Locally-processed local fall detection, voice recognition and face ID will be running on its ‘GPX10 Pro’ IC.
Inside, AI processing is performed in two sets of five cores (called ‘MX8’) in two separate power domains.
“One set is in an always-on block which supports ultra-lower power sensor interfacing and fusion – for instance, when performing always-on keyword spotting, the chip consumes less than 100µW,” according to the company. “10 MX8 cores perform up to 2,560 multiply-accumulate operations per cycle, producing total peak AI throughput of 512Gop/s.”
Power is saved compared with all digital processors by executing multiply-accumulate in sram-like analogue multiplier cells, the outputs of which are analogue-summed.
Either side of the analogue layers are DACs and ADCs that push data in and extract results.
“GX10 Pro’s compute function is supported by 2Mbte of on-chip sram, ten times more than in the existing GPX10, to enable implementation of larger and more complex AI models,” said the company.
Classic control functions are taken care of by an Arm Cortex-M4F CPU.
Interfaces are provided for up to eight simultaneous analogue sensors and 20 digital sensors.
Support comes from a tool-chain dubbed ‘Nebula’ which is compatible with model training frameworks including TensorFlow, Keras and Onnx.
Beyond this is ‘SenseMesh’, a hardware sensor fusion layer that produces responses to trigger events, off-loading sensor polling from the CPU.
The device is sampling now, with volume production expected in Q1 2026.
Find the GPX10 Pro demonstrations on stand 5.E31 at Electronica in Bangalore (17 – 19 September).
Ambient Scientific headquartered in Santa Clara, with development there and in Bangalore.
