Synaptics' SR100 combines a microcontroller, the low-power processor that runs a device's real-time controls, with a neural processing unit for local AI models. Synaptics rates the AI block above 100 billion operations per second per watt and offers four versions sized for jobs such as wake-word detection, sensor analysis, and simple vision. This gives a product team a plausible path to run the model beside the control code instead of adding a separate accelerator and connecting the two chips.
That physical simplification matters more than the peak compute number. A separate accelerator needs its own power rails, board space, memory path, software package, and tests for every message crossing between processors. The SR100 keeps the model, a RISC-V processor, memory, and common device interfaces in one part. Synaptics also provides a software development kit with model conversion, profiling, and debugging tools, so the evaluation starts from an executable model rather than a data sheet and a bare chip.
Pre-production samples are available now, with mass production planned for early 2027. The open question is model support: efficiency on paper won't save time if a team's operators or network format need manual conversion. Embedded teams evaluating the part should bring one existing model and measure conversion effort, memory use, and control-loop timing before they redesign the board around it.