Small devices are getting a chip family designed to process vision, audio and AI locally. Synaptics introduced its Astra SR200 series on October 20, using Google's open Coral neural-processing architecture.
Local processing matters when a device must react continuously without sending every sensor reading to a remote server. Cameras, wearables and industrial equipment also face limits on battery power, heat and space.
The family has three models: SR210, SR230 and SR250. Synaptics lists neural-compute capacities of 200, 300 and 400 billion operations per second respectively. Those are hardware throughput specifications, not measurements of application accuracy or battery life.
The chips combine Arm processing cores with specialized vision and audio hardware and the Torq neural accelerator. A neural-processing unit, or NPU, handles operations used repeatedly by machine-learning models.
All three models use the same pin-compatible package, according to Synaptics. A manufacturer can therefore design a board around the family and change the performance tier without redesigning every physical connection.
Google describes the underlying Coral work as an open standard supported by the IREE and MLIR software toolchain. That aims to reduce the work required to move a model from software into a power-constrained device.
The launch does not establish that a finished pair of glasses or a camera will achieve a particular runtime. Sensors, memory traffic and the chosen model also affect a product's behavior.
The concrete change is a path from an open accelerator design to announced commercial silicon. The device still needs its own engineering.
