# NVIDIA pitches Vera for AI agents, but benchmark baselines matter

> Vera agent benchmarks use different x86 baselines.

By BIGCHANGE.AI Editorial

Published: 2026-11-06T01:25:38.770Z
Updated: 2026-11-06T01:25:38.770Z
Canonical: https://bigchange.ai/blog/nvidia-vera-cpu-ai-agent-benchmarks

![Conceptual illustration of a server and processor board.](https://bigchange.ai/api/media/file/vera-concept-original.png)
AI-generated conceptual illustration by BIG CHANGE; no actual NVIDIA hardware or test is depicted.

NVIDIA says its Vera CPU can support more AI-agent work per watt than competing x86 processors, citing partner tests and a 100-megawatt comparison.

## The big change

- Agents repeatedly move between model inference and host-side tools, so the CPU can affect how long a task takes even when a GPU runs the model.
- A partner’s measured comparison shows why naming the x86 baseline matters: its largest quoted speed ratio is not against its fastest tested x86 processor.
- NVIDIA’s energy comparison lacks enough published detail to reproduce.

## Why a CPU matters to an AI agent

An agent can call a model, then search files, run code or ask a tool for information before making another model call. The host processor handles orchestration and tool execution around those calls. In an [independent study](https://arxiv.org/abs/2608.04458), researchers examined production Azure workloads and open-source frameworks and found that this repeated CPU–GPU handoff produces fragmented, bursty resource demand.

A slow tool step can delay the next stage even if model inference is fast. Usefulness depends on concurrency and acceptable latency; a single unloaded task is only part of the picture.

## NVIDIA’s comparison and its missing details

In its [November 5 post](https://blogs.nvidia.com/blog/vera-cpu-highest-throughput-fastest-agents/), NVIDIA claims 1.6 times as many agents per watt versus its fastest x86 comparator, and twice the agent speed versus its most energy-efficient x86 comparator. The baselines differ.

It does not identify those CPUs, workload or power-accounting method, or say whether the facility-scale comparison was measured or projected.

## The 2.2-times baseline

NVIDIA calls DeepInfra’s result 2.2 times the best x86 speed. But [DeepInfra’s own report](https://deepinfra.com/blog/nvidia-vera-cpu-benchmark) gives 29 milliseconds for Vera versus 64 milliseconds for Intel Sapphire Rapids—about 2.2 times by latency—while its fastest tested x86, AMD Zen 5 Turin, took 35 milliseconds, making Vera about 1.2 times faster.

DeepInfra says it replayed captured production-agent traffic at fixed pacing without generating live model tokens. Each chip used 20 physical cores, one thread per core and a single NUMA node. The 29/64/35-millisecond figures are for near-bare-metal orchestration; its isolated microVM results are different. The report tests host-side work under a controlled configuration, not an entire live model interaction.

## Starting and sharing sandboxes

[Daytona](https://www.daytona.io/dotfiles/when-2000-agents-share-one-cpu) tested the same coding workload across its sandboxes. It reports 1,000 environments ready in 10.8 seconds on Vera, versus 32.3 seconds on AMD EPYC 9575F and 88.4 seconds on EPYC 9755. Startup time is distinct from sustained job throughput; the reported 12.7 jobs per second at 2,000 concurrent Vera sandboxes measures a different stage.

[Modal](https://modal.com/blog/nvidia-vera-sandboxes) built a test from 50 Terminal-Bench task solutions and verifiers, with network calls removed and containers prewarmed. It reports up to twice the task throughput of an equivalent x86 Granite Rapids machine under load. Modal also says it would need infrastructure and image-build changes to support Arm after operating x86 systems.

For a provider, migration and operating costs matter alongside throughput at the concurrency and latency it needs. BIG CHANGE did not test the hardware.

## Sources

- [NVIDIA Vera comparison](https://blogs.nvidia.com/blog/vera-cpu-highest-throughput-fastest-agents/) — November 5 vendor claims and partner summary.
- [DeepInfra Vera benchmark](https://deepinfra.com/blog/nvidia-vera-cpu-benchmark) — Partner methodology and measured baselines.
- [Modal sandbox benchmark](https://modal.com/blog/nvidia-vera-sandboxes) — Partner workload and migration conditions.
- [Daytona sandbox benchmark](https://www.daytona.io/dotfiles/when-2000-agents-share-one-cpu) — Partner startup and throughput measurements.
- [Agentic workflow architecture study](https://arxiv.org/abs/2608.04458) — Independent research on CPU and GPU orchestration.