Nvidia’s Grace CPU
Another step towards ARM ending the x86 era?
When Nvidia announced it was buying ARM, the intent was clear: Nvidia wanted a way to compete in the CPU space. That’s set to happen in 2023 with the arrival of the Grace CPUs. It’s still a way off, but the specs already look impressive. The ARM acquisition paves the way for Nvidia to attempt to do in the CPU space what it’s been doing with GPUs for over 20 years.
A decade ago, we first heard about Project Denver, a CPU architecture that would use binary translation to provide x86 compatibility. Intel blocked the x86 compatible aspect, but the chip did see the light of day in a few products, such as the Nexus 9 tablet. With the rise of ARM processors that power nearly all phones and tablets, plus some laptops and even servers, x86 compatibility is a must-have no more.
Grace focuses on Nvidia’s strong points: AI and deep learning. The company claims it delivers 10 times more performance than the fastest servers, at least in AI and HPC workloads. That will almost certainly come via vector extensions that boost performance in matrix computations, but Grace will also boost higher bandwidth and efficiency, thanks to LPDDR5x, and promises some of the fastest interconnects between CPUs and GPUs via NVLink: over 900GB/s cache coherent CPU-toGPU NVLink, and more than 600GB/s CPU-to-CPU.
Nvidia hasn’t talked core counts or clock frequencies yet, and while the rendered image suggests it might have 96 CPU cores, it’s likely just a placeholder. Prior to the Nvidia acquisition, ARM had already disclosed its future high-performance Zeus CPU cores, which likely play a role in Grace given the time frame, though Nvidia now calls these ARM Neoverse cores. The ARM V1 platform includes support for all the latest whiz-bang tech: PCIe Gen5, DDR5, HBM2e or HMB3, and a CCIX 1.1 interconnect. Swap out CCIX for NVLink and link the CPU up with an Ampere Next GPU, and you have a potent starting point for a supercomputer node that can scale up to thousands of nodes.
Incidentally, the CPU is named after Grace Hopper, one of the
“Nvidia hasn’t talked core counts or clock frequencies yet, and while the rendered image suggests it might have 96 CPU cores, it’s likely just a placeholder.”
pioneers of computer programming. Current indications are that Nvidia’s post-Ampere GPU architecture will be code-named Hopper, so it will combine CPUs and GPUs to give double honours.
Perhaps more interesting is where else we might see the Grace Hopper combo show up. It’s a safe bet that Nvidia’s plans don’t end with supercomputers. Look at the laptop world and Nvidia already discusses that as an almost entirely separate market from desktop gaming PCs. We’ve also seen a few laptops testing the Windows on ARM possibilities – and, of course, there’s Apple’s M1 that eschews Windows and x86 in favour of an all-Apple design.
Officially, there’s no word of consumer variants of Grace or Hopper yet, but we know they’re coming. Nvidia still has its roots in the gaming market, and building a better gaming laptop – one that can play games untethered for many hours – would be impressive. Getting game developers to jump ship from x86 to ARM would have been practically unthinkable five years back, but with Nvidia, Apple, and others involved, it becomes less a matter of “if” and more “when” we’ll see high-end gaming laptops running allNvidia processors.