Recent work has proposed a new strategy for optimizing the energy efficiency of Internet of Things devices: exploiting the tolerance to constrained errors that pertains to many algorithms, from online search to multimedia compression in smartphones. This strategy is at the heart of the promising AxC paradigm, which explores making deliberate mistakes in exchange for disproportionate savings.
As different applications have varying degrees of tolerance, Damsgaard’s doctoral dissertation explores the use of AxC in reconfigurable hardware architectures. Unlike the processor in your smartphone that can only execute a set of operations determined when it is designed, reconfigurable architectures can be altered to support different operations in the field. As a result, these architectures can quickly adapt to various applications and degrees of error tolerance.
Framework available in open source
Damsgaard’s work includes the development of an open-source framework for modeling reconfigurable hardware architectures with integrated AxC features. This framework vastly reduces the effort needed to experiment with new architectures or approximation techniques. Initial results show that simple approximations have limited overhead and may reduce power consumption by up to 20 %.
“Our proposed framework does not merely underline the promises of power savings arising from AxC; it is also a flexible, extensible foundation for further development,” says Damsgaard.
His work and findings have contributed to 13 publications in international conferences and journals. The framework and its related tools are all available on GitHub.
Hans Jakob Damsgaard conducted his research between 2021 and 2024 under the APROPOS project (Approximate Computing for Power and Energy Optimisation), a Marie Sklodowska-Curie European Innovative Training Network, at Tampere University, IS Wireless in Poland, and AMD Research in Germany. APROPOS received funding through the Horizon 2020 Research and Innovation Programme under Grant Agreement No. 956090. He is currently an SoC Verification Engineer with Nokia.
Public defence on Thursday 19 December
The doctoral dissertation of M.Sc. (Eng.) Hans Jakob Damsgaard in the field of electrical engineering titled Reconfigurable Approximating Accelerators for Edge Computing will be publicly examined at the Faculty of Information Technology and Communication Sciences at Tampere University at 12 o’clock on Thursday 19 December 2024 in the TB104 auditorium, Tietotalo, Hervanta campus (address: Korkeakoulunkatu 1, 33720 Tampere).
The Opponent will be Professor Juha Plosila from Turku University. The Custos will be Professor Jari Nurmi from Tampere University. The work has been co-supervised by D.Sc. (Tech.) Aleksandr Ometov from Tampere University.
The doctoral dissertation is available online.
The public defense can be followed via remote connection.