On October 20, at its Tech Day event, chipmaker Synaptics unveiled a new family of processors that could quietly change where artificial intelligence runs. The Astra SR200 series packs Google's open-source Coral neural processing unit (NPU) into ultra-low-power chips aimed at smart glasses, wearables, battery cameras and industrial sensors. The goal is simple but ambitious: run real AI — vision, voice and even transformer models — directly on tiny devices, without a constant link to the cloud.
It is the first production implementation of Google's open-source Coral NPU design, and it marks a milestone in how AI hardware is evolving: instead of every chip company designing its own AI accelerator from scratch, the industry is starting to rally around shared, open building blocks.
What Synaptics Actually Announced
The Astra SR200 is a new family of three chips, which Synaptics brands as "xMCUs" — processors meant to deliver microcontroller-level power efficiency alongside microprocessor-class AI performance. The three parts share one hardware foundation and one software development kit:
- Astra SR210 — 200 GOPS (giga-operations per second) of AI compute, positioned as the power-optimized entry point for always-on voice and sensing.
- Astra SR230 — 300 GOPS, adding vision, display and optional audio support for mainstream multimodal devices.
- Astra SR250 — 400 GOPS with dual vision and display paths, aimed at premium edge applications that need to process more than one camera feed or run richer on-device models.
A clever detail: the three chips are pin-compatible. A hardware team can design a single circuit board and swap in more powerful silicon as a product moves from prototype to shipping unit — no board redesign needed.
The Coral Connection: Google's Open-Source AI Chip Inside
The most interesting part of the SR200 is not the chip itself, but whose technology sits at its heart. Synaptics' Torq AI engine integrates the Coral NPU from Google Research — an AI accelerator core that Google describes as open-source intellectual property. According to Google's developer materials, Synaptics' implementation is the first production use of that open-source design.
This is not a brand-new relationship. Google and Synaptics announced their partnership in January 2025, combining the Coral NPU with Synaptics' Astra processor line. Developer boards based on the earlier Astra SL2610 silicon were showcased at Google I/O 2026, and the technology was previewed at Google Research's Research@ event on October 14 — days before the formal demonstration at Synaptics' Tech Day.
The open-source angle extends to software. The companies describe an open stack built on MLIR and IREE compiler tools, with support for the Zephyr real-time operating system — an effort to standardize machine-learning frameworks and reduce the fragmentation that has long plagued edge computing for IoT devices.
Google's strategy is becoming two-track: from our coverage of Google's Project Suncatcher, which puts AI data centers in space to open NPUs in pocket-sized chips, the company is pushing AI both beyond the data center and deep inside everyday devices.
Under the Hood: The Specs That Matter
Each SR200 chip pairs two Arm processor cores with the Torq AI engine. An 800 MHz Arm Cortex-M85 handles demanding tasks, while a smaller Cortex-M52 stays on to read sensors and run light AI work at low power. An optional Cadence HiFi 5 audio processor handles sound, and a dedicated always-on section listens for voice commands to wake the chip.
Memory is where the design breaks from typical microcontroller conventions. The chips pack up to 8 MB of integrated SRAM alongside custom DRAM that scales up to 256 MB — all on the chip itself. That removes the need for an external memory chip, which saves board space, cost and power. Synaptics says the architecture uses substantially less power than typical system DRAM, though real-world battery performance has not yet been independently measured.
On security, Synaptics lists hardware isolation, cryptographic acceleration and secure boot. The company describes PSA Level 3 and CC EAL4+ as certification targets — worth noting that these are targets, not completed certifications, so buyers should verify the status rather than assume it.
What These Chips Are Actually For
Synaptics is aiming the family squarely at devices where power, size and connectivity are constraints: AI wearables, smart glasses and hearables, battery-powered sensing devices, connected cameras, streamers, and industrial vision and voice systems.
A concrete example, drawn from analyst coverage of the announcement, shows why this matters. Take a battery-powered video doorbell. Today, such a doorbell wakes on motion, then often asks the cloud whether it saw a person or just a swaying branch. With an SR250-class chip, the always-on domain wakes the processor, and the on-device NPU classifies the event itself. Dual vision paths and on-chip DRAM can hold two camera streams without external memory. Only the clips that actually matter leave the house — which cuts upload traffic, reduces false alerts, and keeps most of your video inside your home.
In factories and warehouses, the same logic applies to machine vision and predictive maintenance: sensors that can interpret what they see locally don't need constant connectivity to a cloud AI service, and they keep working when the network goes down.
The Developer Angle: Coralboards You Can Buy Now
Chips announced at events can take a long time to reach real products, so Synaptics also moved on the developer side. The Coralboard SL — a development kit running the earlier Astra SL2619 processor with a 1-TOPS implementation of the Coral NPU, dual 2 GHz Arm Cortex-A55 cores and 2 GB of DDR4 memory — reached general availability through distributor DigiKey on October 20, moving from a limited-edition preview to a board anyone can order.
A second board, the Coralboard SR, was announced as a new product carrying the SR250 chip for low-power on-device AI development, though Synaptics has not disclosed a general sales date or price for it.
The boards and the chips play different roles: the Coralboards are for prototyping, letting engineers experiment with running models such as Google's Gemma locally; the SR200 chips are what will end up inside shipping products. That developer-first approach is how Google and Synaptics plan to build an ecosystem around the open Coral NPU before the silicon itself ships at volume.
Why This Launch Matters: Three Shifts
Beyond the product details, the SR200 announcement points to three broader trends in how AI hardware is developing.
1. Open-source silicon goes mainstream. For years, every edge AI chip vendor designed its own neural accelerator — a costly duplication of effort. Google open-sourcing the Coral NPU, and Synaptics adopting it instead of building a bespoke core, signals that edge AI silicon is consolidating around shared accelerator IP, much like Arm's CPU cores standardized mobile computing a decade ago.
2. The xMCU idea fills a real gap. There has long been a gulf between simple microcontrollers (cheap, miserly with power, but too weak for real AI) and application processors (powerful, but hungry). A chip that runs vision, voice and a display from a microcontroller-style power budget — and lets designers scale up through pin-compatible tiers — targets exactly the devices where that gap hurts most: things that live on batteries and wrists.
3. Privacy by architecture. On-device inference means data doesn't have to travel. A camera that decides locally whether a clip is worth sending is inherently more private than one that streams everything to the cloud — an argument that is getting harder to ignore as homes fill with connected cameras and microphones.
The Caveats, Honestly Stated
Balanced coverage requires noting what Synaptics has not said. The company published no pricing for the SR210, SR230 or SR250, and no firm production schedule — the chips remain in the sampling stage. The performance and efficiency figures are company claims rather than independent measurements; analysts note they should be treated as unverified until official datasheets and third-party testing appear. The same applies to the security certification targets.
It also helps to place the SR200 in the wider chip landscape. This is the edge of the edge — the opposite end of the spectrum from data-center silicon like Microsoft's Maia 300 AI chip, built to cut inference costs — and one of several competing visions for low-power AI hardware. GlobalFoundries' FDX Fusion chips, for instance, take a different route to powering physical AI. Which approach wins will be decided by developers and design wins, not launch-day presentations.
What Comes Next
The announcements to watch now are practical ones: independent power measurements, public pricing, and the first shipping products — likely wearables, smart cameras and industrial sensors — built around the SR200 family. Also worth watching is whether more chip vendors adopt Google's open Coral NPU design, which would confirm the consolidation trend this launch hints at.
Google's Coral project began years ago as a developer curiosity. With Synaptics putting it into commercial silicon, open-source AI hardware has crossed a threshold: from experiment to product. If the SR200 family delivers on its claims, the next wave of smart devices may owe less to the cloud — and more to a tiny, open chip running quietly inside.