A Look into the Performance of the Raspberry Pi AI HAT+ for Computer Vision Applications

The Raspberry Pi series of computers provides a great starting point for embedded Artificial Intelligence (eAI) applications, especially for students and beginners. However, even the latest release from Raspberry Pi Ltd., Raspberry Pi 5, severely underperforms compared to other options available in the market. Raspberry Pi Ltd. released the Raspberry Pi AI HAT+ add-on accessory capable of 26 Tera Operations per Second (TOPS) inferencing performance to address this lack of performance. This article highlights the performance improvements observed in the Raspberry Pi 5 while running a pre-trained You Only Look Once (YOLOv8) object detection model. We compare inference and other related parameters with and without the AI HAT+. We also explore the difference the AI HAT+ makes in CPU utilization, memory utilization, thermal performance, and power consumption. Ultimately, we observed that a Raspberry Pi 5 with the AI HAT+ offered improvements in inferencing and related parameters of greater than 80% compared to running the model natively on the base Raspberry Pi 5 while utilizing 2 W less power, generating less heat on the CPU, and also reducing CPU core utilization by 24.6% for core 0, and at least 60% for cores 1, 2, and 3. The improved performance, simplicity of using the Raspberry Pi 5, affordable price, and excellent community support means that the Raspberry Pi will continue to be the eAI computer of choice even when compared to newer competitors like the Nvidia Jetson boards, particularly for applications that rely on inferencing.