Nvidia says its Vera CPU is built to stop agentic AI from turning server CPUs into expensive traffic cones.
The chip centres on the Olympus core, which Nvidia claims was designed for the branch-heavy, latency-sensitive code paths used by agents, tools and reinforcement-learning systems.
Those workloads run sandboxes, execute code, call tools, query databases, analyse results and then hand the mess back to a model.
That shifts more of the useful work onto the CPU, where single-thread performance, memory bandwidth per core and predictable latency start to matter. Olympus tries to keep instructions flowing with a neural branch predictor, high-bandwidth instruction fetch and a 10-wide decode engine.
The mid-core design uses deep out-of-order execution, a large reorder buffer, memory renaming and value prediction to chew through dependency chains and pointer-heavy code.
The execution engine then spreads work across integer, branch, vector, floating-point, cryptographic, load and store resources, because agent workloads rarely behave like tidy benchmarks.
Nvidia’s cache subsystem adds multiple hardware prefetch engines and a graph prefetcher to reduce stalls in messy data structures, retrieval indexes and runtime state.
Vera uses Nvidia Spatial Multithreading, which partitions the wide Olympus core across two hardware threads rather than letting them fight like pigeons over decode bandwidth.
With 88 Olympus cores and 176 SMT threads, Vera is meant to keep many agent sandboxes running without turning tail latency into a lottery.
The CPU connects those cores through Nvidia’s Scalable Coherency Fabric, offering up to 3.4TB/s of bisectional bandwidth and a 164MB unified L3 cache.
It uses SOCAMM2 LPDDR5X memory for up to 1.2 TB/s of aggregate bandwidth, or up to 14 GB/s per core, while keeping memory power in check.
Vera can scale into a coherent dual-socket setup using second-generation NVLink-C2C, while presenting one NUMA domain per socket to keep software tuning less miserable.
The platform supports PCIe 6.4, CXL 3.1 and Confidential Computing, including Arm CCA/RME, per-VM encryption keys and authenticated C2C encryption.
Nvidia claims Vera can deliver up to 1.8 times higher performance on agentic workloads than an x86 CPU, based on internal SPEC CPU 2026 results from July 2026, so the independent benchmark lads still have work to do.
