Attentions Under the Microscope: A Comparative Study of Resource Utilization for Variants of Self-Attention

TL;DR

This paper benchmarks eight attention mechanisms in training GPT-2, evaluating their resource efficiency and energy consumption, revealing that optimized kernels like Flash Attention are most energy-efficient.

Contribution

It provides a comprehensive empirical comparison of attention variants on resource usage and energy efficiency during GPT-2 training.

Findings

Optimized kernel attention mechanisms are most energy-efficient.

Lower GPU power does not always mean lower energy consumption.

Training time significantly impacts overall energy use.

Abstract

As large language models (LLMs) and visual language models (VLMs) grow in scale and application, attention mechanisms have become a central computational bottleneck due to their high memory and time complexity. While many efficient attention variants have been proposed, there remains a lack of rigorous evaluation on their actual energy usage and hardware resource demands during training. In this work, we benchmark eight attention mechanisms in training GPT-2 architecture, measuring key metrics including training time, GPU memory usage, FLOPS, CPU usage, and power consumption. Our results reveal that attention mechanisms with optimized kernel implementations, including Flash Attention, Locality-Sensitive Hashing (LSH) Attention, and Multi-Head Latent Attention (MLA), achieve the best energy efficiency. We further show that lower GPU power alone does not guarantee reduced energy use, as training time plays an equally important role. Our study highlights the importance of energy-aware benchmarking in attention design and provides a practical insight for selecting resource-efficient mechanisms. All our codes are available at GitHub.