Unconventional AI reckons ordinary chips are too thirsty, so it wants to rebuild AI computing from the floorboards up.
According to TechCrunch Unconventional AI boss Naveen Rao, formerly Databricks head of AI, claims it can make inference processing vastly more power efficient. Its secret sauce is an oscillator-based computer architecture.
On Thursday, the company released its first AI model, called Un-0. It is an image-generation tool meant to show that its strange new architecture can mimic conventional AI systems.
The accompanying paper explains how its researchers built a working image-generation model using a software simulation of the new architecture. The result performs at a similar level to state-of-the-art diffusion models, according to the company.
“This is the ‘hello world’ of a new kind of computer. Over the next year, you’re going to start seeing some pretty interesting news around this,” Rao said.
Un-0 produces images similar to models such as Stable Diffusion or OpenAI’s GPT Image 1. The interesting bit is not the pictures, but the weird route it takes to make them.
The model runs on an oscillator-based architecture, rather than the usual chips powering conventional computing and traditional LLMs. Rao thinks the approach could eventually cut power use by as much as 1,000 times.
The current Un-0 still runs on a software simulation, not a working chip you can chuck into a data centre. Unconventional AI plans to release schematics for an actual chip soon. After that, it wants to build a full inference stack from scratch and sell compute capacity like any other provider.
Rao said: “We will build a new kind of system composed of our chips. We will run AI models there, and we will have a network cable where prompts come in and inferences go out, but it’ll be done at 1/1000 of power.”
That is ambitious for a company with less than 50 staff. Still, with AI data centres slurping electricity like a stag party at a free bar, at least the problem is real.
Rao argues that energy will become the hard limit for AI scaling in the next few years. The industry can keep ordering more chips, but it cannot magic up endless power without annoying grids, governments and anyone with a monthly bill.
“AI scaling is hard because of energy. It’s going to be the fundamental limit in the next few years. You just can’t go past it. It’s going to be an energy-limited problem, at the end of the day,” he said.
