Skip to content
Theme
OnDevice Tools

GadgetsLaptops & PCs

Nvidia Vera CPU design was not aimed at agentic AI

Nvidia has rejected AMD's criticism over its Vera processor, saying the chip was shaped by design choices rather than a forecast of AI agents.

Nvidia Vera data centre CPU illustration showing memory bandwidth and AI agent workloads

Key takeaways

  • Nvidia rejected AMD's claim that Vera was designed specifically for agentic AI.
  • Ian Finder said Vera's design came from co-design with Nvidia accelerators and clean-sheet choices.
  • Finder expects memory bandwidth to matter more as AI agents use more tools and services.
  • Nvidia has announced the Rosa CPU with the Rigel microarchitecture as a follow-up to Vera.
In this article
  1. Nvidia responds to AMD criticism
  2. How Vera was designed
  3. Why memory bandwidth matters for AI agents
  4. Benchmarks and what comes next

Nvidia responds to AMD criticism

Nvidia Vera CPU design has become the subject of a public dispute over whether the processor was built specifically for agentic AI. In this context, agentic AI refers to AI agents that can call tools and perform tasks. According to Tom's Hardware, AMD has criticised Nvidia over marketing claims that describe its CPUs as suited to AI agents, while Arm has also faced some criticism. Nvidia rejected that criticism during a conversation with Tom's Hardware Premium at the SC26 event.

Ian Finder, who leads Nvidia's data centre CPU business, said it would be absurd to suggest that Nvidia foresaw the current agentic wave when work on Vera started four years ago. He acknowledged that the team had begun to see early signs of tool calling, but he would not claim that the full agentic moment was obvious. Instead, he described Vera as the result of design principles, product intuition and established computer science ideas.

How Vera was designed

Finder said Vera was shaped by co-design with Nvidia's accelerators. The approach began with the Grace data centre CPU and continued with Vera. In simple terms, the team asked what kind of processor would keep an accelerator busy and supplied with useful work.

He repeatedly described elements of the design as opinionated. Grace introduced a memory system built around Nvidia's Scalable Coherency Fabric. Vera extends that memory approach and adds Nvidia's Olympus core microarchitecture, which is the basic internal design of the processor's cores.

Finder contrasted this with what he described as legacy constraints at established CPU makers. Nvidia, he said, could start from scratch. That allowed the team to focus on strong single-thread performance, or how well it handles one task at a time, and concurrency, or handling many tasks at once. It also targeted loaded memory latency, which is how quickly the processor can access data under load, and high bandwidth, rather than being limited by conventional cost-per-core economics.

Why memory bandwidth matters for AI agents

Illustration: AI, privacy, fiber network, graphics card and the labels AI Agents, AMD, NVIDIA

A central part of Finder's argument is that memory bandwidth will matter more as AI agents become more capable. He used a coding agent as an example. After a user gives a prompt, the model may make tiny requests to open files or find code. The model then spends most of its time running on a GPU, which is a processor designed for parallel graphics and AI work.

The workload can then shift to compiling and testing the project, which places more demand on the CPU. Finder said the GPU is not idle during that period because it can be used for another task. He suggested that a similar pattern may emerge on the CPU as agents perform more work.

This creates a need to start and stop isolated agent environments quickly. Tom's Hardware notes that startups such as Daytona and E2B are working on bringing these sandboxes up and shutting them down in milliseconds. Serving more agents would put pressure on the memory system, an area where Nvidia says Vera has an advantage.

Benchmarks and what comes next

Coverage from Tom's Hardware indicates that Vera does not lead every workload. It also says AMD's EPYC Venice has reclaimed some top benchmark positions. That makes the debate about Vera's purpose more than a marketing argument. The cited benchmarks show that performance varies by workload.

Nvidia has already announced that it will follow Vera with the Rosa CPU, which will use the Rigel microarchitecture. Finder said Vera was designed to perform well even before the industry fully adapts to these agentic workloads, but that it will make greater use of them as software improves.

The exchange shows how AI agents are influencing data centre CPU design. Instead of focusing only on clock frequency, Nvidia is emphasising memory systems, accelerator pairing and the ability to handle many small, tool-driven tasks.

Sources

  1. Nvidia says it’s ‘absurd’ to assume it predicted the agentic moment during Vera’s development – ‘product intuition’ helped drive CPU design, superior memory bandwidth will play critical role in future AI agents, CPU boss claimsTom's Hardware

Tags

Spotted a mistake? See our corrections policy. Read how News by OnDevice works in our editorial policy.