Axera AX650N: An Edge AI Chip for Local Robotics, Pitched as a Jetson Alternative

Axera has released the AX650N, an Edge AI chip: a specialized processor that runs neural networks locally on robots and IoT devices. This is the first detailed technical teardown of the chip's architecture, with real-world tests of YOLO, multimodal LLMs and other popular computer vision models.

What is the Axera AX650N

The AX650N is a SoC (System-on-Chip) with an integrated NPU (Neural Processing Unit), meaning the chip is purpose-built for neural network workloads. It differs from the general-purpose NVIDIA Jetson in that it targets the edge specifically: running models locally on the device, without sending data to the cloud.

In the tests, the chip ran on Sipeed's Maix4 Hat board: a SoM (System-on-Module) with 8 GB of RAM and a baseboard, connected to a Raspberry Pi 5 over PCIe 2.0. In this configuration the AX650N acts as an external ML accelerator, much like the popular Hailo, but it is built into a single monolithic SoC instead of a separate device.

Chip architecture

The AX650N contains several specialized blocks:

• CPU for system control and conventional computing
• NPU for accelerating neural networks (the main block, optimized for CNNs, transformers and VLMs)
• DSP for audio and real-time signal processing
• ISP for working with cameras and preprocessing images
• Two separate DDR4 memory controllers for parallel access

Memory is critical: 8 GB lets the system keep the model, image batches and neuron activations in RAM, avoiding slow storage access.

The chip was tested on real models: YOLO v8, Depth Anything, SuperPoint and multimodal Qwen3. All tasks ran locally, with no data sent to servers.

Why Edge AI

Robotics, drones and IoT devices need local compute:

• Response speed — instant response, no cloud latency
• Privacy — video, audio and sensor data stay on the device
• Autonomy — works without internet (critical for drones and field robots)
• Economics — no need to pay for cloud API requests

The market has three classes of solutions. The first is the expensive NVIDIA Jetson with CUDA ($300+). The second is external accelerators such as Hailo. The third is NPUs built into SoCs (often Chinese: Axera, MediaTek, Snapdragon). The AX650N belongs to the third class. Jetson is more versatile, but it costs more and needs more power and space. The AX650N is specialized for neural networks, more compact, cheaper and more power-efficient.

What this means

Edge AI is moving out of the niche of high-budget industrial projects and into the mass market. Previously the choice was stark: an expensive Jetson or the cloud with its latency. Now affordable chips like the AX650N are appearing that make it possible to run powerful neural networks locally.

This opens opportunities for startups in agricultural robotics, agri-drones, industrial automation and security. Developers can experiment with AI without heavy infrastructure spending.

The second part of the teardown promises detailed benchmarks and a comparison with Jetson on performance and power consumption.