Unconventional AI unveils oscillator-based architecture with 1,000x power efficiency claim for inference

Unconventional AI, a startup led by former Databricks AI chief Naveen Rao, released Un-0, an image-generation model intended to demonstrate the feasibility of its oscillator-based computing architecture. The model replicates the performance of state-of-the-art diffusion models while running on a software simulation of the company’s planned hardware, according to the company.

In an accompanying research paper, the team claims the Un-0 model performs comparably to conventional diffusion models despite operating on a fundamentally different architecture. Rao described the release as the “hello world” of a new kind of computer and indicated that further developments are expected over the next year.

The core innovation is an oscillator-based architecture, which differs from the transistor-based chips used in conventional computing and traditional large language models. Rao argues this approach will ultimately reduce power consumption by as much as 1,000 times compared to existing AI hardware.

Currently, Un-0 runs on a software simulation of Unconventional AI’s oscillator chips. The company plans to release schematics for an actual chip and build an inference stack from the ground up, with the goal of providing compute capacity where prompts enter and inferences exit at approximately 1/1000th the power of current systems.

Unconventional AI emphasizes energy efficiency as a limiting factor for AI scaling. Rao stated that energy constraints will be a fundamental barrier in the coming years, framing the company’s work as a response to an anticipated energy-limited future for AI.