In this week’s edition of The Prototype, we look at chips designed for AI, building modular nuclear reactors, creating real-life web fluid, the Nobel Prize and more. You can sign up to get The Prototype in your inbox here.

Let’s get started, shall we?

Walter Goodwin, CEO of Fractile

Last week, OpenAI announced that it had raised $6.6 billion in new investments. But despite this staggering amount of money, The Information reported that the company expects to spend up to $9.5 billion annually in computing costs over the next few years to train new models for research.

A big part of this expense has to do with the underlying architecture of computers, which keeps memory and processing in separate parts of a chip, requiring a constant shuffling of data between those two parts. When it comes to AI models that are crunching massive amounts of data, this constant moving between memory and processing accounts for a significant amount of the energy consumed by chips.

U.K.-based hardware startup Fractile, recently emerged from stealth with $15 million in backing, is developing an AI co-processor that integrates memory and processing together. While it’s not the first company to work on a new hardware architecture–competitors like Groq and Cerebras have raised hundreds of millions between them–CEO Walter Goodwin told Forbes, one thing that distinguishes his company’s approach is that the chips will only handle the “very small set of operations” that AI algorithms utilize. So rather than a general-purpose chip, Fractile’s hardware is specific for LLMs. The company said this means it could run models up to 100 times faster at one-tenth of the current cost.

Fractile still has a ways to go, though. Currently, Goodwin said it has “a bunch of prototyping efforts” for its new chips, and added that “in terms of getting to a full first product, I think we have a very clear path to get there.”

Stay........

© Forbes