NVIDIA Vera: Why Agentic AI Needs Faster Cores, Not More Cores

NVIDIA has detailed the architectural philosophy behind the Vera CPU — the processor at the heart of the Vera Rubin superchip. The central thesis is not core count, but the speed of each individual core: agentic AI loops are inherently sequential, making single-threaded throughput as important as parallelism.

Why Are Agent Loops Different From Classical Workloads?

The conventional approach to scaling CPU performance focuses on adding cores to parallelize independent tasks. An agentic AI workflow operates differently: the agent plans a step, executes it, receives a result, and then plans the next step based on that result. Each step depends on the previous one, meaning the loop cannot be trivially parallelized.

In such a scenario, the speed of a single core directly determines the latency of the entire agent loop. A processor with twice the per-core speed but half the core count can be faster for agentic tasks than a processor with more, slower cores.

Olympus Core Specifications

The Vera CPU is built on Olympus cores that deliver 50% higher IPC compared with NVIDIA Grace cores. The processor supports up to 88 cores and offers memory bandwidth of up to 1.2 TB/s through LPDDR5X memory, with sub-40 W power consumption for the memory subsystem.

Core-to-core bandwidth is 3.4 TB/s, which according to NVIDIA data is 3x more than any other datacenter CPU. This metric is especially relevant for workloads in which cores must exchange data intensively.

In heavy workloads that typically do not favor ARM architectures, Vera achieves 1.8x higher sustained per-core performance compared with x86 processors. NVIDIA attributes this result to the monolithic construction that eliminates the so-called “chiplet tax” — latencies that arise when data crosses between chips within multi-die processors.

Perplexity as the Primary Early Adopter

Perplexity is the primary publicly named customer to have shared concrete performance data. The company tested the Vera CPU on coding agent workflows involving repository cloning and test suite execution.

Results: approximately 1.5x faster job completion compared with an x86 alternative, and up to 1.9x faster concurrent sandbox startup. The latter result is especially relevant for scenarios where multiple agent sessions need to start quickly in parallel, a common requirement in modern coding assistant systems.

Roadmap: Rosa and the Rigel Core

NVIDIA has announced a successor called the Rosa CPU that will use the Rigel core. Rigel brings improved instruction delivery, a larger L2 cache, and more efficient memory management while retaining the same silicon footprint as the Olympus core.

In addition to Perplexity’s data, NVIDIA cites results with other partners: Starburst reports 3x faster SQL analytics on large datasets, while Redpanda achieves up to 6x lower latency on real-time data processing workloads. These results indicate that Olympus cores carry advantages beyond narrowly agentic scenarios.

The Vera CPU is part of NVIDIA’s broader infrastructure strategy to serve growing demand for agentic AI systems. At this point, the announcement functions more as a first-customer validation than a classic product launch — Perplexity provides concrete market evidence for the architectural decisions NVIDIA made in designing the Olympus core.