Optimization of GPU thermal management via response surface methodology (RSM) analysis

Lim, Chin Hui (2024) Optimization of GPU thermal management via response surface methodology (RSM) analysis. Project Report. Universiti Teknikal Malaysia Melaka, Melaka, Malaysia. (Submitted)

Text (Full Text) Optimization of GPU thermal management via response surface methodology (RSM) analysis.pdf - Submitted Version Download (6MB)

Abstract

The recent rapid progress in information technology (IT) has highlighted the negative impact of excessive temperatures on computer systems, particularly the GPU. This can significantly impair the device's performance and reduce its lifespan. Presently, GPU cooling systems utilise two distinct cooling methods: active and passive cooling. Therefore, it is imperative to have a robust cooling system in order to effectively dissipate heat. Hence, it is imperative to do thorough research and innovate novel cooling systems that are more efficient. This study aims to address the knowledge gaps regarding GPU thermal regulation by specifically examining the influence of fan speed on GPU temperatures. The investigation will be conducted under identical GPU load situations. The GPU stress indicated pertains to the Dagger-Hashimoto mining algorithm. Additionally, there are three distinct thermal pad materials: silicone, nickel, and copper. These materials were examined to analyse the thermal regulation of the GPU while also considering the impact of fan speed. The employed model consists of a sandwich structure consisting of a microchip, a thermal pad, and a heat sink. The study utilised the response surface methodology (RSM) with a single factor to forecast the ideal fan speeds necessary to achieve certain temperatures. To ensure the accuracy of the model, we conducted additional experiments to assess how the suggested GPU temperature affects the fan speed, as predicted by the RSM model. For the given scenario, the most accurate prediction for the GPU temperature of the stock thermal pad, specifically the Arctic TP-3, is achieved using a quartic model. Conversely, the most accurate prediction for the nickel thermal pad can be achieved by employing a quartic model for the GPU core.

Actions (login required)

View Item