A Versatile Edge Machine Learning Test Bench for High Bandwidth Instrumentation

New scientific experiments and instruments generate large volumes of data that need to be transferred to storage or processing. Edge machine learning could allow us to reduce the amount of data before transmission, thus reducing the cost of experiments. Data compression can be done in various ways depending on the targeted experiment such as image compression or rejection of unwanted events using a veto system. The high throughput, complexity and operational costs of the instruments make debugging in the real system a difficult challenge. To facilitate debugging and caracterisation of edge machine learning systems, we developed a test bench that can be used for fast testing using both simulated and measured data. This paper details this test bench and the methodology used for caracterizing the performance of a real time data compression system based on an FPGA. In our current configuration, the test bench acquires 51.2 Gbps from a 6.4 GSps analog to digital converter and is designed to integrate data pre-processing and machine learning inside an FPGA. All elements are modular and this test bench could be used for analog or digital detectors but also time to digital converters based systems.