NVIDIA has unveiled more details about its custom Armv9.2 "Vera" CPU, engineered for NVL rackscale systems and independent enterprise deployment. The "Vera" CPU is optimized for data analysis, single-threaded performance, agentic AI, and large-scale software operations. Its "Olympus" core design features a 10-wide decode engine, a neural branch predictor, and a mid-core section for uncovering independent work and enabling deep out-of-order execution.

The execution engine efficiently distributes scalar, vector, floating-point, cryptographic, and load/store microoperations with low latency. The cache subsystem boasts a deep hierarchy, multiple prefetch engines, and dedicated graph data prefetchers to handle demanding agentic AI workloads requiring high throughput for sparse data retrieval. NVIDIA's simultaneous multithreading partitions resources to prevent the "noisy neighbor" effect common in traditional CPUs.

Connecting the system is the Scalable Coherency Fabric with 3.4 TB/s bandwidth and a 164 MB unified L3 cache. External connections support up to 1.5 TB of LPDDR5X memory per CPU at 1.2 TB/s aggregate bandwidth. Preliminary comparisons suggest up to 1.8x performance boost over AMD EPYC "Turin" in select workloads.