Last week brought a sensational piece of news. Ever since AMD acquired GPU maker ATi almost twenty years ago, people have speculated whether Intel and Nvidia might one day strike a similar alliance. It seems that time has finally come, though only now that Intel is in trouble and has to push its pride aside. The two companies announced not a merger, but a far-reaching partnership. It will be interesting to watch where this leads.
Both companies revealed yesterday that they will combine Intel’s processor technology with Nvidia’s GPU technology in jointly developed products. The result will be Intel x86 processors paired with Nvidia graphics, or more precisely, Nvidia GPU accelerators.
Custom processor for Nvidia’s AI servers
Nvidia hasn’t said much about these products, but Intel has been more open. For now, two projects have been announced. The first will be a special x86-based server processor designed for Nvidia, meant for use in its AI platforms. Over the years, Nvidia has deployed AMD Epyc CPUs, Intel Xeons, and now its own Arm Grace CPUs (though they simply use the licensed ARM Neoverse architecture, not fully in-house developed cores) in its DGX and HGX servers. Intel will apparently supply special models or entirely custom designs tailored specifically for Nvidia’s use to keep their spot in Nvidia’s racks.
This likely means NVLink support baked into the CPU, allowing direct coherent connections with compute GPUs. These processors themselves won’t contain Nvidia GPUs; those will stay separate. But the key advantage is having that native CPU–GPU link, something other CPUs can’t offer (though AMD offers something similar in their Infinity Fabric, which links Epyc CPUs to Instinct GPUs).
X86 RTX SoC: Intel processors with GeForce GPUs integrated
What’s probably more interesting for us is the second product. It’s already confirmed that this partnership will also deliver processors for PCs and laptops (the consumer market). These CPUs will use Intel’s x86 architecture but feature integrated Nvidia graphics – Intel calls them “x86 RTX SoCs.” They’re pitched as solutions for a broad range of computers requiring top-tier GPU and CPU performance.
This GPU will come as a chiplet, so it won’t (yet) be a monolithic integration of both companies’ technology on the same die. However, Intel already uses a separate GPU chiplet for its own Arc graphics in today’s Core Ultra CPUs anyway so this isn’t entirely different. It’s possible the final product may resemble MediaTek’s GB10 CPU, which we looked at recently:
GB10 is made up of a chiplet that provides practically the complete functionality of the processor, chipset, and so on, but the integrated graphics is added in the form of another chiplet developed and supplied by Nvidia. The logic blocks in the MediaTek chiplet include implementation of the NVLink-C2C interface (the abbreviation means it is a derivative of NVLink for connecting two chiplets over a short distance), which makes it possible to join the two parts from different manufacturers into one. Intel can do the same thing as MediaTek; in theory it might be possible to design an Intel processor supporting the usage of the same GPU as the GB10. In reality, though, the Intel-Nvidia collaboration will probably feature a different GPU chiplet. This joint product is also not particularly close to release yet, so by the time it reaches the market, the next generation GPU architecture may already be around to be used instead of Blackwell.
Further details are still emerging, so in the coming days we’ll get a clearer picture of whether this is just a “side business” or the start of a deeper, lasting integration of Nvidia GPU tech into Intel CPUs.
Intel once tried this with AMD
Some of you may remember that Intel already tried something similar back in 2018, but with AMD. The result was the Intel Kaby Lake-G processor, which paired a 14nm Intel quad-core CPU with a Radeon GPU. This GPU had its own HBM2 memory and was essentially a standalone graphics chip mounted directly onto the CPU package alongside its dedicated VRAM.
That product didn’t succeed, shipping only in limited number of systems before the partnership was dissolved, and Kaby Lake-G never had a successor. Something similar could, in theory, happen to this partnership too. But given Nvidia’s far greater clout, it feels more likely that the joint x86 RTX SoC will gain traction.
Nvidia to own a slice of Intel
The seriousness of this partnership is underscored by the fact that Nvidia will now literally own a piece of Intel. Alongside the technical collaboration discussed so far, Nvidia has also become an Intel investor, purchasing 5 billion USD worth of shares.
Nvidia CEO Jen-Hsun Huang is basically taking advantage of Intel’s current weakness that collapsed the stock price, which makes the whole company relatively cheap. Nvidia paid 23.28 USD per share, giving it an estimated 5% stake (these are apparently newly issued shares, which will dilute existing shareholders).
It wasn’t so long ago that Intel was a much bigger, stronger company than Nvidia. The idea that a mere GPU maker could not only surpass Intel but overshadow it so overwhelmingly that it could effectively buy into the company with pocket change (in the last quarter alone, Nvidia made 26 billion USD in net profit, enough to buy a quarter of Intel at the current market cap outright) would once have seemed like wild fantasy. But the AI boom—or bubble—has funnelled so much money to Nvidia that even such unbelievable scenarios are now becoming reality.
⠀
