A number of the company’s founders come from mixed-signal IC design service house Kapik Integration Inc. (Toronto, Ontario) and the company was founded in February 2017.
Untether AI claims to be developing the highest-performance neural net inference engine, based on bus-free near-memory computing and enabled by software capable of preplacing data when necessary. The architecture is said to be scalable from mobile devices to data centres.
The so-called Kensington architecture gets data to the processors at 2.5 petabits per second, a rate 1,000 times better than what is possible in traditional architectures, the company said.
“We’re at a historic moment in computing right now. AI is changing the nature of what computers do. They can now interact with the real world based on recognizing patterns instead of following procedures,” said Martin Snelgrove, CEO of Untether AI, in a statement. “Now is the time to build a new kind of ultra-efficient chip for new frontiers in AI applications — from autonomous vehicles, data centers, mobile, vision processing and beyond — that will ultimately change how everything works.”
Dave Flanagan, senior managing director of Intel Capital, said that within five months of receiving seed capital Untether AI was able to build successful prototype based on its processing-near-memory architecture. “This is a team with decades of experience in chip design, and they’re now developing a commercial product for neural net inference – in an AI inference market that’s projected to grow to $22 billion by 2022.”
Darrick Wiebe, CTO of Untether AI, said: “Cloud AI inference has an efficiency problem; automotive AI chips are underpowered; and sophisticated AI at the edge will flatten batteries in no time. By introducing a completely scalable and ultra-efficient inference architecture, Untether AI is uniquely positioned to address each of those market segments with what will be by far the most energy-efficient AI accelerator.”