Vera has barely shipped, and NVIDIA is already teasing its next-gen CPU — three generations in three years, who can keep up with that pace?
It's been a while since we've seen NVIDIA this aggressive on the CPU front.
Right as Vera just entered mass production and the first batches landed in the hands of Anthropic and OpenAI, NVIDIA suddenly dropped a new bombshell — the next-generation Rosa CPU.

💡 What You Will Learn

Vera has barely shipped, and NVIDIA is already teasing its next-gen CPU — three generations in three years, who can keep up with that pace? It's been a while since we've seen NVIDIA this aggressive

📜 Table of Contents

• From Grace to Rosa, NVIDIA's CPU Journey Took Only Three Years
• Vera's Performance Is Already Making x86 Sweat
• The Rigel Core: Same Die Area, More Single-Core Muscle Through Architecture
• Agentic AI Is the Real Driving Force Behind the Scenes
• This Roadmap Goes Way Beyond 2028
• Final Thoughts

Vera Just Started Shipping, and NVIDIA Already Leaked the Next-Gen CPU — Three Generations in Three Years, Who Can Keep Up With This Pace

It's been a long time since we've seen NVIDIA this aggressive on CPUs.

Right as Vera entered mass production and the first batches landed in the hands of Anthropic and OpenAI, NVIDIA suddenly dropped a new bombshell — core architecture details for the next-gen Rosa CPU, codenamed Rigel.

Same die area, no increase in chip size, but single-core performance keeps climbing. And it's not targeting the general-purpose computing market at all — it's aimed squarely at the exploding new赛道: Agentic AI.

From Grace to Rosa, NVIDIA's CPU Journey Took Only Three Years

Let's lay out the timeline first, and you'll see just how fast NVIDIA has been moving:

Three years ago, Grace was still using Arm's stock Neoverse cores — basically a "rebranded" solution. By Vera, NVIDIA had already switched to fully custom Olympus cores. And for Rosa in 2028, they're going straight to the second-generation custom core, Rigel.

From rebranding to custom silicon, and then iterating on that custom design — all in just five years. That pace is rare even by the standards of the entire semiconductor industry.

Vera's Performance Is Already Making x86 Sweat

Before talking about Rigel, let's look at how impressive its predecessor Olympus has been.

Phoronix recently ran the first independent benchmarks on Vera, and the results were explosive:

• ✅ ~11% higher than AMD EPYC 9575F (geometric mean)
• ✅ 63% performance improvement over the previous-gen Grace
• ✅ ~50% IPC uplift on single-core
• ✅ Memory bandwidth per-core efficiency is 4x+ that of x86 competitors

And these numbers were achieved under a 450W TDP power envelope, with the memory subsystem drawing under 40W. High performance + low power — hard not to be nervous.

The Rigel Core: Same Die Area, More Single-Core Muscle Through Architecture

Rosa's Rigel core is based on Arm v9.2, same instruction set family as Olympus, but NVIDIA has been explicit about several key upgrades:

⏸️ Larger L2 cache — Vera already has 2MB L2 per core (2x Grace), and Rigel will push even higher, dramatically improving cache hit rates for single-threaded workloads

⏸️ Better instruction delivery mechanism — optimized specifically for AI Agent scenarios. Agents constantly cycle through "think → call tools → return results," and that kind of control-flow switching puts extreme pressure on instruction scheduling

⏸️ More efficient memory handling — paired with next-gen LPDDR6/LPDDR6X, further reducing per-core power draw

The key takeaway: same silicon area, but single-core performance keeps rising. This means NVIDIA isn't taking the old "just pile on more cores" approach — they're actually doing real work at the architecture level.

Agentic AI Is the Real Driving Force Behind the Scenes

You might be thinking, isn't NVIDIA making CPUs a bit "off-brand"?

But once you understand what Agentic AI workloads actually look like, it all makes sense:

When an AI Agent runs, 80% of the scheduling, orchestration, tool calling, and concurrency control happens on the CPU. The GPU only handles the inference portion. The smarter the Agent, the more tools it calls, and the more critical single-threaded performance and cache size become.

NVIDIA's data shows Vera outperforming competitors by 50% in Agent sandbox scenarios. And Rosa pushing single-core performance even higher is essentially saying one thing:

In the Agent era, the CPU can't be the bottleneck holding back the GPU.

This Roadmap Goes Way Beyond 2028

According to NVIDIA's public roadmap:

• 2026: Vera Rubin (Vera CPU + Rubin GPU) ramps to volume shipments
• 2027: Rubin Ultra upgraded version
• 2028: Feynman (3D-stacked GPU + custom HBM) paired with Rosa CPU
• 2030: Rigel core trickles down to RTX Spark desktop-class AI chips

In other words, the Rigel core announced today won't just power 2028's supercomputing clusters — it'll also show up in the AI PC sitting on your desk in the future.

Final Thoughts

What NVIDIA is doing right now is essentially the exact same playbook they used in the GPU market: enter with a "good enough" solution (Grace), then rapidly pivot to custom architecture and iterate (Vera→Rosa), doubling performance within the same product segment every generation.

The difference is, this time the opponent isn't AMD and Intel's GPU businesses — it's their core stronghold: the CPU.

When a company can go from zero to crushing EPYC with a custom CPU in just two years, guess what happens in the next two?

How do you think AMD and Intel will respond? Drop your thoughts in the comments.

Source: Wccftech / NVIDIA Blog