Interview Naveen Rao founded AI businesses and sold them to Intel and Databricks. He’s now turned his attention to satisfying AI’s thirst for power and believes his new company, Unconventional AI, can do it by building chips inspired by nature.

On Monday, Rao revealed Unconventional AI raised $475 million in seed funding from Andreessen Horowitz, Lightspeed, Jeff Bezos, and others, to answer the question.

"AI is intrinsically linked to hardware and hardware is intrinsically linked to power. We can’t scale beyond a certain number of inferences per unit time because of the energy problem. We can’t produce that much more energy in the next 10 years," Rao told The Register.

With Unconventional AI, Rao makes the case we’re using the wrong tools for the job.

"Natural learning systems never used numerics. They didn’t simulate the dynamics of learning. They use the intrinsic physics of whatever substrate they’re on to build a learning system," Rao said. "We believe we can recapitulate that behavior in silicon."

Rao is no stranger to this concept. Prior to founding MosaicML and Nervana Systems, which were acquired by Databricks and Intel, respectively, Rao studied electrical engineering at Stanford and earned a PhD in neuroscience at Brown University.

The idea that biological systems shaped by millions of years of evolution may offer clues to more efficient computer architecture is not new: The likes of IBM and Intel have been chasing it for years. If our brains run on just 20 watts of bioelectric energy, imagine what we could do with a megawatt, never mind the gigawatt-scale datacenters now being built.

This class of computers is known as “Neuromorphics” and their designers aim to reverse engineer the inner workings of the brain and implement them in silicon. Despite decades of research in the field, only a handful of working prototypes have been built. None get remotely close to the performance and efficiency of the human brain, never mind lesser creatures like owls.

Slow progress doesn’t mean this approach is wrong. "Some of these things don’t work until they do. Neural networks were considered sort of a backwater until the mid 2000s," Rao said. That changed as compute became more plentiful.

Unconventional AI isn’t solely focused on neuromorphic computing. "The problem with neuromorphics is it has to work like the brain. But, why does it have to work like the brain," Rao said. "There’s probably concepts from the brain that are useful in building such a [learning] system. That’s the way we look at it. It’s not that it must work like the brain."

Instead, Rao tells us Unconventional AI’s lab is exploring several different approaches to improving the efficiency of machine learning accelerators. He declined to detail the company’s research, but what we do know is they’ll be fabbed in silicon and will likely be an analog chip rather than a digital device.

"These are nonlinear dynamics of circuits. That’s inherently an analog thing," he said. "All devices are analog, even ‘digital’ devices. We just engineer those circuits to behave digitally, but we’re largely erasing the richness of what those circuits can do by making them one and zero."

For a lot of computational workloads, the determinism afforded by digital systems is desirable. For example, you wouldn’t want a piece of accounting software that spits out a different answer every time.

However, machine learning is often nondeterministic in nature and so you don’t necessarily need a deterministic compute platform. Rao envisions scenarios where a combination of non-deterministic analog and deterministic digital logic are used to accelerate different aspects of machine learning workloads.

• Meta and Google turn to NextEra to feed insatiable datacenter power hunger
• IBM drops $11B on Confluent to feed next-gen AI ambitions
• Datacenters are hoarding grid power just in case, says Uptime Institute
• Amazon’s Trainium3 is the latest to conform to Nvidia’s mold

According to Rao, certain models are more amenable to the kinds of non-linear dynamics that Unconventional is targeting. "Things like diffusion models, flow models, energy-based models are things that inherently have dynamics," he said.

The CEO thinks solving this problem will take time.

"We’re not going to have a product in two years," he said. "This is largely a research effort for the next several years, and we’re really trying to crack a new paradigm."

Having said that, Rao does plan to share Unconventional AI’s findings along the way, potentially as soon as next year. "This is not something that we go off in a lab for four years and emerge with the solution," he said. "Over the next several months, we’re going to start releasing things."

And while Rao’s initial focus is on research, his long term aspiration is to build a systems company. ®