Bachelor’s Thesis / Master’s Thesis / Student Research Project

Abstract

Generally speaking, benachmarking of embedded devices is very time-consuming, as they are often not as powerful as a desktop computer. In their paper “Learning-based analytical cross-platform performance prediction” [1] Zheng et al. use different machine learning models (regression models) to predict the runtime of a program on an embedded device by using performance counter values collected on a more powerful host-machine.

Benchmarking an embedded machine learning accelerator can be even more time-consuming. Therefore, the goal of this student project is to use and adapt the methods proposed in [1] to make cross-architecture performance predictions of DNN models running on embedded AI accelerators by using information (e.g. runtime) collected on a more powerful host system with a GPGPU.

References

[1] X. Zheng, P. Ravikumar, L. K. John, und A. Gerstlauer, „Learning-based analytical cross-platform performance prediction“, in 2015 International Conference on Embedded Computer Systems: Architectures, Modeling, and Simulation (SAMOS), Samos, Greece: IEEE, Juli 2015, S. 52–59. doi: 10.1109/SAMOS.2015.7363659.

Requirements

• Python
  • Scikit Learn, PyTorch, Pandas
• Training of classic Machine Learning and Deep Learning Models
• Linux (optional)

Contact

Jung, Alexander

Lübeck, Konstantin

Bringmann, Oliver