New Purpose-Built CPU for AI Workloads
NVIDIA announced the Vera CPU, a custom processor engineered to address a critical bottleneck in modern AI systems. While GPUs have dominated token generation and training, CPU-bound serial tasks in agentic loops—where reinforcement learning models must evaluate outputs and agents coordinate complex tool usage—have become performance limiters. Vera is designed to tackle this problem head-on.
Key Technical Features
The Vera CPU incorporates several innovations:
- 88 Custom Olympus Cores: Purpose-built cores optimized for sustained single-threaded performance required by individual sandbox environments
- 1.2 TB/s Memory Bandwidth: Ensures consistent SLAs under high concurrency with efficient data movement for real-time analysis and context switching
- NVIDIA Spatial Multithreading (SMT): Enables task-level concurrency across thousands of concurrent environments
- Second-Generation Scalable Coherency Fabric: Supports deterministic latency and rack-scale deployment
- Monolithic Die Design: 14 GB/s memory bandwidth per core with LPDDR5X SOCAMM modules for predictable performance
Performance Metrics & Deployment Options
Vera delivers up to 50% faster sandbox performance compared to competitive x86-based platforms, with 4x greater sandbox density and 2x performance-per-watt efficiency. The platform supports flexible deployment through tightly coupled Vera Rubin NVL72 racks, liquid-cooled CPU-only racks, and single/dual-socket server configurations.
Commercial availability is expected in H2 2026 through major OEM partners, positioning Vera as part of NVIDIA's broader Rubin AI infrastructure platform.