Skip to main content

NASA alumnus made a 256-core processor just to handle AI

knuedge 256 core stackable processors knupath01
Although no one is sure whether the AI devices of the future are going to kill us, take our jobs or just help us with mundane tasks, one thing we can agree on is that they will need power and lots of it. That’s why a number of companies have developed processors designed with processing tasks in mind, like ex-NASA engineer Daniel Goldin, who’s created a 256-core chip that’s helped him raise over $100 million in funding.

Goldin knows his way around running a company too, since he holds the record as the longest running NASA administrator, with his tenure extending through much of the Shuttle era, between 1992 and 2001. He has brought that expertise and engineering background to bear on KnuPath, his monstrous chip design.

This chip is being combined with another product known as KnuVerse, which we’re told is a neural computing system capable of “military grade” voice recognition: no matter how loud the environment you’re in, it will be able to understand you. Background interference is one of the biggest issues facing speech recognition systems — doing away with that could really change the industry.

Related: Elon Musk thinks we should insert ‘neural laces’ into our brains so we can become cyborgs

Combined with the KnuPath processor, which can be stacked with over half a million of its siblings, it’s hoped that this will create the foundation for some of the world’s first real AI, capable of understanding humans and completing their tasks for them.

Details on the chip’s capabilities are quite thin, though TechSpot claims it’s capable of “320Gbps of throughput,” and that its inter-rack latency when built into a server farm is around 400 milliseconds.

What’s really exciting about this technology though, is that it will be free – or at least, the APIs will be. Buying the hardware likely won’t be cheap, but if you can develop new software to make use of it for free, the applications for such a computational system could come thick and fast when KnuPath and KnuEdge are released sometime in the latter half of this year.

Editors' Recommendations