Thermal load of a single-fan graphic processing unit (GPU) on a server configuration

Mohd Khir, Muhammad Sufi Ridhwan (2022) Thermal load of a single-fan graphic processing unit (GPU) on a server configuration. Project Report. Universiti Teknikal Malaysia Melaka, Melaka, Malaysia. (Submitted)

	Text (24 Pages) Thermal load of a single-fan graphic processing unit (GPU) on a server configuration.pdf - Submitted Version Download (433kB)
	Text (Full text) Thermal load of a single-fan graphic processing unit (GPU) on a server configuration.pdf - Submitted Version Restricted to Repository staff only Download (4MB)

Abstract

Graphics Processing Unit (GPU) has gone through many revolutionary changes throughout the decade. GPU’s processing power is able to challenge the existing Central Processing Unit (CPU) in the task of running high-profile software. However, any electronics that flow current into it will generate heat. The thermal problem is the major problem for GPUs. The optimal operating GPU requires minimum fan speed and core temperature with the highest efficiency (hashing power). This research focused on obtaining the optimal response at a particular core clock and memory clock via the optimization tool of Design-Expert software. The responses were recorded while applying a constant amount of power supplied to the GPU. The GPU used is in the Single-Fan Graphic Processing Unit, ASUS NVIDIA® GeForce TURBO-GTX1070-8G. The relationship between the clocking (core and memory) and GPU responses (fan speed, core temperature, hash rate) was observed. By that, Central Composite Design (CCD) generated one equation for each fan speed, core temperature, and hash rate. Next, the optimization process suggests several new clocks settings that give the best performance than the current setting. Next, the optimization process offers several new clocks settings that give the best performance than the current setting. The best core and memory clock was selected for the validation and confirmation process. The results from the validation process prove that the predicted response was precise with less than a 2% deviation.

Actions (login required)

View Item