Nvidia unveiled its Vera CPU, a standalone AI‑centric processor featuring 88 custom ARMv9.2 cores and 1.5 TB of memory, targeting hyperscalers and cloud providers.

Nvidia’s latest breakthrough, the Vera CPU, marks a decisive shift toward a standalone, AI‑centric processor architecture that promises to reshape hyperscale and cloud computing workloads.

What Makes Vera Different

Unlike Nvidia’s previous GPU‑focused offerings, Vera is built around 88 custom‑designed ARMv9.2 cores, collectively dubbed the Olympus family, which are tightly integrated with a massive 1.5 TB of on‑chip memory.

The Olympus cores are engineered for low‑latency tensor operations, featuring dedicated matrix‑multiply units and a unified cache hierarchy that reduces data movement—a critical factor for large‑scale AI inference and training.

Architecture Highlights

  • 88 custom ARMv9.2 cores, each with extended SIMD extensions for AI workloads
  • 1.5 TB of high‑bandwidth memory (HBM) directly attached to the die
  • Integrated tensor accelerators that share the same instruction set as the cores
  • A unified cache system that delivers up to 2 TB/s memory bandwidth

The design also incorporates a novel interconnect fabric that allows the cores to communicate with the tensor units and memory subsystem without the overhead typical of traditional CPU‑GPU pairings.

Target Markets and Use Cases

Nvidia is positioning Vera for hyperscalers and cloud providers that need to run massive AI models at scale while maintaining energy efficiency. The processor’s ability to handle both general‑purpose compute and AI‑specific tasks makes it a compelling alternative to deploying separate CPU and GPU clusters.

Potential deployments include large language model serving, real‑time recommendation engines, and high‑throughput video analytics, where the reduced data‑transfer latency can translate into measurable cost savings.

Vera represents a new class of processor that blurs the line between traditional CPUs and AI accelerators, delivering unprecedented performance per watt for cloud workloads.

While detailed performance benchmarks are still forthcoming, early internal tests suggest that Vera can outperform comparable x86‑based AI servers by a significant margin in both throughput and power consumption.

The processor will be available to select partners later this year, with broader commercial rollout expected in early 2027.

The Register’s deep dive on Nvidia’s Vera CPU and the Olympus cores