Nvidia has unveiled its groundbreaking Vera CPU a significant departure from traditional datacenter design philosophies. This new processor, built around the Olympus core marks Nvidia’s first custom CPU core designed specifically for datacenter applications since the Tegra era. The Vera CPU is not just an attachment to Nvidia’s existing product line but a formidable contender in the datacenter processing arena.
The Vera CPU is engineered to address the evolving demands of agentic AI where AI agents perform complex tasks by executing code, invoking tools, and interacting with databases. This shift requires a different approach to CPU design, prioritizing single-thread performancememory bandwidth and latency predictability under heavy loads.
Architectural Innovations of the Vera CPU
The Vera CPU features 88 Olympus cores on a monolithic compute die, supporting 176 threads through a unique partitioning scheme called Spatial Multithreading. This design choice ensures efficient resource utilization and minimizes latency. The monolithic compute die is connected by a second-generation scalable coherency fabric, delivering approximately 3.4 terabytes per second of bisection bandwidth.
The memory subsystem is another standout feature, utilizing LPDDR5X memory hardened for datacenter use. This subsystem provides up to 1.2 TB/s of bandwidth, offering roughly 3x the memory bandwidth per core and about 5x the bandwidth per watt compared to conventional DDR-based server designs. The Vera CPU also boasts 40% lower memory latency under load, enhancing
Performance and Real-World Applications
Nvidia claims that the Olympus core delivers roughly 2x faster performance compared to previous generations. The Vera CPU is available in two configurations: a dense liquid-cooled rack packing 256 CPUs and more than 22,000 cores, and a conventional air-cooled 2U with two sockets. Dell has committed to integrating the Vera CPU into its PowerEdge systems, highlighting its potential in the market.
Early customer data shows promising results. For instance, Perplexity reported completing coding sandboxes 1.5x faster on Vera CPUs compared to their production Xeon fleet. The New York Stock Exchange tested Vera with the Redpanda streaming engine and measured 6x lower p99 latency versus Epyc Turin. Los Alamos National Laboratory saw significant improvements in agentic workloads and radiation transport simulations.
Accelerating Chip Design with Vera CPUs
Nvidia is also leveraging its Vera CPUs to accelerate chip design processes. By partnering with Cadence Systems and Synopsys, Nvidia has optimized electronic design automation (EDA) software to run on Vera CPUs. Early tests showed a 1.5x performance improvement for Cadence’s Jasper verification platform and Synopsys VCS logical simulation tool.
Nvidia is integrating its PhysicsNeMo physics-AI libraries and GPU math libraries into the Nvidia Agent Toolkit. This integration allows AI agents to call accelerated solvers, enhancing the automation of chip design processes. The new libraries support iterative sparse solvers on GPUs, enabling faster simulations for fluid flow, structural stress, and electromagnetics.
Keysight Technologies reported a 10x speedup in electromagnetic simulations using the new cuDSS libraries. Silvaco Group achieved a 3.2 billion-mesh-node photonic edge coupler simulation in under four hours on a cluster of 32 GPUs. These advancements underscore the potential of Vera CPUs in accelerating complex design workflows.
