The Real Test Comes in 2028! What’s the Goal of NVIDIA CPU’s Two-Year Iteration Cycle?

When a company that once treated CPUs as a supporting role starts compressing its update cadence to one cycle every two years, this move is no longer about patching a weakness—it is about proactively entering the middlegame. In my view, the Vera and Rosa roadmap NVIDIA unveiled this time looks like seizing the initiative in an endgame: first, use its self-developed Arm CPU to secure its footing in the ecosystem, then package the capabilities of the entire platform and bring them to market, directly competing with Intel and AMD on cadence.

Let’s start with Vera, which has already been played. The known information is very clear: it is based on the Arm instruction set, uses NVIDIA’s in-house architecture, has 72 cores, and is no longer just for “internal use” but is beginning to serve both internal and external needs, including commercial sales. One line in the original text puts it very bluntly: “NVIDIA CPU processors are shifting from internal self-use to serving both internal and external needs—used by itself and sold externally as well.” This means it is no longer content with merely supporting its own platform, but intends to truly enter the arena and compete head-on in the general-purpose processor market.

From a system design perspective, Vera emphasizes not only piling on core counts, but also foundational capabilities such as spatial multithreading and high-speed interconnects. NVIDIA officially claims a 1.5x IPC performance increase, with clear gains in both single-threaded and multithreaded efficiency. This signal is crucial, because server and AI platforms are not judged by any single chip alone, but by whether the CPU, GPU, DPU, and network can work together like a finely coordinated tactical system—remove any one link, and it can turn into wasted resources.

Even more intriguing is Rosa. According to the roadmap, it will debut in 2028 as part of the next-generation computing platform, and it will continue to place emphasis on even more extreme single-thread performance. It will also connect next-generation network cards, the ConnectX-10 SuperNIC, through the BlueField-5 DPU, while the platform’s core GPU will be the previously announced Feynman. The latter has two new highlights: one is die stacking, and the other is custom HBM4. Reports also mention that it will likely be derived from C-HBM4E, with single-package capacity expected to exceed 1TB and bandwidth continuing to climb. Put plainly, NVIDIA is not making a move with a single CPU piece, but unleashing a combination attack built around an entire computing platform.

In terms of strategic depth, the most noteworthy point is the change in development cycle. What used to be one cycle every four years has now been shortened to two, which essentially brings NVIDIA closer to the mainstream cadence of Intel and AMD. The original text also states clearly: “This also shows that NVIDIA’s CPU development cycle is shrinking from 4 years to 2 years, aligning itself with Intel, AMD, and others, which generally introduce new architectures about every two years.” The significance of a two-year refresh cycle is not just livelier product launches—it means the ecosystem, customer expectations, and R&D organization must all accelerate in step. If NVIDIA can execute this steadily, the threat will become very real.

In addition, the name Rosa was not chosen casually. It comes from Rosalyn, referring to Rosalyn Sussman Yalow. She was born on July 19, 1921, and passed away on May 30, 2011. She received the 1977 Nobel Prize in Physiology or Medicine for developing the radioimmunoassay of peptide hormones. This technique made it possible to detect virus, hormone, and drug levels in serum without the need for bioassays, and it was used in hepatitis blood screening, antibiotic efficacy testing, and drug dosage measurement. She also gave up patent rights, allowing the technology to be promoted free of charge. This naming continues NVIDIA’s convention of naming CPU/GPU codenames after renowned physicists and scientists, adding a layer of humanistic meaning to an otherwise cold hardware roadmap.

But the issue is that a beautiful roadmap does not guarantee a winning position in the market. The CPU sector is not like graphics cards, where a single breakout can easily turn the tide—especially when rivals have already built thick moats across desktops, servers, and the broader ecosystem. Any misstep comes at a price, like an exchange in a strategic game: on the surface you may not lose, but you can end up surrendering the initiative entirely. Still, judging from the trend, NVIDIA clearly does not want to be only the strongest GPU supplier of the AI era. It wants to hold the CPU, GPU, DPU, and network all in its own hands. For industry users, this will bring more platform choices; for Intel and AMD, this match has already moved beyond probing at the edges and formally entered a fierce middlegame clash.