Ex-Databricks AI chief aims to cut AI energy use 1,000x
▼ Summary
– Unconventional AI, led by former Databricks AI head Naveen Rao, aims to rebuild computing architecture to drastically improve inference power efficiency.
– The company released its first model, Un0, an image-generation system that replicates conventional AI performance using a new oscillator-based architecture.
– Un0 runs on a software simulation of the oscillator chips, but the company plans to release actual chip schematics and build a full inference stack.
– Rao claims the oscillator-based architecture could reduce power use by up to 1000 times compared to conventional AI chips.
– The company believes energy supply will become a fundamental limit for AI scaling, and its technology is designed to address this constraint.
The quest to define the next frontier in artificial intelligence has fueled some extraordinarily ambitious ventures, but one startup is seizing the moment to completely reimagine the underlying computing architecture. Led by Naveen Rao, the former head of AI at Databricks, Unconventional AI is promising to make inference processing dramatically more energy-efficient through a novel oscillator-based computer architecture.
On Thursday, the company unveiled its first model, called Un0, an image-generation system that demonstrates how its technology can mimic conventional AI systems. A newly published paper from the company’s research team explains how they built a fully functional image generation model using a software simulation of this new architecture. The model performs on par with state-of-the-art diffusion models, marking a significant proof of concept.
“This is the ‘hello world’ of a new kind of computer,” Rao told TechCrunch. “Over the next year, you’re going to start seeing some pretty interesting news around this.”
The output from the Un0 model is comparable to popular systems like Stable Diffusion or OpenAI’s GPT Image 1. What sets it apart is the path it takes to reach that performance. The model relies on an oscillator-based architecture that is fundamentally different from the chips powering conventional computing and traditional large language models. While the technical advantages are complex, Rao believes this approach could ultimately cut power consumption by as much as 1,000 times.
Much of the necessary infrastructure is still under development. The current version of Un0 runs on a software simulation of Unconventional’s oscillator chips, but the company plans to release schematics for a physical chip soon. From there, the goal is to build an entire inference stack from the ground up, eventually offering compute capacity like any other cloud provider.
“We will build a new kind of system composed of our chips,” says Rao. “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.”
This is a staggeringly ambitious target, especially for a company with fewer than 50 employees. Yet, given the massive scale of the AI buildout and the soaring costs of meeting inference demand, it may be one of the few projects that matches the scope of the challenge. Rao argues that available power will become a hard limit for AI growth in the coming years, and Unconventional is uniquely positioned to address that bottleneck.
“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 says.
(Source: TechCrunch)
