The goal of the eki project is to increase the energy efficiency of AI systems for deep neural network (DNN) inference through approximation techniques and mapping to high-end FPGA systems. DNNs have emerged in recent years as an essential approach to statistical machine learning. In the inference phase, classifications or regressions are computed on typically very large datasets, which already provides a significant computational load and associated CO2 emissions. As the demand for DNN inference will continue to grow strongly, there is a high need for action. FPGAs are a particularly well-suited technology for DNN inference because their hardware reconfigurability allows them to be optimally adapted to the application. In this project, based on the open source tool FINN, a software tooflow is developed that automates, optimizes and hardware-adapts DNNs. The approaches followed are the approximation techniques of network pruning and low precision quantization, as well as parallelization on an FPGA cluster. Subsequently, the achieved energy savings will be characterized by precise measurements in real server systems. Other aspects of the project include the development of an AutoML method for energy optimization and experimental evaluation using test DNNs and two use case studies from the areas of natural language processing and optimization in agriculture.