Energy-Efficient CNN Inferencing on GPUs with Dynamic Frequency Scaling

To ensure that emerging technologies such as autonomous driving and application-specific Internet of Things devices work correctly, fast and accurate calculations must be performed by algorithms like Machine Learning (ML). One essential algorithm in these systems is Convolutional Neural Network (CNN), which requires a lot of computational resources. Designers often use ML accelerators like General-Purpose Graphic Processing Units (GPGPUs) to keep up with design requirements, but choosing the right accelerator and accelerator configuration can be time-consuming and difficult. Our research analyzes the power consumption and execution time of CNNs on GPGPUs with different frequency settings. We found that changing the frequency significantly impacted power consumption but only had a marginal effect on computation time. Furthermore, increasing the frequency beyond 1200 MHz shows no improvement in computation time anymore. Therefore, a lower frequency can help create an energy-efficient CNN inference system without sacrificing performance.