Purpose-Built for AI Infrastructure
NVIDIA has unveiled the Vera CPU, a custom-designed processor purpose-built for modern AI factory workloads. The chip addresses a critical bottleneck in AI infrastructure: CPU-bound serial tasks that limit throughput despite increasingly powerful GPUs. Vera represents NVIDIA's first fully custom data center CPU core, the Olympus core, marking a significant expansion beyond GPU-centric infrastructure.
Key Specifications and Performance
The Vera CPU features 88 Olympus cores with NVIDIA Spatial Multithreading (SMT) technology and delivers:
- 1.2 TB/s memory bandwidth for efficient data movement across concurrent environments
- Up to 50% faster agentic sandbox performance compared to competitive platforms
- 14 GB/s uniform memory bandwidth per core ensuring deterministic latency
- Monolithic die design with adjacent dielets for tight coupling of compute, cache, and memory
- 4x sandbox density and 2x performance per watt over x86-based racks
Optimized for RL and Agentic Workloads
Vera is specifically engineered for two critical AI workloads: reinforcement learning post-training and agentic inference. In RL scenarios, models generate code on accelerators that must be compiled, tested, and evaluated on CPU sandboxes in tight feedback loops. Delayed evaluation creates wasted tokens and slows model training. Vera's extreme single-core performance and massive memory bandwidth enable tight SLA adherence, ensuring evaluation results complete in time to influence the next training iteration.
Deployment and Availability
NVIDIA will offer multiple platform options including tightly coupled Vera Rubin NVL72 racks, liquid-cooled CPU racks, and flexible single/dual-socket servers. Commercial availability from major OEMs is expected in the second half of 2026, enabling AI factories to rapidly deploy and scale agentic capacity.