Inference Energy Analysis in Context of Hardware-Aware NAS

Hardware-aware neural architecture search (HW-aware NAS) methods are crucial for
designing and optimizing deep neural networks (DNNs) for efficient deployment on
hardware accelerators.
In this work, we analyze two HW-aware NAS methods, EvoApproxNAS and ApproxDARTS,
and investigate the impact of precise hardware parameters (such as energy)
measurement using Timeloop on their performance.
In particular, we compare this precise measurement approach with the original
approach employed by EvoApproxNAS and ApproxDARTS, which relied on a simple
analytical energy estimation based on the number of multiplications performed
during the inference phase of the convolutional neural network (CNN). Our
analysis demonstrates how the improved energy measurements enhance the search
process of HW-aware NAS methods, resulting in more energy-efficient
architectures.
Furthermore, we highlight the importance of precise hardware parameters
measurement, showing that accurate hardware modeling is critical for obtaining
CNNs with good accuracy-energy trade-offs. Our results show, that without precise
hardware parameter measurement, the HW-aware NAS can produce acceptable results
but may fail to fully exploit the potential of hardware accelerator, especially
if the 8xN-bit approximate multipliers are considered, ultimately limiting the
efficiency of designed architectures.