Ambarella N1-655 Edge GenAI SoC

Models & Pricing

Specifications

Overview

The Ambarella N1-655 is the second member of the company’s N1 Edge GenAI family and is aimed at a specific and increasingly common problem: running multimodal large models on-premise, on many camera channels at once, without a datacenter budget. It combines eight Arm Cortex-A78AE application cores on Samsung’s 5nm process with an Ambarella Neural Vector Processor tuned for AI performance per watt.

The headline capability is multi-modal inference — processing vision-language models that reason jointly over images, video and text, at a scale of roughly eight billion parameters, while simultaneously decoding up to twelve 1080p video streams, all inside a power envelope under 20 watts. That combination is what separates it from both classical vision SoCs and from edge GPU boards: the video pipeline, the NPU and the CPU cores are dimensioned to work at the same time rather than forcing the designer to choose which workload gets the silicon.

Ambarella explicitly targets autonomous mobile robots and on-premise AI box applications with this part, which tracks with where physical AI demand is growing. A robot or an on-premise analytics appliance needs local reasoning — LLM or VLM inference — plus enough concurrent stream handling to feed the model. The N1-655 is designed to do both without cloud round-trips, which addresses latency, bandwidth cost and data-governance objections simultaneously.

Customers needing more headroom can step up to the flagship Ambarella N1, which doubles the CPU complex to sixteen Cortex-A78AE cores, adds a GPU, and brings a General Vector Processor for offloading classical computer vision, radar processing and float-intensive algorithms, plus a dense stereo and optical flow engine. For narrower vision tasks, the CV72S and CV75S vision SoCs in the same 5nm family cover 4KP60+ and 4KP30+ imaging respectively with the same CVflow CNN toolkit and multi-stream encoding. The N1-655 sits in the middle: more AI capability than a vision SoC, far less power than an edge GPU.

Target Applications

Autonomous mobile robots and AMRs, on-premise AI box appliances, multi-camera AI processing hubs, smart city and traffic video analytics, industrial automation and inspection, intelligent healthcare imaging, smart retail analytics, and on-premise multi-channel vision-language model deployments.

Related Products