Intel is building a dGPU—are they serious?

Do they know how hard that is?

Jon Peddie
Photo credit Cottonbro

When Intel launched their Larrabee project back in 2009 to build a revolutionary MIMD style GPU they hired just about anyone who could spell GPU. Some of those people left when the project was canceled, but many of them stayed, scars and all. Then in 2018 when the company valiantly decided to take another run at it but failing back to a traditional SIMD approach this time, they again went on a hiring binge. As a result, Intel probably has more GPU experts than all – yes ALL the other GPU and GPU IP suppliers in the world.

Intel has been building and shipping a GPU since 2001. Not a frame-busting super-resolution heat-sink melting GPU like AMD and Nvidia build, but a functional one that hundreds of millions of people are using every day all day.

And those hundreds of millions of users have provided, ahem, some feedback. So, in addition to being the biggest GPU expert employer in the world, and the biggest GPU shipper in the world, they also have the biggest focus group in the world. Now, what could a company do with those kinds of resources?  Oh, and let’s not forget they have a gazillion dollars in the bank and some of the best fabs in the world.

So maybe, just maybe they are really really REALLY  serious about building a discrete GPU. But not just a GPU— a whole family of GPUs. They call it (as if you didn’t know)—Xe.

Xe, according to Intel’s chief spokesperson Raja Koduri is a scalable and extendable architecture that can satisfy the requirements for a free GPU found in an x86 processor (like Sky Lake) but also a  GPU that can do the heavy lifting of compute acceleration for supercomputers like the Ponte Vecchio Intel is building for Argonne National Laboratories—and everything in between including the take-no-prisoners gamers. That’s a tall order, and not what one would expect from a start-up or newbie—but then, Intel is neither.

If you were a defense attorney and your client was on the stand, Intel would be the perfect witness because they’d say all the right things. It’d be a checklist: the right amount of shaders per block, check; right amount tiles, check; correct fabric, check; power management, check; best API and driver, check; right fab, ah, well, yeah, check.

Not only that, but the company has the audacity to suggest that they will actually ship some of these Xe things this year.

So, if Intel was a poker player, and they said they were going to build a scalable GPU they’d be a pretty good bluffer. Don’t think I’d bet against them.

Image credit: Julián del Nogal

What if Intel bought Arm?