GRC AI Add-on for CrowPanel on HX6538

Run AI models on your CrowPanel — without cloud, complexity, or extra boards.

Build interactive, intelligent devices that generate stories, classify inputs, or run TinyLM models — all offline and on-device.

Just plug it in.

GRC AI Add-on is a plug-in NPU module designed for CrowPanel, powered by the ultra-efficient Himax HX6538 microcontroller with Arm Cortex-M55 and Ethos-U55 NPU.
It enables local neural network inference — perfect for applications like text generation, classification, and voice interfaces.

Who It's For

Smart Device Developers
→ Rapid prototyping and hypothesis testing with TinyML

Schools, Makerspaces, Tech Hubs
→ Practical demonstration of TinyLM/TinyML applications

DIY Makers & AI Enthusiasts
→ Build offline AI projects with Tiny Language Models (TinyStories, command assistants, etc.)

Key Features

Demo : TinyStories

An interactive storytelling prototype based on the TinyStories Language Model. It showcases how compact hardware can be used for interactive natural language generation — all at the edge, with low power consumption and no internet required for model inference.

1. The user types or choose a prompt on the CrowPanel touchscreen — for example, “Tell me a story about a princess.”
2. The request is processed locally by a Tiny Language Model running on the GRC AI Add-on.
3. Within seconds, a complete story is generated and displayed on the screen — with no internet connection or cloud services involved.

To run the TinyStories demo, a CrowPanel Advance is required for input and display. Download the CrowPanel firmware here. The same link also provides the latest firmware for the Add-on and full documentation.

Specs

AI Core

• Microcontroller: Himax HX6538-A06TDFG
• CPU: Arm Cortex-M55
• NPU: Arm Ethos-U55
• SRAM: 2 MB
• ROM: 4 MB
• External Flash: 128 Mbit (16 MB) QSPI NOR Flash

Form Factor & Pinout

• Add-on layout: 14-pin dual-row (7×2) header
• Pitch: 2.54 mm
• Board size: ~20×25 mm

Power & Performance

• Power Supply: 3.3 V (regulated by CrowPanel)
• Typical Consumption:
• Idle: <1 mA
• Active inference: ~10–20 mA depending on model
• Designed for: battery- and solar-powered deployments

Storage & Model Deployment

• Flash for models: 16 MB QSPI Flash
• Preloaded example: TinyStories (text-generation LLM)
• Model update: via firmware flashing or QSPI update tools
• Frameworks supported: CMSIS-NN, GRC SDK, TensorFlow Lite for Micro (adapted)

Links

https://github.com/Grovety/grc_ai_add-on