Get started
Work

Evaluating the Impact of Tensor Cores on YOLOv8's Performance as a Pedestrian Detection System

Year: 2024
Type: dissertation
Abstract: The rapid advancement of autonomous vehicles underscores the need for improved pedestrian detection systems to enhance road safety. Traditional GPUs accelerate the detection process but can be prohibitively expensive and slow for widespread deployment. This study investigates the performance of GPUs equipped with Tensor Cores, which can accelerate matrix multiplication operations in pedestrian detection systems but may also introduce precision loss. Focusing on speed, precision, recall, and mean Average Precision, the impact of Tensor Cores on a YOLOv8 model was evaluated using the Caltech pedestrian dataset. Tests were conducted on both high-end NVIDIA A100 and more affordable NVIDIA RTX 3070 GPUs, with and without Tensor Cores activated. The results revealed a noteworthy 11.8% improvement in inference speed on the RTX 3070 with Tensor Cores activated, whereas performance enhancement on the A100 was less pronounced. Notably, enabling Tensor Cores did not lead to significant precision loss on either GPU model, with both exhibiting a decrease in precision of less than one percent, which is negligible. However, an 8.9% decrease in recall was observed on the A100, while the RTX 3070 experienced a 3% decrease. These findings underscore the need for careful optimisation when implementing Tensor Cores to balance speed enhancements with detection accuracy. Furthermore, this research raises ethical considerations regarding the acceptable trade-off between precision loss and speed improvement. It underscores the advantages and challenges of integrating advanced GPU technologies into autonomous vehicles, highlighting the essential need to balance enhanced performance with reliable operation.
Authors: Sasikumar, Abhinav, Syed, Farhan
Language: English
Open Access status: green
OpenAlex ID: W7019130069
Cites: 19
Cited by:
Related to: 10