Martin Stiborski has been with Bressner Technology since 2004. During his eight years as a sales manager, he acquired a broad range of industry expertise. He has been Managing Director since 2015 and has also been part of the executive team at parent company One Stop Systems in the USA since 2018.

The Industrial Internet of Things is generating more and more data that needs to be managed. This data complexity needs to be mastered. What role does edge computing play in this?

Stiborski: Edge computing is the ideal form of data processing when it comes to calculating large volumes of data in real time with the lowest possible latency. Data from sensors, routers or gateways is not sent to central data centers, but processed in workstations close to the plant.

Edge computing relieves industrial IoT (IIoT) networks immensely and is often used for big data and M2M communication in warehouses and production halls as well as for artificial intelligence. However, the technology also enables interesting applications such as autonomous drones, advanced driver assistance systems (ADAS) and intelligent robotic systems.

At the edge, sensor data is pre-selected, software applications are processed or even AI calculations are carried out. What should users bear in mind with regard to the edge software landscape? Which technologies should/must be taken into account?

Stiborski: The framework that is used for edge computing hardware should correspond to the application. Depending on whether the hardware uses CPU, TPU, GPU, FPGA or MXM modules, there are complete software packages that complete the platform.

The demands on the hardware grow with the tasks. What should users look out for when selecting the right edge computing hardware?

Stiborski : There is no 'universal' solution when selecting edge computing hardware: FPGAs, TPUs, GPUs or CPUs offer different advantages and disadvantages. Users must therefore ensure that they create a suitable architecture for their applications.

The advantage of CPUs is that they are very easy to program using C/C++, Scala, Java, Python or other languages and can be integrated into any framework. In edge computing, they are more suitable for simpler AI models with short training phases. Compared to CPUs, GPUs are based on much simpler processor structures, but can process data in parallel due to the high number of cores. In combination, GPUs become a powerful tool for facial recognition or similar applications. TPUs were developed to perform extremely fast applications such as the calculation of vectors and matrices. However, this architecture offers less flexibility than CPUs and GPUs and should only be used on TensorFlow-based models. The decisive factor with FPGAs is that the hardware for solving a problem can be customized to the processor structure, which is not possible with CPUs. This enables high performance to be achieved at low cost and low power consumption. FPGAs are used in particular for special HPC algorithms and can be programmed with OpenCL and HLS.