) architecture. This platform is heterogeneous and provides developers with scalar/vector processing units and programmable logic for mission-specific customization. In this paper, the performance and power-efficiency tradeoffs of different Versal ACAP processing subsystems (i.e., AI Engines, ARM Cortex-A72, ARM Cortex-R5F, and programmable logic) were evaluated by comparing execution of representative space applications. These experiments included three convolutional neural network (CNN)-based image-classification applications (i.e., MobileNetV1, ResNet-50, and GoogLeNet) for AI and a multitaper spectral estimation application for communication. Notably, the vector-based Versal AI Engines showed a promising performance improvement over the ARM Cortex-A72 and Cortex-R5F while usually being outperformed by the programmable logic. However, the AI Engines typically consumed significantly less power, with some AI engine applications being 1.48× more energy efficient compared to the programmable logic while not losing static power to excessive resource utilization. Overall, this evaluation demonstrates the versatility and heterogeneity of the Versal architecture and the tradeoffs between the on-chip subsystems. Therefore, the Versal will be considered for a future single-board computer to address challenges set by demanding next-gen space applications.