Jonathan Wang

EECS Department, University of California, Berkeley

Technical Report No. UCB/EECS-2025-124

May 19, 2025

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-124.pdf

Energy-efficient software development on domain-specific processors necessitates a high-level model to quickly and accurately estimate software workload energy consumption. However, previous work in energy characterization has mostly focused on architectural design space exploration for domain-specific processors or ISA-based energy models for CPUs. Currently, no work has been done to create ISA-based energy models targeting DNN accelerators.

In this thesis, we present a methodology for creating instruction-level energy models specific to DNN accelerators, focusing on architectures that use systolic arrays for computation. For energy characterization, we determine the DNN accelerator's energy consumption for both its systolic array and private SRAMs for each accelerator instruction. Based on this characterization, we estimate the energy cost per accelerator instruction, which is used to predict the overall energy consumption of DNN accelerator workloads. By using this methodology, our energy models make predictions that are 90% accurate to the actual energy consumption of real-world benchmarks.

Advisors: Sophia Shao

---

BibTeX citation:

@mastersthesis{Wang:EECS-2025-124,
    Author= {Wang, Jonathan},
    Title= {Improving Energy Efficiency of Machine Learning Software with an Instruction-Level Dynamic Energy Model for DNN Accelerators},
    School= {EECS Department, University of California, Berkeley},
    Year= {2025},
    Month= {May},
    Url= {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-124.html},
    Number= {UCB/EECS-2025-124},
    Abstract= {Energy-efficient software development on domain-specific processors necessitates a high-level model to quickly and accurately estimate software workload energy consumption. However, previous work in energy characterization has mostly focused on architectural design space exploration for domain-specific processors or ISA-based energy models for CPUs. Currently, no work has been done to create ISA-based energy models targeting DNN accelerators. 

In this thesis, we present a methodology for creating instruction-level energy models specific to DNN accelerators, focusing on architectures that use systolic arrays for computation. For energy characterization, we determine the DNN accelerator's energy consumption for both its systolic array and private SRAMs for each accelerator instruction. Based on this characterization, we estimate the energy cost per accelerator instruction, which is used to predict the overall energy consumption of DNN accelerator workloads. By using this methodology, our energy models make predictions that are 90% accurate to the actual energy consumption of real-world benchmarks.},
}

EndNote citation:

%0 Thesis
%A Wang, Jonathan 
%T Improving Energy Efficiency of Machine Learning Software with an Instruction-Level Dynamic Energy Model for DNN Accelerators
%I EECS Department, University of California, Berkeley
%D 2025
%8 May 19
%@ UCB/EECS-2025-124
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-124.html
%F Wang:EECS-2025-124