New Vera CPU Architecture for AI Workloads

NVIDIA has announced the Vera CPU, a processor purpose-built for AI infrastructure workloads including reinforcement learning post-training and agentic inference. The chip addresses a critical bottleneck in modern AI systems where single-threaded CPU tasks become the limiting factor—a principle known as Amdahl's law—as GPU throughput continues to increase.

Key Performance Features

The Vera CPU features 88 custom Olympus cores with NVIDIA Spatial Multithreading (SMT) and the second-generation NVIDIA Scalable Coherency Fabric. Key specifications include:

• Up to 50% faster sandbox performance compared to competitive x86-based platforms
• 1.2 TB/s of memory bandwidth for efficient data movement
• 14 GB/s per core of uniform memory bandwidth
• Monolithic die design with adjacent dielets for optimal performance and density

Design Optimizations

The architecture prioritizes three critical requirements for AI workloads: extreme single-core performance for executing complex tasks like code compilation and testing, high memory and fabric bandwidth per core for consistent service-level agreements under heavy load, and efficient rack-scale co-design for data center deployment. The design delivers 4x higher sandbox density and 2x better performance per watt compared to x86-based racks.

Platform and Availability

Commercial deployment options include tightly coupled Vera Rubin NVL72 racks, liquid-cooled CPU racks, and flexible single/dual-socket server configurations. The processors are expected to be available from major OEMs in the second half of 2026.