Energy Efficiency of Locally Deployed LLMs: A Preliminary Quantitative GPU Power Benchmark on Consumer Hardware
06:00 · August 4, 2026 · arXiv cs.AI RSS
The local deployment of large language models (LLMs) is gaining traction due to privacy concerns and the desire for on-premise inference. However, the energy costs on consumer hardware remain poorly characterized, as most benchmarks focus solely on accuracy. This paper presents a reproducible, hardware-level energy benchmark of nine open-source LLMs (1B to 7B parameters) executed on a single consumer GPU (RTX 4060Ti 16GB). Using the Ollama inference engine, GPU power draw was sampled at 2Hz via nvidia-smi across a fixed prompt set. We evaluate mean/peak power, total energy per prompt (J/prompt), energy per output token (J/token), and throughput (tok/s). Our findings suggest that factors beyond raw parameter count, including model architecture and quantization strategy, drive energy efficiency. Specifically, gemma3:1b and llama3.2:1b achieve the lowest energy cost (0.56 J/token and 0.65 J/token) and the highest throughput (>170 tok/s). In contrast, the 7B-Mistral model consumes up to 4.4x more energy per token than the most efficient model. Notably, qwen3.5:2b exhibits anomalously high per-prompt energy due to extended internal reasoning, highlighting the need to distinguish between token generation modes in efficiency metrics.
Summary
A recent arXiv preprint examines the energy consumption of local LLM inference on consumer-grade hardware, an area that has received less attention than training costs or cloud-scale deployments. The authors ran nine open-source models ranging from 1B to 7B parameters on an RTX 4060 Ti 16 GB GPU under the Ollama runtime, sampling power draw at 2 Hz through nvidia-smi while executing a fixed set of 15 prompts that varied in length and reasoning demand. They recorded mean and peak power, energy per prompt, energy per generated token, and throughput, producing a reproducible measurement harness and openly available dataset.
Results show that energy efficiency is not determined solely by parameter count. Model architecture and the quantization scheme supplied by Ollama exert stronger influence. The smallest models, gemma3:1b and llama3.2:1b, delivered the lowest energy figures at 0.56 J/token and 0.65 J/token respectively, while sustaining throughput above 170 tokens per second. In contrast, the 7B Mistral variant required up to 4.4 times more energy per token than the most efficient model under identical conditions.
One notable outlier was qwen3.5:2b, whose per-prompt energy consumption rose sharply because of extended internal reasoning traces that increased output length without a corresponding gain in external token count. The study therefore distinguishes between token-generation modes when interpreting efficiency metrics and underscores the practical value of architecture-aware model selection for sustained local use.
Why it matters
Directly actionable for Dutch researchers and SMEs deploying local LLMs, aligning with EU sustainability goals, privacy-driven on-premise inference, and Green AI priorities in the Netherlands.
