HeadsUpAI

NVIDIA Launches Vera CPU Purpose-Built for Agentic AI Workloads

· Updated

NVIDIA Vera CPU runs on 88 custom-designed Olympus cores with LPDDR5X memory delivering up to 1.2 TB/s of bandwidth — twice the bandwidth at half the power of general-purpose CPUs. Each core handles two tasks via Spatial Multithreading, targeting the concurrent workloads agentic AI demands. A single rack fits 256 liquid-cooled Vera CPUs sustaining over 22,500 concurrent environments at full performance.

Vera represents a new class of CPU, with Jensen Huang noting the CPU is now driving the model rather than simply supporting it. It pairs with Rubin GPUs via NVLink C2C at 1.8 TB/s — 7x PCIe Gen 6 bandwidth. Cursor, Meta, Alibaba Cloud, CoreWeave, and Oracle Cloud Infrastructure are among early adopters, alongside national labs like Los Alamos and TACC.

Evaluate Vera for agentic inference and reinforcement learning workloads — 256 liquid-cooled CPUs per rack sustaining 22,500+ concurrent environments changes how you'd architect large-scale distributed RL training.

View the full update on nvidianews.nvidia.com
NVIDIA Newsroom
NVIDIA Newsroom
@nvidianewsroom
X

Agentic AI is only as fast as the CPU behind it. #NVIDIAGTC news: Introducing the NVIDIA Vera CPU, the first CPU purpose built for agentic AI and reinforcement learning. Delivering 2x efficiency and 50% faster performance than traditional rack scale CPUs. https://t.co/Dd76FDFlmR

48retweets
View on X

Share this update