When Eta Compute was first founded in 2015, it was one of very few that adhered to the claim that spiking neural networks (SNNs) would be the low-power path to AI for small and battery-powered sensors and devices.
Even when the company was demonstrating its successes with SNNs back in 2018 with its new TENSAI SNN chip, the development of which was supported by tens of millions in venture capital, Eta knew that the technology was not yet ready for market. Despite being able to handle unsupervised learning while remaining ultra-power-efficient, the technology core to TENSAI needed further research and development.
This is no longer the case, however.
A solar-powered sensing board with Eta Compute’s TENSAI chip. Image used courtesy of Eta Compute via GloveNewswire.
Eta’s First SNN-backed Chips Are Now Shipping
In early February at the tinyML Summit when Eva announced that its first production chips using SNNs are now shipping. With these chips, the company has focussed on bringing to market a product that is aligned with more traditional neural networks such as deep learning.
Like its predecessor, TENSAI, Eva’s new chip, the ECM3532, is extremely power efficient. It consumes as little as 100 microwatts and has been designed to perform AI-enabled tasks like object recognition, data analysis, and sound recognition, helping it find applications across a broad range of both consumer and industrial applications.
Eva’s CEO, Ted Tewksbury, said that “We’ve got a number of customers lined up working on a number of different projects…. These customers have just been waiting for the production silicon and as soon as we get it in their hands, we are confident that we’re going to start to sell.”
Eta Compute’s ECM3532 is an AI multicore processor developed for embedded sensor applications. Image used courtesy of Eta Compute.
Driving AI to ‘Edge’ Devices
The ECM 3532 is a system-on-chip (SoC) that has been built around an Arm Cortex-M3 processor core. What makes the chip work, however, is Eta’s own proprietary technology known as CVFS (continuous voltage and frequency scaling). This allows the system to independently throttle the voltage and frequency of each core, lowering the operating voltage of a circuit and thus saving power.
Although CVFS may sound similar to an existing technology called dynamic voltage frequency scaling (DVFS), there is a difference – DVFS only allows a set of discrete voltages and frequencies whereas CVFS voltages and frequencies can range over a continuum thanks to algorithmic analysis.
Through its power efficiency, the ECM 3532 may help bring AI to more so-called “edge” devices. This is because the chip helps batteries conserve power by eliminating the need for devices to send a continuous stream of data to the cloud.
An example provided by Eva was that of smart building control. By deploying low-power cameras powered by AI to determine whether a room is occupied and if it is, how many people are in it. The use of AI negates the need for the camera to send a video stream to the cloud because it can simply report the number of people that it is detecting.