SSD Thermal Throttling Profile Prediction Using Neural Network

Abstract

In the past decades, the growing capacity and performance of Solid-State Drives (SSDs) have resulted in higher power consumption and increasing thermal challenges. Thermal throttling has been widely adopted in SSD products to keep critical component temperatures within limits. While thermal throttling profiles are essential for SSD performance in firmware developments, using the traditional Computational Fluid Dynamics (CFD) approach to simulate thermal throttling profiles can be time-consuming. In this paper, a fast-prediction method using a neural network approach is proposed for predicting the thermal throttling profiles under different workloads. The Long Short-Term Memory (LSTM) neural network framework has been adopted in this method. The neural network will learn from the testing data of a single workload to model the thermal performance and hence predict thermal throttling profiles for various workloads. Results have shown that the predicted thermal throttling profiles for various workloads align with experimental test data with good accuracy. Prediction time is significantly reduced to a few minutes. This work has demonstrated that machine learning can be effectively applied to SSD thermal test data to model thermal performances with different test parameters, including power, ambient temperature, and throttling temperature limits.