Vision Language Models Cut AI Energy Costs by Encoding Time-Series as Images
A recent study published on arXiv (2608.07427) indicates that Vision-Language Models (VLMs) can significantly lower AI energy usage by transforming time-series data into 2D visualizations. This approach results in a token reduction of 3.6-10.4x for architectures such as Llama-3.2-90B, Qwen2.5-VL-72B, and Pixtral-12B, leading to a 1.8-2.5x decrease in inference energy. This equates to a daily energy saving of around 7.2 MJ in telecom edge deployments and CloudRAN monitoring of 200 cells every 15 minutes. Importantly, a fine-tuned Llama-3.2-90B-Vision VLM demonstrates 220.7% greater precision compared to its text-only version and surpasses LSTM and ARIMA benchmarks by over 144% in telecom anomaly detection, addressing the significant inefficiencies in LLM inference that account for over 90% of AI's operational energy.
Key facts
- LLM inference accounts for over 90% of AI operational energy.
- VLMs reduce input tokens by 3.6-10.4x by encoding time-series as 2D plots.
- Energy reduction of 1.8-2.5x in inference is achieved.
- Saves approximately 7.2 MJ/day at telecom edge deployments and CloudRAN.
- Fine-tuned Llama-3.2-90B-Vision VLM achieves 220.7% higher precision than text-only counterpart.
- Outperforms LSTM and ARIMA baselines by over 144% on telecom anomaly detection.
- Tested on Llama-3.2-90B, Qwen2.5-VL-72B, and Pixtral-12B architectures.
- Addresses inefficiency in telecom network analytics and numerical time-series data analysis (NTSDA).
Entities
Institutions
- arXiv
- Llama
- Qwen
- Pixtral