NVIDIA formally took the wraps off of Vera, its second-generation custom CPU based on the Arm microarchitecture and engineered to power AI and high performance computing (HPC).

After decades of hyping GPUs and downplaying the role of the CPU, NVIDIA is firmly in the CPU business. However, its CPU business is designed in conjunction with its GPU business and the two are tightly paired.

Right off the bat, it’s clear where Vera differentiates itself from Grace. Where Grace used 72 Neoverse cores designed by Arm, Vera comes with 88 custom cores designed entirely by NVIDIA code-named Olympus.

Olympus features high single-thread performance, optimized for agentic AI workloads involving orchestration, scripting languages, databases, compilers, and runtime environments.

It also features wide out-of-order architecture with deep instruction scheduling, advanced prefetching, and a neural branch predictor, spatial multithreading capable of running two threads per core by partitioning hardware resources, and support high memory bandwidth.

Vera features deep integration with NVIDIA’s accelerated computing stack. The architecture leverages high-bandwidth connections that allow CPUs and GPUs to share memory and exchange data with significantly lower latency.

Over the course of the last two decades, NVIDIA has evolved from a GPU provider to a full-stack compute and networking provider, offering CPU, GPU, networking, and software to run the gamut of AI and HPC computing.

Vera is expected to play a central role in future NVIDIA rack-scale platforms, replacing x86 CPUs, where thousands of CPUs and GPUs operate as a unified computing system and interconnected by NVIDIA’s high speed interconnects.

Whereas the Grace processor was paired with Hopper and later Blackwell-era GPUs, Vera will ship either paired with a GPU, code named Rubin, but it may also be shipped as a standalone CPU without a GPU. This will be in desktop and edge devices, new markets for NVIDIA. Up to now, NVIDIA has stuck strictly to data center deployments. But with Vera, it is branching out into end points, so you will be seeing Viera processors in everything from laptops to servers.

By developing custom CPUs alongside GPUs, networking hardware, and software platforms, the company can optimize end-to-end system-level performance. It also helps in NVIDIA reduce dependence on third-party silicon suppliers.

Although Vera is primarily targeted at servers, and NVIDIA is also targeting the desktop. In his keynote speech introducing the Vera platform, CEO Jensen Huang described Vera as being the CPU to power AI agents on the desktop and the server. He referred to the desktop as “a personal agent box.”

NVIDIA says Vera systems will be available from system builders and cloud partners starting this fall, with production shipments of Vera Rubin also set to begin in that timeframe.

Originally published by Techstrong.IT. Republished with attribution.