The AI ​​accelerator has a processing power of 13 TOPS (13 tera-operations per second) for the inference of its neural network. Thanks to the PCIe communication channel of the Pi 5, the accelerator can efficiently transfer information with the cores of the Raspberry-Pi. This also optimizes the transmission of data from the camera modules.

The AI ​​accelerator is capable of producing results at a rate of 13 frames/second without using the SoC of the Pi 5. The 4 cores of the Pi 5 are still available for your own applications.

Conversely, without an AI accelerator, the inference of a neural network would be supported by the 4 cores of the Pi 5. This would result in 2 to 3 frames/second with all the cores of the SoC monopolized by the inference. There would be very few/no resources left for your own applications.

It seems obvious that most applications will use the AI ​​module with the camera module (or two cameras for a stereoscopic view).

Many pre-trained AI models are available, this allows you to quickly take advantage of the Kit's capabilities. Note that the Hailo 8L module does not have enough power to allow the creation and training of your own models.

A version of Raspberry-Pi OS for Pi 5 will automatically detect the Hailo and the NPU (Neuronal Processing Unit) will be available for AI processing.

The rpicam-apps camera applications integrated into Raspberry Pi OS support the AI ​​module (native support), and will automatically use the NPU to execute post-processing tasks (tasks compatible with the NPU).

The AI ​​Kit offers a particularly efficient scheduler for the hardware accelerator. It allows you to run several neural networks on a single camera -OR- one or more neural networks on two cameras (concurrent processing) see source.

Note: this product requires a Raspberry-Pi 5 and a camera to work.

Need to use GPIO with the AI ​​Kit? Check out our GPIO Replicator which can be soldered onto the stacking header before placing the AI ​​Kit HAT.