Accurate and Stable CPU Power Modeling
Accurately estimating CPU power consumption is a key requirement for controlling CPUs – e.g. for implementing energy saving techniques – and for exploring the CPU design space. This project has developed PowMon, which uses models built and validated from real, measured data from an actual device. This means that the tool accuracy is known and therefore the power figures can be trusted.
Obtaining accurate data from mobile devices can be challenging and more time-consuming than using a simulator or desktop/server devices. For this reason, PRiME is making available its experimental platform software tools which allow workloads to be automatically run on a mobile device. Metrics such as Performance Monitoring Counters (PMCs), temperature, CPU utilisation, CPU power and CPU voltage are collected and analysed to produce accurate power models.
The PowMon software has two main features:
- It allows users to automate the model building methodology for their hardware of choice, producing accurate and stable models
- It provides power models for specific CPUs, e.g. quad core Cortex-A7 and quad core Cortex-A15, which can be used as reference models for this hardware
Further information and the PowMon tool software download is available at http://www.powmon.ecs.soton.ac.uk/powermodeling/
The project led to a number of co-authored publications between researchers at Arm and ECS, including a HiPEAC workshop paper (on the initial findings), a PATMOS conference paper (on the thermally compensated model), and an ISPASS paper on the gemStone tools. The IEEE TCAD paper, 2017, was nominated for an IEEE TCAD best paper award, has had >2000 downloads, and was still the 33rd most accessed paper in TCAD 2 years after publication. Tutorials were run with Arm at MICRO 2015 and ISPASS 2016, training >40 academics/engineers, and a number of industrial internships totalling >1yr support knowledge transfer.