Accurate Device Temperature Forecasting using Recurrent Neural Network for Smartphone Thermal Management

Abstract

Technological innovations such as 5G, multiplayer online gaming, 8k video recording and incorporation of more camera sensors in smartphones have resulted in heat generation causing poor performance and high power consumption. Given that acceptable range of the device when in contact with human skin is 40~45°C, the generated heat can be uncomfortable for users. In order to have a balance between high performance experience and thermal hazard in smartphones, effectual thermal management strategies are needed. To address this challenge, in this paper we propose Thermal Foresight Engine (TFE) model which predicts the device skin temperature in advance, for prediction windows of 30 seconds, 1 minutes & 3 minutes with accuracy of 97%, 94% and 86% respectively with an average inference time of 17 ms, providing thermal foresight to smartphone performance management & thermal mitigation systems in use cases such as 5G, video recording and games. It uses a Long Short-Term Memory (LSTM) [1] model, which has an RNN[1] based architecture. It takes input as current & past workload conditions in the form of a system-wide contextual information. Applications of this research will assist mitigation strategy used in smartphone proactively. This will effectively handle thermal anomalies in smartphones without heat up by improving thermal awareness, preventing performance impact due to thermal anomalies.