“CV75AX provides up to 2x the AI performance over Ambarella’s prior-generation SoC, enabling the latest transformer neural networks for improved accuracy and reduced false positives without having to train for every object. [It] also provides the AI horsepower for more complex driver-scoring models,” according to the company. “Suited for next-gen vehicle gateways, CV72AX supports up to 10 cameras and offers up to 6x the AI performance over its predecessor, enabling both vision transformers and multi-modal vision language models, which provide instant event alerts, as well as multi-channel video pre-categorisation and natural-language searches for more efficient video analysis – all without the need for training.”

Little further information has been revealed so far, beyond that the ICs support LPDDR5 memory and USB 3.2 interfaces for 5G modems.

On board is an AI accelerator, a CPU, an image signal processor and an H.264/5 video encoder, and they are capable of “multi-channel recording for evidence collection”, said Ambarella.

One snippet is that the CV75AX’s has a <5mW mode for monitoring the vehicle’s sensors when parked.

Fabrication is on a finfet process at a Samsung foundry, and operation is over -40 to 105°C. Both parts are sampling now.

Development support comes from a platform called Cooper, with Cooper Metal providing hardware tools and Cooper Foundry providing a multi-layer software stack.