FuzzPower: Leveraging the Power of Fuzzing to Uncover the Extremes of Processor Power

Loading...
Thumbnail Image

TR Number

Date

2026-05-07

Journal Title

Journal ISSN

Volume Title

Publisher

Virginia Tech

Abstract

Power is a first-class design consideration for all modern hardware, impacting everything from performance to thermal management and system longevity. Accurate identification of peak power consumption scenarios is crucial for power delivery network design and thermal validation. However, as hardware designs scale to billions of transistors at modern single-nanometer process nodes, manually identifying, constrained-random simulation, or exhaustive exploration of the operational conditions that lead to these power peaks becomes intractable. Human engineers have no hope of divining extreme test sequences, and conventional verification techniques struggle to provide sufficient coverage to reliably trigger worst-case power events. Simulations potentially require days per scenario, hindering rapid design iteration and comprehensive analysis. Furthermore, the inability to uncover true extremes forces designers to employ unnecessarily large guard bands, leaving performance on the table and systems vulnerable to power viruses. These limitations create a critical gap in pre-silicon power validation, especially for complex designs. This paper introduces an open-source, fast, and scalable approach to identify extreme (both high and low) power consumption scenarios in hardware designs by repurposing techniques from hardware fuzzing to create a power-aware hardware fuzzer. Instead of performing a detailed power simulation for every test case, we observe that the size of the execution trace files generated during fuzzing correlates with circuit-switching activity and, consequently, with dynamic power. By analyzing trace file size variations across a large set of stimuli generated using power as a feedback mechanism for mutational test case generation, our method efficiently uncovers test sequences that induce power extremes. We build a power-aware hardware fuzzer that uses fuzzing and execution trace file size to evaluate multiple peripherals and processor cores. Our implementation identifies test cases that, on average, increase maximum dynamic power by 44%, decrease minimum dynamic power by 14%, and increase overall range by 13.6x compared with traditional dynamic power estimation methods.

Description

Keywords

fuzzing, simulation, power estimation

Citation

Collections