Fraunhofer Institute for Integrated Circuits IIS, Division Engineering of Adaptive Systems EASNewsletter 04/2024
The search for suitable hardware for an AI application is often an expensive and time-consuming process. Now, a team from the Fraunhofer Institute for Integrated Circuits IIS in Dresden has developed a method that shows which AI runs more efficiently on which hardware. For example, this allows SMEs, industrial companies, building services companies, or building operators with wear-critical plant technology to minimize investment risks and downtime, guarantee process stability and operational reliability, and simultaneously reduce their energy consumption and costs.
Many companies choose not to use the cloud to analyze the data for their AI applications, as it may not meet stringent data protection requirements. Moreover, data processing in the cloud is slower than in local applications.
The high sampling rates in sensor data capture call for “edge devices,” particularly when it comes to the recording and analysis of vibrations in machines. Edge devices are hardware in which the AI is incorporated directly into the terminal device. The question, however, is which hardware provides the appropriate performance for a specific AI application at the lowest cost.
Systematic performance evaluation of edge devices
A team from the Engineering of Adaptive Systems division of Fraunhofer IIS in on hand to help with decision-making. Comprehensive comparative testing of current AI hardware allows the AI experts to quickly, comprehensively, and independently determine which electronics are best suited to a customer’s intended AI application. Alternatively, the experts are able to determine how AI models can be tailored to the characteristics of the hardware architecture in order to reduce energy and resource consumption.
Moreover, customers’ pilot applications of AI can be implemented and evaluated as prototypes at the AI Application and Test Center (ATKI) of Fraunhofer IIS. “We have access to multiple edge devices in different performance categories,” says André Schneider from the Computational Analytics working group. “For example, our laboratory has already evaluated an IoT device for condition monitoring of an ultrapure water supply used in semiconductor manufacturing.” To this end, the researchers reconstructed the real-world situation and tested the hardware, which was assessed as being economically viable for widespread utilization at the user’s facilities. Lastly, they were able to quantitatively assess the suitability of the chosen sensor–edge device combination for use with AI-based condition monitoring and the extent to which this combination meets the stipulated requirements.
Relocating the validation steps to the suitably equipped laboratory paved the way for shortening the expensive and time-consuming tests in the real application environment. Above all, however, the extensive laboratory analyses mean that it is possible to systematically recognize patterns in the recorded data and then identify effects in system and machine behavior that become relevant in the event of faults and wear.
Sound decisions on AI applications
Soon, the aim is to automatically analyze parameters for customers’ prototype applications of AI. After measurement, all relevant performance parameters – such as inference time, latencies, and resource and energy requirements – will be provided to customers and evaluated, allowing companies to make well-founded decisions regarding their specific application of AI.