Nvidia unveils the Grace CPU Superchip, an Arm-based discrete data center CPU coming next year, with 144 high-performance cores and 1TB of secondary memory
Nvidia offered details on its Grace central processing unit (CPU) “Superchip” during CEO Jensen Huang's keynote speech at its virtual Nvidia GTC 2022 event.
Context & Ripple Effects
Grace first surfaced a year ago as an Arm-based server CPU aimed at large-scale neural network workloads; today's GTC keynote turns that announcement into concrete silicon specs — 144 high-performance cores and 1TB of secondary memory on a discrete Superchip shipping next year.
The timing matters because Arm has since positioned Neoverse V2 explicitly for cloud, hyperscale, and HPC workloads, with those cores set to power Grace — making this the most concrete test yet of Arm's push out of mobile and into the data center.
First-order effects
- Hyperscale buyers running large-scale neural network workloads gain a discrete Arm CPU option purpose-built for their pipelines, rather than waiting for Grace only as a component inside Nvidia systems.
- Nvidia extends its data center franchise beyond GPUs, giving Jensen Huang's company a CPU it controls end-to-end for the first time.
Second-order effects
- Arm's Neoverse V2 gets a flagship hyperscale customer in Grace, strengthening its case against incumbent x86 server CPUs just as cloud operators weigh second sources.
- A discrete Nvidia CPU pressures the traditional separation between server-CPU vendors and accelerator vendors, forcing each side to consider bundling the other's job.
Third-order effects
- If the pattern holds, CPUs and GPUs converge into single-package superchips — the later GH200 Grace Hopper pairing shows exactly where this discrete CPU was headed.
- The longer arc runs from licensed Arm cores toward in-house design: by 2026 Nvidia ships Vera, its first CPU built on fully custom cores, suggesting Grace was the bridge product in that transition.
The trend: Data center compute is consolidating around tightly coupled CPU-GPU superchips, with Nvidia using licensed Arm cores as a stepping stone to fully custom silicon.