At Sensor Expo 2018, Paul Washkewicz, Vice-President of Marketing for Eta Compute explains how his company’s application-specific integrated circuit (ASIC) uses asynchronous logic to lower the power to the microcontroller unit (MCU) to extend battery life. The ASIC is based on an ARM M3 and has a CoolFlux digital signal processor (DSP) from NXP in it as well.
Two artificial intelligence (AI) demonstrations using asynchronous technology allows them to support a spiking neural network (SSN) which is a very power efficient AI engine. Command recognition provides one of the demos. Building on that capability, in the Always-on Wake Up Word Demo, selecting the command words that are recognized initiates activity. The demo normally only draws a milliwatt until the voice command words wake it up. In addition to command recognition, it also performs noise cancellation and echo cancellation.
The energy efficient ASICs and SNN-based AI software developed by Eta Compute avoid need for many training samples for edge device applications where the amount of resources (both memory and computing) are limited.
Using their technology that includes process insensitive and low-voltage design, the company recently achieved the benchmark of an 8-order of magnitude improvement in model efficiency compared to convolutional neural networks (CNNs) and deep neural networks (DNNs) for key word recognition while consuming just 2 mW of power for inferencing.
For sensing applications, especially for motion and environmental sensors based on accelerometers, chemical sensors and gyros, Eta Compute’s methodology enables sensor hubs to perform more extensive sensor algorithms providing real time data and updates from mobile and Internet of Things (IoT) networked devices.
Filed Under: AI • machine learning, Sensor Tips