Google and Synaptics Put Open-Source Coral AI Chips Into Devices: What Owners Should Ask Before the Demo

A $100 board that runs a small language model without the cloud

Synaptics announced the Astra SR200 series on October 20, 2026, describing it as a new class of microcontroller it calls the xMCU. The three pin-compatible parts, the SR210, SR230 and SR250, deliver 200, 300 and 400 billion operations per second of AI processing respectively, and combine Arm Cortex-M85 and Cortex-M52 cores with the Coral neural processing unit from Google Research. The company also said the Coralboard SL development board, built with Google Research and powered by the earlier Astra SL2619, now ships generally from the distributor DigiKey, while a second board, the Coralboard SR, based on the new SR250, is forthcoming.

The connection to Google matters more than the clock speed. Google open-sourced its Coral NPU architecture in October 2025 under the Apache 2.0 license, which means chip makers can incorporate the accelerator design into commercial processors without paying Google licensing fees. Synaptics is the first commercial silicon partner, and its Torq compiler, built on the open-source MLIR and IREE projects, is intended to let developers move models trained in common frameworks such as PyTorch, JAX and Google's LiteRT onto this hardware. VentureBeat, which reported the launch, quoted Billy Rutledge of Google Research describing the goal as minimizing developer pain across many proprietary chips.

What the vendors claim, and what is still unproven

The numbers on the table are vendor figures. Synaptics reports that Google's Gemma 3 270M model can generate roughly 12 to 16 tokens per second on the platform depending on compression settings, and claims its custom memory architecture consumes about one third the power of conventional system DRAM in comparable workloads. VentureBeat notes plainly that independent benchmarks have not established the overall battery-life advantage, and that peak operation counts do not show which processor runs actual models faster once memory, power and software compatibility are considered.

Availability is also partial. The SR200 chips are in the sampling stage, Synaptics has not disclosed individual chip prices or a firm volume production schedule, and the Coralboard SR board is described as forthcoming. The board you can buy today is the Coralboard SL at $99.79 according to VentureBeat's reporting of the DigiKey listing. The launch also arrives while Synaptics awaits a proposed acquisition by onsemi that both companies expect to close in mid-2027, a corporate change any long-term hardware commitment should weigh.

What this means for your business

For a manufacturer, warehouse or service firm, the practical question is not whether a chip exists but whether a small model can run on your own equipment, on your own network, without sending recordings, images or sensor readings to a cloud service. A $100 board that runs a speech or vision model locally is a cheap way to answer that question for one workflow. It can test whether on-device inference is accurate enough, fast enough and private enough before anyone commits to product hardware.

Detailed In Design could build that kind of bounded evaluation as custom work: a Coralboard-based prototype that runs a small open model against your own sample data, such as recorded phone inquiries, inspection photos or sensor logs, with a staff-reviewed workflow around the output. The result would be a written finding on accuracy, latency and privacy for your specific case, not a product promise. The limitations are real. These are tiny processors, the vendor performance figures are not independently verified, the chip family is still in sampling, and any useful deployment still needs a person who owns what the system is allowed to do with the results.

A small next step

Before evaluating any edge AI hardware, write down one workflow where keeping data local matters, what decision the software would support, and who reviews its output. If you cannot name those three things, a development board is an expensive toy. If you can, a small evaluation project is a reasonable way to find out whether the local-inference conversation applies to your business at all.