← Back
NVIDIA
NVIDIA debuts Vera CPU with 88 cores; claims 50% faster AI sandbox performance for post-training
· releasefeatureplatformperformance · developer.nvidia.com ↗

The Vera CPU: Purpose-Built for AI Infrastructure

NVIDIA has unveiled the Vera CPU, a new processor architecture engineered to address the specific demands of modern AI infrastructure. The chip is designed to solve a critical bottleneck in AI factories: while GPUs are optimized for parallel token generation and model training, CPUs have become the limiting factor in reinforcement learning (RL) post-training and agentic inference workflows, where CPU-bound serial tasks gate overall system performance.

Key Architecture and Performance Features

The Vera CPU delivers several technical improvements tailored for AI workloads:

  • 88 Custom Olympus Cores: High-performance cores with NVIDIA Spatial Multithreading (SMT) to handle thousands of concurrent tasks required by modern AI agents and RL evaluation loops.
  • Massive Memory Bandwidth: 1.2 TB/s total bandwidth with 14 GB/s per core, supporting deterministic latency and sustained performance under constant load.
  • Monolithic Die Design: Adjacent dielets connected via NVIDIA's second-generation Scalable Coherency Fabric (SCF) for efficient, low-latency communication.
  • Performance Gains: Up to 50% faster sandbox performance over existing x86-based systems, 4x higher sandbox density per rack, and 2x better performance-per-watt.

Real-World Workload Optimization

The Vera CPU addresses two critical AI workload patterns. In RL post-training, models generate code on accelerators while CPUs compile, execute, and test it in sandboxed environments—creating a tight feedback loop. Agentic inference similarly requires CPUs to manage tool execution (code interpreters, databases, web browsers) while GPUs run inference. The Vera CPU's design prioritizes single-core performance for executing individual sandbox tasks quickly and consistently, combined with the throughput capacity to run thousands concurrently.

Deployment and Availability

NVIDIA is offering multiple deployment options, including the Vera Rubin NVL72 rack (tightly coupled to accelerators), standalone CPU racks with liquid cooling, and flexible single/dual-socket server configurations. Commercial availability through major OEMs is expected in the second half of 2026.