Mythic has ported its low-power analogue machine learning chip to the Sentinel UAV reference design from Modal AI (writes Nick Flaherty).

Sentinel is a ready-to-fly craft with 2.4/5 GHz wi-fi or 1.8/2.3 or 2.4Ghz point-to-point wireless links and an Analog Matrix Processor (AMP) machine learning chip from Mythic. The chip has been added to the UAV on an M.2 format board measuring 22 x 80 mm for image processing.

The Sentinel uses the latest PX4 autopilot running on Qualcomm’s QRB5165 chipset, which also runs ROS 1/2 or Ubuntu18.04 Linux. There is also a Docker build environment for the CPU, graphics GPU and digital signal processor on the chipset to run machine learning models.

The first AMP chip is the MNP1076, which has 76 tiles that handle the machine learning frameworks in a more power-efficient way than a digital GPU. This gives a typical power consumption of 3 W, making it suitable for UAV applications.

Because the chip handles machine learning frameworks in a different way, the key is the software compiler. Deep neural network models developed in standard frameworks such as Pytorch, Caffe and TensorFlow are optimised for the chip, reduced in size and then retrained for the Mythic Analog Compute Engine before being processed through Mythic’s compiler.

The resulting binary code and model weights, typically 80 m parameters, are then programmed into the AMP for inference. All the weights are stored on the chip, which avoids having to use off-chip memory, saving a lot of power.

Pre-qualified models such as the YOLOv3image recognition framework are also available for developers.