Deep neural networks (DNNs) are computationally and memory intensive, which makes them difficult to deploy on traditional hardware environments. Therefore, many dedicated solutions have been proposed in the literature and market. However, most of them remain proprietary or lack maturity, thus preventing the adoption of deep-learning (DL) based software in new application domains. The Nvidia Deep-Learning Accelerator (NVDLA) is a free and open architecture that aims at promoting a standard way of designing deep neural network (DNN) inference engines. Following an analogy with open-source software, which is downloaded and executed, open hardware is likely to use FPGAs as reference implementation platform. However, tailoring accelerator configuration to the capacity of cost-effective reconfigurable logic remains a fundamental challenge for their actual deployment in system-level designs. This chapter presents an overview of the hardware and software components of the NVDLA inference framework, and reports on the exploration of its configuration space. It explores the resource utilization-performance trade-offs spanned by the main precompiled NVDLA accelerator configurations on top of the mainstream Zynq UltraScale+ MPSoC. For the sake of comprehensive end-to-end performance characterization, the inference rate of the software stack and of the accelerator hardware are matched, thus identifying current bottlenecks and promising optimization directions.