Reproduce “Trends in Energy Estimates for Computing in AI/Machine Learning Accelerators, Supercomputers, and Compute-Intensive Applications” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Sadasivan Shankar, Albert Reuther
This paper quantifies computational energy requirements across AI/ML accelerators, supercomputers, and compute-intensive applications from 2010-2022, revealing that hardware efficiency gains are significantly outpaced by the escalating demands of large AI/ML models and system-level overheads. The analysis shows that application-level energy consumption can negate hardware-level efficiency improvements, challenging current scaling trends.
Key insights
- Traditional energy efficiency gains from geometrical scaling are demonstrably slowing, with increased transistor counts plateauing in power and clock frequency, suggesting gains are more from active energy management than inherent scaling.
- A significant discrepancy exists between fundamental bit-level energy efficiency and higher-level instruction/system efficiency, with EPI being orders of magnitude higher than EPB due to architectural overheads and increasing complexity per instruction.
- The rapidly escalating computational demands of AI/ML applications, particularly in training large NLP models, can completely offset hardware-level energy efficiency improvements, leading to unsustainable overall energy consumption.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.