← Back
NVIDIA
NVIDIA unveils Vera CPU with 88 Olympus cores for AI workloads; claims 50% faster sandbox performance
· releaseplatformperformancefeature · developer.nvidia.com ↗

NVIDIA Vera CPU: Purpose-Built for AI Infrastructure

NVIDIA introduced the Vera CPU, a custom processor architecture designed to address the emerging bottleneck in modern AI systems: CPU-bound serial tasks within agentic loops. As reasoning models increase token demand and complex evaluation tasks, the CPU becomes a critical constraint alongside accelerators—a challenge exemplified by Amdahl's law in distributed AI workloads.

Key Architecture and Specifications

The Vera CPU features:

  • 88 custom Olympus cores with NVIDIA Spatial Multithreading (SMT) for task concurrency
  • 1.2 TB/s memory bandwidth with uniform 14 GB/s per core allocation
  • Monolithic die design with adjacent dielets connected via NVIDIA Scalable Coherency Fabric
  • LPDDR5X SOCAMM modules for sustained high instruction-per-cycle (IPC) and deterministic latency

The architecture delivers up to 50% faster agentic sandbox performance compared to competitive x86 platforms and achieves 2x performance-per-watt efficiency. Data centers can achieve 4x sandbox density using Vera-based infrastructure.

Optimized for AI Workloads

The Vera CPU targets two specific use cases:

Reinforcement Learning Post-Training: Models generate code on accelerators, which is compiled, tested, and evaluated on CPUs in feedback-reward loops. Each CPU core executes lightly-threaded sandboxes end-to-end—requiring both extreme single-core performance and massive concurrent throughput.

Agentic Inference: AI agents require CPUs to orchestrate tool usage (web browsers, databases, code interpreters) while moving large volumes of context data efficiently. Deterministic latency and consistent SLAs are critical for real-time performance.

Deployment and Availability

NVIDIA offers multiple Vera platform configurations:

  • Vera Rubin NVL72 racks with tight CPU-GPU coupling for integrated workloads
  • Liquid-cooled CPU-only racks for standalone agentic compute
  • Single/dual-socket server options for flexible deployment models

Commercial availability from major OEMs is expected in the second half of 2026.