Research note: ETA compute makes Edge AI assumptions a reality

ETA Compute, a company created in 2015 by experienced start-up entrepreneur Gopal Raghavan, the co-founder of Inphi Corporation, a $ 4 billion publicly-traded company, and with US$ 8 million in funding from an investor syndicate including Venture Capitalist and Cadence CEO Lip-Bu Tan, has just announced the release of its system on a chip dedicated to Edge AI, a new emerging trend in the connected world industry.

For the last ten years, companies have been using IOT to connect their devices to the Internet. Just connecting the device was not enough, companies wanted to the device to be smart, able to work remotely and disconnected. Furthermore, IOT applications were too expensive to manage, required less latency and more privacy and security. To solve all those problems, more processing needed to happen on or close to the devices. The bellwether example for this trend is cameras, which now send to the cloud only anomalies or just what needs to be tracked within a video stream, reducing by two orders of magnitude the amount of information sent to the cloud, and thereby lowering power consumption significantly. Edge AI , as it is called, is a nascent but appears to be one of the key core technologies of the 4th industrial revolution. In terms of the number of units, this vibrant market is forecasted to triple between 2019 and 2024, Gartner and our own research predict a 2 to 5 years to mainstream adoption as well.

This market is also very ripe for start-ups as creating AI chips that are low cost, low power, able to act independently requires radical new thinking. And it has been very rewarding. Xnor.ai got acquired by Apple for US$ 200 million and many start-ups were successful at fundraising. Dedicated conference are emerging such as TinyML who gathered more than 500 attendees last week.

An example of successful start-up is ETA Compute, which released its chips last week at the TinyML conference. ECM 3532, is a system on a chip, built around an Arm Cortex-M3 processor core and an NXP CoolFlux DSP core. It has everything needed to run an AI processor. ETA Compute secret sauce is a technology called CVFS that throttles the voltage and frequency of each core independently, which saves power but lowers computation as well. ECM 3532 works very well in many IOT scenarios as the stream of data collected in the case of sensors, audio, or video is usually discrete, and the chip runs at low voltage and low frequencies while waiting for an anomaly to occur.

ETA compute chips only process inference, so algorithm training happens somewhere else. Once new rules emerge, the company has built the infrastructure to update firmware and software on the chip through its delivery platform. The company leverages Tensorflow and tinyML developers’ ecosystem for software development.

ECM 3532 seems to have many attributes needed for market adoption. Jim Feldhan from Semico Research thinks that Eta Compute’s technology “is orders of magnitude more power-efficient than any other technology and should make AI at the edge a reality.” Furthermore, Semir Haddad, ETA Compute Senior Director of Product Marketing, also indicated the total cost of ownership is low enough to fit into the budget of an existing product.

Use cases ETA compute is targeting are asset Tracking (motion sensors, sound sensors), vision applications in industries such as smart building, smart retail, and consumer goods.