Complex AI technologies are held back by the limitations of traditional chip architecture. Untether AI is a Canadian group that develops extremely efficient chips that allows more rapid movement of data, opening new doors for AI applications.
Untether AI’s runAI200 chip minimizes the distance that data must travel to process information, thereby creating a far more efficient and cost-effective solution.
The broader field of machine learning could benefit greatly from this since computational data sets tend to be hundreds of gigabytes in size, and future applications include next-generation medical devices and consumer cell phones.
Founded in Toronto in 2018, the budding start-up is funded by Radical Ventures and Intel Capital and raised $26 million CAD for its Series A round in 2019.
“A new breed of AI accelerators”
In traditional chip architecture (AKA von Neumann architecture), 90% of the energy for AI workloads is taken up by data movement as it transfers between external memory, on-chip caches, and finally onto the computing element itself. Modern devices operate with strict control of their energy usage to prevent overheating, so they are generally unable to keep up with the increasing demands of AI computing workloads.
Not only are traditional chips costly in terms of energy and performance, but they’re also limited by transistor and frequency scaling, ultimately causing bottlenecks in data movement.
Eliminating as much complexity as possible while maintaining peak operational performance and placing that processing as close as can be to the system’s memory should theoretically use as little power as possible.
Untether AI’s “at-memory” architecture was designed with this principle in mind. By maximizing output while minimizing power usage, the chip provides “unrivalled compute density” and allows data to move 1,000 times faster.
The runAI200 chip can be found in Untether AI devices like the tsunAImi accelerator card, which contains four of them. This reportedly provides double the computing power of any available competing product.
“As AI compute requirements continue to explode, new architectures are needed to meet these demands,” said Linley Gwennap from The Linley Group, host of the industry’s premier processor conference.
“Untether AI’s runAI200 devices, with their innovative at-memory compute architecture, break through traditional von Neumann architecture bottlenecks and represent a new breed of AI accelerators.”
In an interview with Venture Beat in late 2019, CEO Arun Iyengar noted his excitement over the new investment partnerships and how they will help to advance the field: “The history for computing and AI chips is being written right now, and Untether AI is emerging as a pioneer in the space.”
“We look forward to working with world-class investors Intel Capital and Radical Ventures as we prepare to introduce groundbreaking chip innovations that will enable new frontiers in AI applications.”