NVIDIA Claims Vera CPU Doubles Agent Speed of AMD's 256-Core EPYC Venice

NVIDIA has published a set of CPU benchmarks built around AI agents, claiming that its Vera CPU runs agent workloads twice as fast as AMD's 256-core EPYC Venice. It is the company's most direct attempt yet to argue that the CPU, not only the GPU, decides how quickly an agent fleet gets work done.

The argument rests on how agents actually behave. An agent does not simply generate an answer: it searches for information, queries databases, writes and tests code, and takes actions in external systems, bouncing between GPU reasoning and CPU execution at every step. As businesses put more agents into production, the CPU becomes the queue every task has to pass through.

NVIDIA frames the contest on two axes rather than one. Beyond raw speed, it is selling capacity per watt, meaning how many agents a data center can keep running inside a fixed power budget. In a 100-megawatt comparison, the company says Vera delivered twice the agent speed of the fastest x86 processor and 1.6 times more agents per watt than the most energy-efficient x86 part.

Those figures are vendor-run, and the workloads, baselines and 100-megawatt framing were all chosen by NVIDIA. How much of the gap survives third-party replication is still an open question.

NVIDIA also lined up early customers to argue the gains show up outside a lab. Perplexity says it ran agentic coding tasks 1.5 times faster on Vera and plans to deploy the chip in an upcoming production system, while sandbox and database vendors including Daytona, Modal and ClickHouse report faster parallel job starts and near-linear scaling across 176 cores.

The timing matters because AMD's EPYC Venice is built on the opposite bet: 256 Zen 6 cores aimed at maximum throughput, backed by AMD benchmarks that put Venice ahead. The two vendors are effectively describing different customers, one optimizing the latency of a single agent chain and the other the total volume of parallel work.

When power is the binding constraint, the question buyers actually ask has shifted from how many cores a CPU has to how many agents a single watt can keep busy. That reframing favors whoever can prove the metric on real workloads rather than vendor slides.

What to watch next is delivery and independent testing: whether Vera ships in the volumes NVIDIA's partners describe, how the first production deployments perform under mixed agent traffic, and how Venice answers once it is in customers' hands.