Skip to main content

NVIDIA Vera CPU

NVIDIA Vera CPU is NVIDIA's AI-specific CPU announced at GTC 2026, featuring 88 custom Olympus cores, supporting ARMv9.2 instruction set, equipped with up to 1.5TB LPDDR5X memory and 1.2TB/s bandwidth, serving as the host CPU for the Vera Rubin platform, responsible for data movement scheduling, memory management, and system control orchestration.

Key Specifications​

SpecificationValue
CPU ArchitectureARM architecture (Olympus cores)
Instruction SetARMv9.2 (fully compatible)
Core Count88 Olympus cores
Thread Count176 threads (spatial multithreading)
Single-Core Performance2× previous generation
Max Memory Capacity1.5 TB (LPDDR5X)
Memory Bandwidth1.2 TB/s
InterconnectNVLink-C2C (1.8 TB/s)
On-Chip Interconnect2nd-gen NVIDIA SCF (3.4 TB/s bisection bandwidth)
TDPNot disclosed (estimated 350-500W)
Release DateMarch 17, 2026
Mass ProductionSecond half of 2026

Architecture & Specifications​

Vera CPU adopts a monolithic compute chip design, avoiding cross-chiplet communication latency, maintaining stable latency and throughput under all-core load, with predictable performance.

Key Technical Innovations​

  1. 88 Custom Olympus Cores

    • Supports spatial multithreading (176 threads)
    • Single-core performance 2× previous generation
    • Industry-leading energy efficiency

  2. World's First CPU Supporting FP8 Precision

    • Fully compatible with ARMv9.2 instruction set
    • Hardware-level FP8 compute support

  3. 2nd-Gen NVIDIA SCF (Scalable Coherent Fabric)

    • Provides 3.4 TB/s bisection bandwidth
    • On-chip mesh + unified cache
    • Latency-free scaling to 88 cores

  4. NVLink-C2C Interconnect

    • Coherent bandwidth up to 1.8 TB/s
    • Enables seamless data sharing between CPUs and between CPU and GPU
    • Supports unified memory system

  5. Full Confidential Computing

    • Supports hardware-enforced security isolation
    • Protects sensitive data and code

Memory Subsystem​

  • Max Memory Capacity: 1.5 TB (3× previous generation)
  • Memory Type: LPDDR5X
  • Memory Bandwidth: 1.2 TB/s (2× bandwidth, 1/2 power vs traditional CPU)

Companion Platform​

Vera Rubin NVL72​

  • 72× Rubin R200 GPUs
  • 36× Vera CPUs
  • Total Memory: 54 TB LPDDR5X
  • TDP: ~180kW (full liquid cooling required)

HGX Rubin NVL8​

  • 8× Rubin R200 GPUs
  • 2× Vera CPUs
  • For small-to-medium scale AI training and inference

Deployment Formats​

  1. High-Density Liquid-Cooled Vera CPU Rack

    • Based on NVIDIA MGX
    • Supports up to 256 Vera CPUs
    • Supports over 22,500 concurrent environments
    • For AI factory-scale reinforcement learning and agentic AI

  2. Standard Server Configuration

    • Supports dual-socket and single-socket standard configurations
    • Adaptable to general data center needs

  3. Independent CPU Platform

    • Can be used as high-performance independent CPU
    • Supports hyperscale cloud, data analytics, storage, enterprise workloads, HPC

Performance Advantages​

  • Software environment runtime speed: Up to 50% faster than traditional architecture CPUs
  • Efficiency: 2× traditional architecture CPUs
  • RL evaluation cycle: Can be shortened by 50% under full load
  • AI workflow: Seamless collaboration with NVIDIA GPUs, ensuring full-speed AI workflow

Application Scenarios​

Vera CPU is designed for the AI era, suitable for:

  • Reinforcement learning (RL) and agentic AI
  • Data center host CPU (data movement scheduling, memory management, system control orchestration)
  • Hyperscale cloud
  • Data analytics and storage
  • Enterprise workloads and HPC

References​