OpenAI’s fourth large language model (LLM), GPT-4, took an estimated 50 Gigawatt-hours to train, or the equivalent of 5,000 American homes‘ yearly power consumption. That was in 2023. Since then, the computational resources used to train frontier LLMs have only increased, though direct power usage numbers are hard to come by.
Now, a research group at University of Twente in the Netherlands has sho...
The finding that fine-grained control over clock frequency can yield significant energy savings challenges the current paradigm where hardware performance is often treated as a fixed constraint. The core innovation lies in optimizing the hardware for the software by targeting the execution level, rather than relying on coarse-grained adjustments over training iterations. This introduces a critical tension between theoretical savings and practical implementation: the measured 14 percent saving is...