Google Coral NPU + Synaptics: $99 edge AI board ships

Edge AI just got cheaper to try. Synaptics announced its Astra SR200 chips on Tuesday, built with Google's open-source Coral NPU, and the Coralboard SL developer kit is now generally available for $99.79 through DigiKey. The point is simple: small AI models running on the device, next to the camera or microphone, with no cloud call and no per-request bill.

What actually happened

Per VentureBeat and Google's Coral product brief:

• The design is open. Google Research released the Coral NPU as RISC-V-based accelerator IP under the Apache 2.0 license: no license fees, no unit royalties. Google does not make the chips. Synaptics is its first commercial silicon partner.
• The new chips: the Astra SR200 family. The SR210 is rated at 200 GOPS for always-on voice and sensing, the SR230 at 300 GOPS with vision and display, and the SR250 at 400 GOPS for multimodal work. They are sampling now. Synaptics has not given chip prices or a volume production date.
• The board you can buy: the Coralboard SL, with an Astra SL2619 processor, a 1 TOPS AI accelerator, and 2GB of DDR4, at $99.79. A Coralboard SR built on the SR250 comes later.
• The models: named targets include Gemma 3 270M, Liquid AI's LFM2.5, and Ultralytics YOLO. Synaptics says Gemma 3 270M runs at about 12 to 16 tokens per second. Nobody has published an independent benchmark yet.
• The compiler: Synaptics' open-source Torq compiler takes models from PyTorch, JAX, and LiteRT.

Why edge AI matters for your business

Some data should never leave the building. A camera that counts parts on a bench, checks a shelf, or flags a forklift in the wrong lane does not need to stream video to a cloud model. On-device inference keeps footage on site, keeps working when the internet drops, and costs nothing per frame.

Small models handle narrow jobs. You will not run a frontier model on a $99 board. You can run object detection, wake words, a defect check, or a 270M-parameter model that sorts short text. Most physical-world automation is a narrow job.

Open IP lowers the lock-in risk. Apache 2.0 hardware IP plus an open compiler means you are not tied to one vendor's toolchain for the model side. Check that before you build: the board is cheap, but rewriting your pipeline for a closed SDK later is not.

Prototype on the board, not the spec sheet. Peak GOPS numbers do not tell you real speed. Buy one board, run your actual model on your actual camera feed, and measure.

Sources: VentureBeat, Google Coral product brief.