David Rolfes
Accelerating Deep Reinforcement Learning with the Xilinx Versal Architecture

Abstract
In this thesis, a reinforcement learning algorithm is implemented for the ”Versal AI Core Series VCK190 Evaluation Kit”, which is part of the new Versal series and includes a new architecture called AI Engine. The deep Q-network algorithm is implemented, and realistic expectations for performance and neural network architecture are set by a motor control unit used as the evaluation scenario. Despite some shortcomings, the implementation manages to achieve high parallelism and, therefore, high performance. It beats an FPGA-based implementation for more than factor of 10 and is more than fast enough to process all data generated by the scenario. This showcases the impressive AI Engines performance.