A research paper by Software Engineering Research Center (SERC) undergraduate students Hemang Jain, Shailender Goyal, and Divyansh Pandey, co-authored with Dr. Karthik Vaidhyanathan, titled “Dissecting Transformers: A CLEAR Perspective Towards Green AI,” has been accepted as a main conference paper at Conference on Empirical Methods in Natural Language Processing (EMNLP 2026).

The work introduces CLEAR (Component-Level Energy Assessment via Repetitions), a methodology for conducting fine-grained energy analysis of Transformer architectures. While existing Green AI studies often examine energy consumption at the whole-model level, CLEAR focuses on individual Transformer components, including Attention, MLP, normalization, and the LM Head.

The researchers evaluated 15 models across four Transformer architecture types, with CLEAR capturing more than 90% of the total model energy. The study shows that Attention consumes significantly more energy per FLOP compared with the overall model, highlighting that FLOPs alone are not sufficient to characterise the true energy cost of different components.

The work also provides insights into how energy consumption varies with factors such as batch size, attention heads, hidden dimension, KV cache, and different attention implementations. These fine-grained measurements can help researchers and ML practitioners make more informed decisions when designing energy-efficient AI systems.

The research was carried out by SERC undergraduate research students Hemang Jain, Shailender Goyal, and Divyansh Pandey, under the guidance of Dr. Karthik Vaidhyanathan. The work represents more than a year of research and contributes to ongoing efforts towards Green AI and Sustainable AI.

The research was supported by the Anusandhan National Research Foundation (ANRF) PMECRG grant under the SustAInd project.

August 2026