How to Start with Google Coral NPU for Edge AI

Wearables and sensors need AI that runs all day without draining the battery or sending every frame to the cloud. Google Research’s Coral NPU open architecture addresses this by providing a royalty-free RISC-V design and compiler that silicon partners can ship in commercial chips.

On October 20, 2026, Google completed the 32-bit Coral NPU stack, including the Matrix Execution Unit that targets up to 1 TOPS at roughly 10 mW average inference power. Synaptics simultaneously launched the Astra SR200 series of microcontrollers that integrate Coral NPU technology and made the Coralboard SL developer kit generally available.

This guide shows developers how to access the tools, evaluate the hardware, and prepare models for Coral NPU targets.

What Coral NPU delivers

Coral NPU is an open-source neural processing unit based on the RISC-V instruction set. It includes a scalar core for control flow, an RVV vector unit for SIMD work, and a dedicated Matrix Execution Unit for efficient GEMM and GEMV operations. The design prioritizes energy efficiency so that continuous sensing and compact transformer models can run on small batteries.

The matching Coral Compiler v1.0, built on IREE and MLIR, accepts models from JAX, PyTorch, and TensorFlow. It partitions work between host CPU and NPU, fuses operators, and lowers them to Coral-specific instructions. An open MPACT simulator lets teams profile performance before any silicon is fabricated.

Synaptics Astra SR200 chips deliver 200–400 GOPS in an MCU package aimed at always-on voice, sensing, and vision. The Coralboard SL (1 TOPS) is orderable now through DigiKey for Linux-based prototyping, while the Coralboard SR targets MCU-class applications.

Step-by-step: getting started

1. Review the official documentation at developers.google.com/coral. The site hosts architecture guides, the datasheet, IP integration notes, and compiler tutorials.

2. Clone the open repositories. The complete 32-bit RTL lives at github.com/google-coral/coralnpu. The compiler stack is at github.com/google-coral/coralnpu-compiler and the simulator at github.com/google-coral/coralnpu-mpact.

3. Order a development board. The Coralboard SL is listed on DigiKey. Follow the Synaptics Coralboard developer page for SDKs and example applications that already run models such as Moonshine speech-to-text and Gemma variants on-device.

4. Export your model to StableHLO or TOSA. The compiler accepts these intermediate representations and produces Coral NPU binaries. Start with smaller quantized models (int8) for the lowest power envelope.

5. Simulate first. Use the MPACT simulator to measure latency and energy estimates before flashing hardware. This avoids repeated board iterations for unsupported operators.

6. Deploy and measure. Once the Coralboard or a Synaptics SR200 sample is in hand, run the provided examples, then replace the model weights with your own. Profile actual power draw under continuous inference.

For teams already building Android apps that call cloud or on-device Gemini models, Coral NPU offers a complementary path for the always-on sensor tier. See our earlier guide on adding Gemini 4 Argon to Android apps with Firebase AI Logic.

Why the open standard matters

Proprietary edge accelerators force teams to rewrite kernels and retune quantizers for every new chip. Coral NPU’s Apache 2.0 license and shared compiler remove that tax. A model compiled once can target the 1 TOPS Coralboard SL for Linux robotics prototypes or the milliwatt SR200 series for wearables without starting over. Silicon partners such as VeriSilicon and GlobalFoundries are already supporting tape-outs, so the same software stack reaches production devices faster.

Developers who previously used older Coral Edge TPU boards will find the new NPU more flexible for transformers while retaining the focus on low power. The shift from closed toolchains to MLIR-based compilation aligns with how modern ML frameworks already export models.

Practical tips

Quantize aggressively. Coral NPU targets int8 and related low-precision formats for the best energy efficiency. Keep floating-point work on the host CPU when possible.

Profile memory bandwidth. The Matrix Execution Unit reuses weights inside local registers; models that spill frequently to external DRAM will not hit the 10 mW target.

Start with audio or simple vision. Keyword spotting, gesture classification, and compact speech models fit the milliwatt envelope more readily than large multimodal transformers.

Watch the roadmap. Google plans a 64-bit Coral NPU architecture and CHERI security support next. Current 32-bit silicon already supports the vector and matrix units needed for today’s wearable workloads.

Limitations to expect

Not every operator is fully accelerated yet. Unsupported layers fall back to the host CPU and can increase latency and power. Check the compiler logs before committing to a model.

Commercial SR200 chips are still in the sampling phase for many teams. The Coralboard SL is the fastest path to hands-on testing today.

Power numbers are targets under specific conditions. Real-world continuous sensing with sensors and radios active will raise average consumption.

Watch a related demo

The video shows on-device speech-to-text and a compact language model running on earlier Synaptics hardware that shares the Coral NPU lineage.

Conclusion

Google’s completed Coral NPU stack and the first commercial Synaptics silicon give developers a portable, open path to always-on edge AI. By compiling once with the IREE/MLIR toolchain and deploying to either the available Coralboard SL or upcoming SR200 boards, teams can move continuous sensing and compact models off the cloud and onto the device. Begin with the documentation and simulator, order a Coralboard, and measure real power and latency on your target workload.

Sources

• Google Developers Blog: Unifying Edge AI – Coral NPU open standard (20 October 2026)
• Synaptics press release: Astra SR200 series launch (20 October 2026)
• Google Coral documentation: developers.google.com/coral
• GitHub: google-coral/coralnpu, coralnpu-compiler, coralnpu-mpact