At Google I/O 2026, Synaptics and Google Research introduced the Coralboard, a development platform that lets AI models run directly on edge devices without sending data to the cloud. The board combines Synaptics' Astra processor with Google Research's Coral neural processing unit (NPU), a specialized chip designed to accelerate artificial intelligence workloads. This announcement signals a shift in how developers can build AI applications that keep personal data private and respond instantly without relying on internet connectivity.

What Makes the Coralboard Different From Other AI Development Boards?

The Coralboard isn't just another developer kit. It pairs two complementary pieces of hardware: the Synaptics Astra SL2619 dual-core processor running at 2 GHz with 2 GB of memory, and a 1-TOPS NPU subsystem. To put that in perspective, 1 TOPS means the chip can perform one trillion operations per second, enough to run smaller, optimized AI models at low power consumption.

What sets this platform apart is its focus on practical, real-world inference. During the Google I/O showcase, engineers demonstrated "Jellectronica," a live installation that analyzed video footage from the Monterey Bay Aquarium's jellyfish tank and converted the creatures' movements into generative music using Google DeepMind's Lyria Realtime model. This wasn't a theoretical demo; it showed on-device AI handling vision processing and music generation simultaneously, all without cloud connectivity.

How Does the Software Stack Support On-Device AI?

The Coralboard runs on a unified toolchain called Synaptics Torq, which is built on MLIR (Multi-Level Intermediate Representation), a compiler framework that helps translate AI models into efficient code for edge hardware. The platform supports Gemma 3 270M, a lightweight transformer model with 270 million parameters, small enough to fit on the device while remaining capable enough for meaningful tasks.

For developers, this matters because it eliminates a major friction point: they no longer need to stand up cloud infrastructure or manage API calls to run AI inference. Personal data stays on the device, encrypted at the silicon level, which addresses both privacy and security concerns that plague cloud-dependent AI systems.

Steps to Getting Started With Edge AI Development

• Evaluate Your Model Size: Determine whether your AI model can be quantized and compressed to fit within 1 TOPS of compute. Smaller transformer models like Gemma 3 270M are ideal candidates for edge deployment.
• Test the Toolchain: Use the Synaptics Torq MLIR-based toolchain to compile your model and assess operator coverage and quantization fidelity before committing to production.
• Prototype With Reference Hardware: The Coralboard provides a reference design with dual NPU support, allowing teams to validate multimodal workloads (vision, audio, text) before designing custom silicon.
• Plan for Expansion: The board supports optional Wi-Fi and Bluetooth modules, MIPI CSI camera inputs, and audio/display outputs, enabling you to build complete edge AI systems without redesigning from scratch.

What Should Developers Watch For?

The Coralboard was released as a limited-edition board at Google I/O, and broader developer availability timelines have not yet been announced. Practitioners should monitor several factors as the platform matures.

First, the toolchain's ability to handle complex operators beyond the showcase workloads remains an open question. While Jellectronica demonstrated vision-to-music conversion, real-world applications often require custom layers or novel operator combinations that may not be fully supported yet.

Second, independent benchmarks on standard models like YOLOv8 (a popular object detection framework) will help developers understand performance trade-offs between accuracy and speed on the 1-TOPS hardware. Third, the ecosystem's maturity depends on third-party model support and whether the MLIR toolchain can efficiently handle quantization without significant accuracy loss.

"Watching the Synaptics Coralboard come to life at Google I/O is a proud moment for our team," said Robert Otreba, CEO at Grinn, the IoT hardware company that designed the board.

Robert Otreba, CEO at Grinn

The Coralboard represents a meaningful step toward democratizing edge AI development. By pairing Synaptics' processor with Google Research's Coral NPU and providing an open toolchain, the platform lowers the barrier to entry for teams building privacy-first, low-latency AI applications. The real test will come as developers push beyond the showcase demos and attempt to deploy production models on constrained hardware.

For enterprises and makers concerned about data privacy, latency, and cloud costs, this kind of reference design could accelerate adoption of on-device inference. The next milestone will be general availability and real-world deployment metrics from early adopters.