Accelerated Deep-Learning Inference on the Versal adaptive SoC in the Space Domain

Artificial intelligence has found its way into space and necessitates a powerful and flexible hardware platform to keep up with the fast-paced AI domain. With the space-grade variant of the Versal, AMD-Xilinx offers one of the first space-ready AI accelerators that combine multiple compute paradigms, i.e., scalar processing (CPU), adaptive engines (FPGA), and vector processing (AI-Engine array) into an adaptive System-on-Chip. This paper provides a thorough analysis of its AI capabilities with respect to throughput and power efficiency for Multi-Layer Perceptrons and CNNs, and takes a look under the hood by profiling the system’s efficiency on an architectural level based on the idea of the Roofline model. We believe that the gained insights ultimately help to design optimal NN architectures for deployment on the Versal.