A new study published on arXiv investigates the energy consumption of Large Language Models (LLMs) used for code optimization. The research found that while LLMs can improve code performance, the optimized code often consumes more energy than human-written reference solutions. The energy cost of the LLM optimization process itself can be substantial, with inference energy ranging from 0.79 to 8.68 MJ per problem, equivalent to thousands of GPT-4 queries. The study suggests that architectural choices, such as Mixture-of-Experts, have a greater impact on energy efficiency than decoding parameters, and that performance optimization and energy efficiency are not always aligned goals. AI
IMPACT LLM-generated code may be less energy-efficient than human-written code, impacting sustainability efforts.
RANK_REASON Research paper published on arXiv detailing energy costs of LLM code optimization. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
