← Back
NVIDIA
NVIDIA debuts Vera CPU for AI training and inference; delivers 50% faster sandbox performance than competitors
· releaseplatformperformance · developer.nvidia.com ↗

NVIDIA Vera CPU Architecture

NVIDIA has unveiled the Vera CPU, a purpose-built processor designed to address a critical bottleneck in modern AI infrastructure: CPU performance during reinforcement learning post-training and agentic inference workloads. The Vera CPU features 88 custom Olympus cores with NVIDIA Spatial Multithreading (SMT), delivering up to 50% faster agentic sandbox performance compared to competitive x86-based systems.

Key Technical Specifications

The Vera CPU is engineered with several distinguishing features:

  • Extreme single-core performance: Optimized for fast execution of complex code running in individual sandboxed environments
  • 1.2 TB/s memory bandwidth with uniform 14 GB/s per-core bandwidth, enabling efficient context switching and real-time analytics
  • Monolithic die design with adjacent dielets connected via NVIDIA's second-generation Scalable Coherency Fabric for deterministic latency
  • LPDDR5X SOCAMM modules for flexible memory configurations
  • 4x greater sandbox density and 2x performance-per-watt efficiency versus x86-based racks

Why Vera Matters for AI Factories

Modern AI workloads increasingly require tight CPU-GPU coordination. During RL-based post-training, GPUs generate code tokens while CPUs run sandboxed environments (compilers, runtime tests, build systems) that feed results back into training loops. Vera addresses this by combining high single-threaded performance (needed for sequential tasks like code compilation) with massive throughput (needed to orchestrate thousands of concurrent sandbox environments). This solves a classic Amdahl's law bottleneck where CPU-bound serial tasks gate overall system performance.

Deployment Options and Availability

NVIDIA will offer Vera in multiple platform configurations: tightly coupled Vera Rubin NVL72 racks, standalone liquid-cooled CPU racks, and flexible single/dual-socket server options. Commercial availability from major OEMs is expected in the second half of 2026.