Google is taking another step toward moving AI beyond the data center and into the devices people use every day, partnering with semiconductor manufacturer Synaptics on a new generation of processors designed to run AI models directly inside smart glasses, cameras, industrial sensors and other connected devices.
Synaptics announced its Astra SR200 series on Tuesday, October 20, introducing three processors that combine Google's open-source Coral neural processing unit (NPU) technology with specialized memory, audio, vision and computing capabilities.
The companies also announced broader availability of their jointly developed Coralboard platform, which allows engineers to experiment with running models such as Google's Gemma locally before incorporating the technology into commercial products.
For enterprises developing connected products, the announcements address a persistent challenge: running increasingly sophisticated AI without requiring expensive processors, constant internet connections or power-hungry hardware that drains device batteries.
Learn more about the latest in AI infrastructure tech and cost management at VentureBeat’s Agentic Infrastructure Stack Summit in San Francisco on October 28.
Google wants to make AI hardware as accessible as AI models
The initiative represents a potentially significant development in Google's broader AI strategy, which increasingly encompasses not only developing models but also making the underlying technology available for other companies to build upon.
Google Research open-sourced its Coral NPU architecture in October 2025. Rather than manufacturing the chips itself, Google makes the underlying accelerator design available for semiconductor companies to incorporate into their own processors.
The architecture uses RISC-V, an open standard for processor instructions, alongside software designed to help developers move AI applications between different hardware implementations.
In an exclusive briefing with VentureBeat on October 13, Billy Rutledge, director at Google Research and leader of its Systems Research team, described the challenges developers encounter when trying to run the same AI models across the many proprietary chips found in consumer and industrial devices.
"One of our goals was to try to minimize developer pain," Rutledge said, explaining Google's decision to release an open-source NPU design that multiple chip manufacturers could adopt.
Synaptics is Google's first commercial silicon partner for Coral, but Google's ambition extends beyond a single manufacturer.
Under the Apache 2.0 license, other companies can incorporate the Coral NPU architecture into commercial processors without paying Google licensing fees or royalties.
Synaptics combines that architecture with its own proprietary processing technology. Its open-source Torq compiler, built using the MLIR and IREE software projects, aims to help developers deploy models created using popular frameworks, including PyTorch, JAX and Google's LiteRT.
That approach could reduce the engineering work required to adapt AI applications to different hardware platforms.
Powerful enough for AI, small enough for wearables
The new Astra SR200 family consists of three processors with different capabilities.
The entry-level SR210 delivers 200 billion operations per second (GOPS) of AI processing performance, targeting always-on voice recognition and sensing applications. The SR230 increases that to 300 GOPS and adds capabilities for vision and displays, while the flagship SR250 reaches 400 GOPS and supports more demanding multimodal applications.
The processors combine Arm Cortex-M85 and Cortex-M52 cores with Google's Coral NPU, allowing manufacturers to distribute computing tasks across components optimized for different performance and power requirements.
Crucially, Synaptics also integrates up to 8MB of SRAM and custom DRAM configurations reaching hundreds of megabytes, with the forthcoming Coralboard SR supporting up to 256MB.
That additional memory could help devices run larger AI models and process information from cameras, microphones and other sensors without requiring a separate application processor.
Synaptics claims its custom memory architecture consumes approximately one-third the power of conventional system DRAM in comparable workloads, although independent benchmarks have not yet established the overall battery-life advantage.
The company is entering an increasingly competitive market. STMicroelectronics already offers its STM32N6 microcontroller, which delivers up to 600 GOPS through a proprietary AI accelerator, while Renesas introduced its RA8P1 family in 2025 with up to 256 GOPS.
Those peak performance figures do not directly establish which processor runs actual AI models faster or more efficiently. Memory capacity, power consumption, software compatibility and the particular models involved can substantially affect real-world performance.
Synaptics argues that its combination of integrated memory, multimodal capabilities and open developer tools distinguishes its approach.
From Gemma models to factory cameras and smart glasses
Google and Synaptics are targeting a growing range of smaller AI models designed to run locally.
The companies identify Gemma 3 270M, Liquid AI's LFM2.5 models and Ultralytics' YOLO computer vision models among the workloads supported or targeted by their developer ecosystem.
Synaptics reports that Gemma 3 270M can generate approximately 12 to 16 tokens per second using different model compression settings, although it has not provided a comprehensive independent performance comparison for the new SR200 hardware.
Potential enterprise applications include industrial smart glasses that interpret their surroundings, factory cameras that identify objects, and security products that analyze audio and video locally rather than continuously transmitting information to cloud servers.
Synaptics highlighted relationships with audio manufacturer Sonos, security provider Verisure and augmented-reality hardware company Jorjin Technologies in its announcement. Their statements describe existing work with Astra or potential applications for the new processors, rather than confirmed shipments of SR200-powered products.
Rutledge emphasized during the VentureBeat briefing that Google sees the developer ecosystem as a way to encourage entirely new product categories, not simply upgrade existing hardware.
"The most useful devices that will use AI at the Edge are still yet to be designed," he said.
Coralboard available for $99.79; new hardware still rolling out
Developers can begin experimenting with the technology immediately through the Coralboard SL, which uses Synaptics' existing Astra SL2619 processor with a 1-TOPS AI accelerator and 2GB of DDR4 memory.
The board is listed by DigiKey for $99.79, with Synaptics announcing general availability through the distributor on October 20.
A second development board, Coralboard SR, will incorporate the new SR250 processor, but Synaptics describes that model as forthcoming. The SR200 family is in the sampling stage, and the company has not disclosed individual chip prices or a firm schedule for volume production.
The launch also comes as Synaptics prepares for its proposed $5.7 billion acquisition by onsemi, a transaction currently expected to close by mid-2027.
For Google, the longer-term opportunity is potentially broader than any single chip. By making AI accelerator designs openly available while collaborating with manufacturers and model developers, the company is attempting to establish common foundations for an emerging generation of intelligent devices.
Whether the strategy succeeds will depend on how effectively manufacturers turn those foundations into affordable products that deliver meaningful AI capabilities without sacrificing battery life, privacy or reliability.