Ambarella’s N1-665 system-on-chip (Fig. 1) is designed to handle large language models (LLMs) and vision language models (VLMs) on the edge. Allen Liu, Staff System Software Engineer, provides a demonstration of its capabilities (watch the video above).

The N1-665 is built around eight Arm Cortex-A78AE cores but adds a neural vector processor (NVP), essentially a specialized neural processing unit (NPU) designed to handle LLM and VLM chores. It works with models like LLaVa-OneVision, Llama, and Gemma locally while consuming under 20 W.

These features suit the N1-665 for autonomous mobile robots (AMRs) as well as smart-city security video applications. It’s also useful for industrial and automation applications.

Developers can take advantage of the Cooper Pro (Fig. 2) that’s based on the N1-665. The platform includes dual Gigabit Ethernet and 10-Gb Ethernet ports, Wi-Fi, and Bluetooth support. It has quad, gigabit multimedia serial links (GMSLs) to handle video streams. The on-chip decode supports up to 12 simultaneous 1080p30 video streams.

The system also has 64-GB DRAM, 512 GB of flash memory, and an HDMI 2.0 display output. This is very similar to NVIDIA’s Jetson platforms.

The Cooper Pro and N1-665 work with the Cooper Developer Platform (Fig. 3). The latest Cooper Foundry 2.1 supports Cooper Pro.

The Cooper Foundry 2.1 software includes:

• Cooper Core: A Linux-based OS, compiler and SDK
• Cooper Foundation: For building and deploying machine-learning applications
• Cooper Vision: For building multi-modal sensor processing and sensor fusion
• Cooper UX: Provides analytics and development