The research, conducted at the University of Stuttgart (IER), addresses the critical challenge of balancing compute performance with energy efficiency and the quality of waste heat in modern data centers. As hardware power requirements for CPUs and GPUs are projected to rise significantly through 2030 , this project provides a methodology for creating "Operating Maps". These maps visualize how varying CPU frequencies and loads impact system efficiency and thermal characteristics. Key Features of the Methodology Closed-Loop Automation: Utilizes a Python-based stack for centralized monitoring and control of all parameters, including server operation and cooling, by integrating device-specific programs. Full Autonomy & Remote Control: The test bench is designed for fully automated adjustment of operating parameters and offers remote control capabilities, ensuring the system can operate independently without manual intervention. Advanced Thermography: Integration of a FLIR E96 IR Camera via the FLIR Atlas SDK (C#) allows for remote-controlled image capture, parameter setting, and monitoring of heat distribution through an IR-transparent window. Multi-Sensor Data Integration: Combines internal server and CPU sensors with external data loggers (via C++) for the centralized, aggregated recording of all measurement data, providing high-fidelity thermal analysis and high reproducibility. Comprehensive Benchmarking: Employs the SPEC CPU 2017 suite to simulate realistic compute workloads while precisely controlling load via different threads and copies. Integrated Cooling Control: Automated management of cooling parameters, such as fan speed and server air inlet temperature, to map their impact on waste heat quality. The included results demonstrate that peak Compute Efficiency is achieved at a specific "sweet spot" (2GHz at full load, which corresponds to approximately 70 % of the maximum power consumption based on the thermal design power) rather than at maximum clock speeds. The data also shows a "Knee Voltage" threshold at 2.3 GHz, beyond which power consumption and heat generation spike disproportionately to performance gains.
Ott et al. (2026) studied this question.