Energy efficiency of LLM-based database querying

Determine whether the significant energy consumption associated with using large language models to answer natural language queries over tabular data can be reduced to practical levels comparable to relational database engines such as SQLite, and identify configurations or techniques that achieve such energy efficiency without sacrificing accuracy.

Background

The paper benchmarks nine open-source LLMs against SQLite on a synthetic tabular dataset to compare accuracy, execution time, memory, and energy usage for database-style querying. The results show that even small or quantized LLMs incur substantial energy overhead compared to a native SQL engine, while also exhibiting inferior accuracy on direct question answering.

While the authors note that larger models might improve accuracy in the future, they explicitly state that the energy issue remains unresolved—highlighting an open problem regarding how to make LLM-based querying energy-efficient enough to be a viable substitute for relational databases.