Nvidia is making a CPU. But while it’s notable that the jolly green giant of the semiconductor industry has decided to wade into the data center CPU market, Intel and AMD probably don’t have anything to worry about, at least for now.

Nvidia’s Grace CPU — announced at GTC 2021 and named for Grace Hopper, the U.S. computer programming pioneer — took many by surprise. But it's probably not what you think it is. And the chipmaker’s decision to call this a CPU, while technically correct, doesn’t help clear things up either.

You might be thinking a CPU is a CPU, so clearly, Nvidia is picking a fight with team blue and team red. But the answer is actually no.

Grace isn’t like other CPUs, and that's kind of why Nvidia designed it in the first place. But before I explain why Nvidia is making this thing, let me explain what makes it different.

Traditionally CPUs from Intel, AMD, and even newcomers like Marvell and Ampere are designed with a wide array of use cases in mind. These companies make general compute processors that can be deployed in everything from laptops and workstations to 5G base stations and hyperscale data centers.

Grace is a bit more specialized. Scratch that, it’s a whole lot more specialized. And to confuse things even more, it's not even the CPU cores that actually make it unique. Grace will use Arm’s Neoverse N2 core design, which is just an updated version of what Marvell and Ampere are already using in many of their products.

If you look at Grace’s estimated SPECrate 2017 Integer benchmark scores of 300, it won’t even be that fast. AMD’s 2-year-old EPYC Rome 7742, used in Nvidia’s DGX A100 server platform, is actually faster by that measure. So while a score of 300 is honestly respectable by today’s standards, I somehow doubt that spec will hold up nearly as well in two years.

I'm actually surprised Nvidia even threw this number out there. As far as I can tell, CPU performance, while important, isn’t really the point of this chip. Arguably the more important feature is how Grace will communicate with other peripherals, specifically GPUs.

GPUs typically communicate with CPUs over the PCIe standard. And usually, this works out great. PCIe Gen 4.0 and soon PCIe Gen 5.0 are ridiculously fast — and, I’d argue, 99% of the time complete overkill. That is unless you happen to be training a massive artificial intelligence (AI) model with a trillion or more parameters. In this exotic-sounding scenario, even PCIe Gen 5.0, which when commercially available will deliver upwards of 512 Gb/s of bandwidth to a single device, can become a bottleneck.

And clearly, Nvidia thinks this scenario is going to get a whole lot more common in the coming years. It’s this niche use case that Nvidia created Grace to address. Instead of PCIe, Nvidia is directly coupling the CPU cores and critically the system memory to its GPUs over the company’s proprietary NVLink interconnect technology. And it’s really fast: A little back-of-the-napkin math indicates that each NVLink connection will provide 16 Tb/s of throughput per device. Grace will support up to four to-be-announced GPUs connected over NVLink for an aggregate of 64 Tb/s of total throughput.

Because of this, Grace doesn’t make much sense unless you’re using it in a system with at least four of Nvidia’s top-tier GPUs and even then, only when running AI workloads so large they can’t fit in the GPU’s video memory.

So should Intel or AMD be worried about Grace? Maybe, but only in exceedingly large HPC environments where specialized CPUs from companies like Fujitsu aren’t uncommon.

In a press conference last week, Nvidia CEO Jensen Huang basically said Nvidia wasn’t interested in competing in the CPU space and argued that Grace was a special case.

“As a philosophy, we tend not to do anything unless the world needs us to do it and it doesn’t exist. When you look at the Grace architecture, it’s very unique, and it doesn’t look like anything else out there, and it also solves a problem that didn’t used to exist,” Huang said, responding to questions concerning the company’s ambitions in the CPU market.

Let’s not forget, outside of gaming, Nvidia is essentially unchallenged in the data center GPU market. AMD and Intel both have emerging GPUs in the space, but they are relatively small players by comparison. Neither company has anything that comes close to Nvidia’s A100 series GPUs. By comparison, the CPU market is getting rather crowded these days.

So unless the existing crop of CPU vendors suddenly become deficient in their ability to solve the next great computing problem, Huang is essentially saying he’d rather work with these companies than fight them.

Nvidia is already working with Amazon, Ampere, and Marvell to add support for its GPUs and has long-standing relationships with both Intel and AMD.

And given Nvidia’s pending acquisition of Arm Holdings for $40 billion, this makes a certain kind of sense. The chipmaker has already said it plans to license its graphics and networking intellectual property through Arm. I find it far more likely that we’ll see Nvidia graphics and AI capabilities embedded in CPUs from partner companies than a branded general-purpose CPU, at least in the data center.

The consumer space might be a different story, and it will be interesting to see what happens to Nvidia’s Tegra CPUs used in Nintendo’s Switch gaming consoles — yep while Nvidia already makes CPUs, low power ones anyway — moving forward.

So at least in my mind, until such a time comes that Nvidia feels that the CPU vendors have failed to keep up with customer demand and it can do better, I don't see them going toe to toe with AMD’s EPYC or Intel’s Xeon platforms.

If you disagree, you can hit me up on Twitter @Tobias_SDx.