Researchers at the University of Tehran's College of Engineering Develop a Method to Reduce AI Hardware Power Consumption by 80%
According to the Public Relations Office of the College of Engineering, the growing adoption of artificial intelligence has increased demand for processors that are both faster and more energy-efficient. One of the computational operations requiring substantial power and hardware resources in AI workloads is numerical multiplication. In this study, researchers at the University of Tehran sought to perform these computations using a simpler and more energy-efficient approach.
The study introduces a method called ELA-FPM for approximate floating-point multiplication. The central idea is that not every component of a numerical computation needs to be processed with the same level of precision. Components that have a greater impact on the final result are computed more accurately, whereas less significant components are handled using simpler, faster operations or omitted altogether.
To implement this approach, the researchers employed a precomputed lookup table to perform part of the computation. This makes it possible to replace some of the complex, power-intensive circuits traditionally required for full-precision calculations with a simpler hardware structure. As a result, the implementation can achieve a smaller hardware footprint, faster processing, and lower power consumption.
Hardware evaluation results showed that, compared with exact multiplication, ELA-FPM reduced processing time by up to 68.3%, power consumption by 80.3%, and hardware area by 92.1%. The method also demonstrated considerable advantages over the best existing approximate multiplication techniques, achieving reductions of up to 65.6% in processing time, 70% in power consumption, and 71% in hardware area.
To determine whether reducing computational complexity would significantly compromise AI performance, the researchers evaluated the proposed method on several well-known neural network models. The results showed that, in one ELA-FPM configuration, the maximum accuracy degradation across the evaluated models was only 0.05 percentage points. In the same configuration, processing time was reduced by 36%, power consumption by 76%, and hardware area by 88%.
These findings suggest that, in certain AI applications, approximate computation can deliver substantial savings in energy and hardware resources without significantly compromising the final model accuracy.
The proposed method was also evaluated using BERT, a well-known language-processing model. In one ELA-FPM configuration, the model achieved an accuracy of 84.19%, approximately matching its accuracy when exact computations were used.
Overall, the findings indicate that ELA-FPM offers a promising approach to designing AI hardware that consumes less energy, occupies less physical area, and delivers higher processing speeds. These capabilities could be particularly valuable in developing computing hardware for neural network accelerators, AI models, and large language models (LLMs).
The article, entitled “ELA-FPM: An Energy-Efficient LUT-Based Approximate Floating-Point Multiplier for Neural Network Acceleration,” was authored by Mahla Salehi Sheikhali Kelaye, Esma Naseri-Rad, Sahand Divsalar, Omid Salehinia, and Dr. Shaghayegh Vahdat, a faculty member of the School of Electrical and Computer Engineering, University of Tehran. The paper was published in IEEE Micro, a publication of the IEEE Computer Society, in the May–June 2026 issue.
Read the full article:
Your Comment :