A new company, Unconventional AI, is tackling the significant challenge of reducing power consumption in AI inference processing. Led by Naveen Rao, former head of AI at Databricks, the company is developing a novel oscillator-based computer architecture that promises to cut power usage by a factor of 1,000. The company has released its first model AI, Un-0, an image-generation system that demonstrates the potential of this new architecture. Un-0's performance is comparable to state-of-the-art diffusion models, but it achieves this using a fundamentally different approach to computing. The oscillator-based architecture has the potential to greatly reduce power consumption, which is a major limitation in the widespread adoption of AI. As AI continues to scale, power usage is becoming a significant bottleneck, and Unconventional AI's approach may be one of the few viable solutions. The company plans to release schematics for its oscillator chips and build an entire inference stack from the ground up, with the goal of supplying compute capacity at a fraction of the power usage of traditional systems. With less than 50 employees, Unconventional AI is taking on a daunting task, but its ambitious goal may be necessary to meet the growing demand for AI inference. According to Rao, the available supply of power will be a hard limit for AI in the coming years, and Unconventional AI's approach may be one of the few ways to address this challenge.

Source: TechCrunch