Nvidia says it’s ‘absurd’ to assume it predicted the agentic moment during Vera’s development – ‘product intuition’ help

AI & Compute

Nvidia Rejects 'Crystal Ball' Claim Behind Vera CPU's Agentic Rise

Nvidia says it never foresaw the agentic wave when Vera development began four years ago. Design intuition and accelerator co-design shaped the CPU, says data center CPU chief Ian Finder.

By
Tom Whitfield
Filed
Channel
AI & Compute
Read
4 min read

Nvidia began work on its Vera CPU roughly four years before the agentic computing wave arrived, and the company says it never predicted that shift. Ian Finder, head of Nvidia's data center CPU business, called the assumption that Nvidia saw the agentic moment coming "absurd" in a conversation with Tom's Hardware Premium at SC26.

"Am I going to sit here and tell you that we have this magic crystal ball around the agentic moment? That would be absurd, right?" Finder said. "But what we do have as a company, and I think this has been proven out in our track record, is [that] we have design intuition."

The remarks respond to a criticism that has circulated since Nvidia introduced Vera, which is often marketed as an "agentic CPU." Critics argue that neither Nvidia nor, to a lesser degree, Arm with its AGI CPU could have designed a chip four years ago for workloads that only emerged at scale this year.

What actually drove Vera's design?

Finder points to co-design with Nvidia's accelerators, a practice the team has carried forward since its first Grace data center CPU. The starting point, he said, is conceptually simple: what kind of CPU keeps the accelerator fed and happy? Vera, like Grace before it, was built on the assumption that data centers and AI workloads are accelerated — an assumption that largely holds.

The word Finder returned to repeatedly was "opinionated." Grace introduced an opinionated memory system built on Nvidia's Scalable Coherency Fabric. Vera extends that memory system and adds an opinionated microarchitecture, the Olympus core. In plain terms, Nvidia made deliberate design choices rather than trying to serve every customer's requirements.

Finder contrasted this with the "legacy constraints" of Intel and AMD, who must answer to large pools of customers with specific needs. Nvidia's clean-sheet approach let the team reason from first principles about what a CPU beside a GPU should do.

"If we have silicon that needs really, really great single-thread performance, [and] we also need concurrency, then we need to get there with IPC instead of frequency," Finder said. "We need to make sure we have good loaded memory latency. We need to make sure we have a lot of bandwidth, and you start to design something on a clean sheet of paper that's not encumbered by dollar-per-core economics."

Did Nvidia see agents coming at all?

Not explicitly, according to Finder — but the early signals were visible. He said Nvidia was "starting to see the beginnings of tool calling" when the team put pen to paper on Vera four years ago, and that the company knew what it needed from a head node and understood the overall shape of the problem.

Independent benchmarking adds context. Tom's Hardware's Vera CPU testing shows the chip does not excel at every workload, and recent AMD EPYC Venice benchmarking shows AMD reclaiming top spots in several tests. Finder's counterargument is that superior memory bandwidth will matter far more as agentic systems evolve and make fuller use of available compute.

Why the agentic wave lifted CPU demand

When the agentic wave hit earlier this year, some observers framed it as a shift away from GPUs toward CPUs. That reading is wrong. Agentic loops often cannot fully occupy GPUs in a GPU-centric environment, and GPUs are frequently oversubscribed on host nodes. The response is not to buy fewer GPUs; it is to buy enough CPUs to properly leverage the GPU compute already installed.

Finder illustrated the pattern with a coding agent. After a prompt, the model makes several small tool calls to read and locate files, passing results back in between. Those calls are tiny; the vast majority of time is spent in the model on the GPU. The agent then compiles and tests the project, a phase far more bound by the CPU. The GPU is not idle then either — it runs other work. The same packing logic, Finder said, will reach the CPU.

"The idea that we're just going to say, 'OK, the CPU can just chill out here, and we're just going to wait,' that's not going to happen," Finder said. "What's going to happen is we're going to pack multiple environments together, and we're going to subscribe the socket."

He described the challenge of spinning up agent sandboxes within milliseconds — a "very interesting computer science problem" — that startups such as Daytona and E2B are already investigating. Serving more agents on a socket will naturally pressure the memory system, which is exactly where Vera holds its edge.

"We built a product that does really, really well without such optimizations, but as a design point that really will lean into them when the industry starts to catch up to this," Finder said.

What comes after Vera

Vera entered this market as the successor to Grace and has ridden the agentic wave into a much larger role in the CPU business. Nvidia has already confirmed a follow-up: the Rosa CPU, built on the new Rigel microarchitecture. Whether Vera's opinionated design proves correct will depend on how quickly the industry builds workloads that consume its bandwidth advantage.

Source: Tom's Hardware

Share this article:

More from Tom Whitfield

Tom Whitfield

Show full bio

Staff writer covering companies, deals and strategy in Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.

1,077 articles

Related articles

« Previous article