In March, Texas Instruments announced two microcontroller families carrying its TinyEngine neural processing unit: the MSPM0G5187, built on an Arm Cortex-M0+, and the AM13Ex real-time family on a Cortex-M33. The claim is up to 90 times lower inference latency and more than 120 times lower energy per inference than comparable microcontrollers without an accelerator.
The number that changes the conversation is not the speed-up but the price. The MSPM0G5187 is quoted at under $1 in thousand-unit quantities. A Cortex-M0+ is an extremely modest core, and putting a usable inference engine beside one at that price moves machine learning out of the “flagship product” bracket and into ordinary industrial and consumer designs.
The realistic applications are unglamorous and genuinely useful: anomaly detection on a motor from vibration or current signature, keyword spotting, gesture recognition, sensor fusion that would otherwise need a much larger part. The NPU runs alongside the main core, so the control loop is not interrupted by inference.
The remaining constraint is memory, not compute. Flash and RAM on a part like this still set a hard ceiling on model size, so the engineering work moves to quantisation, feature selection and keeping the model small enough to fit — which was always the harder half of the problem.
Source: Texas Instruments