Tech —

Intel to build discrete GPUs, hires AMD’s top graphics guy to do it

Intel is making a third attempt to muscle in on Nvidia and AMD's turf.

Intel

One might have thought that with Monday's announcement that Intel is going to produce processors with embedded AMD GPUs, the two processor companies were on good terms. That's looking a little less likely now. On Tuesday, AMD announced that Raja Koduri, its chief GPU architect, will leave the company. Where is he going? That question was resolved on Wednesday: Intel. And what's he going to do at Intel? He's going to be the senior vice president of a new group—Core and Visual Computing—that will expand Intel's graphics reach both into the low-end, with integrated graphics reaching into the Internet-of-Things space, and more excitingly, at the high end, with discrete GPUs.

Koduri led AMD's Radeon Technologies Group, responsible for both AMD's discrete and integrated GPUs. Before that, he was director of graphics technology at Apple.

Intel has dabbled with discrete GPUs before; its 740 GPU, released in 1998, was a standalone part using the then-new AGP port. A second attempt to build a standalone GPU was the Larrabee project, but that never shipped as a GPU. In 2009, Larrabee was repositioned, with Intel deciding to make it a massively multicore accelerator—the predecessor to the current Xeon Phi chips—rather than a graphics processor.

With neither 740 nor Larrabee delivering the GPU market success that Intel hoped for, the company has stuck with integrated GPUs since. The company's mainstream GPU architecture, called "Gen" for want of a better term, had its ninth major revision in the Skylake processor. It's been marginally revised for Kaby Lake, with current chips being Gen 9.5. Intel has used iterations of Gen across most of its product line, though at the very lowest end it has also shipped third-party GPUs.

The Gen architecture is arguably one of the more advanced GPU designs on the market; the individual execution units (EUs) used for computing shader programs are very flexible and can be independently programmed, features that should make the GPUs readily adaptable to both computational tasks as well as straight graphics. The design also offers some degree of scalability; internally it's organized into "slices" of 24 EUs each, and the company has shipped GPUs with one, two, or three slices.

But in comparison to the discrete GPUs from AMD and Nvidia, Intel's GPUs have tended to be small, with far fewer execution resources than would be found in a discrete part. And while Intel has built some processors with embedded DRAM to help accelerate graphics performance, it has never paired its GPUs with large pools of high-speed dedicated graphics memory.

Intel's render of an 8th-generation H-series processor. The discrete GPU and stacked HBM2 memory are side by side.
Enlarge / Intel's render of an 8th-generation H-series processor. The discrete GPU and stacked HBM2 memory are side by side.
Intel

Nonetheless, the Intel integrated graphics have been squarely in the "good enough" territory for most laptop and a large proportion of desktop users. This "good enough" performance has given Intel something between about 60 and 75 percent of the graphics market—but if the company wants to move beyond that into the smaller but lucrative discrete graphics market, it needs to be building discrete parts with much better performance. Strength in this market would also give Intel a stronger position in fields such as machine learning and supercomputing, as these are markets where GPUs and GPU-like chips have found considerable reach.

What's not clear at this time is precisely how Intel wants to move into this space. An expanded and enlarged version of its Gen architecture would be the quickest win—slap together a bunch of slices, dedicated graphics memory, and a hefty power budget, and you have a discrete GPU. If Koduri has been brought on to develop an all-new architecture, we would be unlikely to see the fruits of that work for a good four years. A hybrid of the two plans is also possible; a beefed up Gen9.5 part now and a new architecture later.

Whatever the plan, hiring Koduri suggests that Intel is taking this market seriously. Intel may have failed to crack the discrete GPU market twice already, but the third time may prove to be the charm. If Intel can find the success that has so far eluded it—certainly possible, given its still strong manufacturing capabilities and new expertise—then it could put the squeeze on both Nvidia and AMD.

Listing image by Intel

Channel Ars Technica