Keeping GPUs Cool: GPU Temperature Prediction Using LSTM

Thermal regulation in GPU-based systems is increasingly critical as rising operational temperatures during high-performance computing tasks can lead to performance degradation, higher cooling demands, and reduced system reliability. Accurate temperature prediction is key to enhancing thermal management strategies and resource allocation. This paper introduces an approach based on Long Short-Term Memory (LSTM) networks developed to estimate GPU temperatures using key GPU operational features. The model delivers accurate predictions by capturing complex temporal patterns in thermal behaviour, achieving an average RMSE of 1.4974 and a mean prediction error of 1.08°C. The findings underscore the accuracy of the approach in predicting temperature trends, forming a foundation for developing more advanced models that can forecast future GPU temperature in real-time.