NVIDIA has, for the first time, revealed details about the internals of its next-generation datacenter CPU, "Rosa." In an official blog post published on July 7 (US time), the company explained that Rosa will adopt a next-generation Arm v9.2 CPU core called "Rigel," which achieves higher per-core performance than "Olympus," the core used in the current Vera, within the same silicon area.
This carries more weight than simply adding another CPU codename. NVIDIA used Arm Neoverse-based CPUs in Grace, then moved to its own in-house designed Olympus with Vera. With Rosa, the company is bringing its next in-house CPU core to the forefront as a server CPU for the AI agent era. Behind the race to add more GPUs lies the view that how fast the CPU can execute tools and code, and return data processing and verification results, determines the utilization rate of the entire AI factory.
Rigel Extends Olympus, Pushing Single-Core Speed Further
The Rosa specifications NVIDIA has revealed so far remain limited. Rigel is an Arm v9.2-generation CPU core that delivers higher per-core performance than Olympus within the same silicon area. The improvements cited include enhanced instruction supply, a larger L2 cache, and more efficient memory handling.
What stands out in this announcement is that NVIDIA is not foregrounding an increase in core count. Vera already features 88 Olympus cores and handles 176 threads via Spatial Multithreading. The changes disclosed for Rosa are design changes aimed at making a single core progress faster within the same area.
AI agent processing does not complete solely through model inference on the GPU. Every time a model decides on its next step, the CPU must call tools, run Python or tests, retrieve data, and verify results. These tasks are prone to becoming a queue of waiting. Adding more cores does not necessarily speed up a single agent's next move.
The Single-Thread Performance Standard Set by Vera
The significance of Rosa becomes easier to see when measured against the design of Vera. NVIDIA describes Vera as belonging to a new CPU category: "max single-threaded CPU at scale." The reasoning is that while high-density datacenter CPUs are moving toward increasing core count per cost, AI agents require each core to maintain speed even under full load and return results with predictable latency.
Vera's Olympus cores reportedly offer 50% higher IPC compared to Grace. At the overall CPU level, Vera combines LPDDR5X memory via SOCAMM, providing up to 1.2TB/s of memory bandwidth and up to 1.5TB of capacity. NVIDIA states that memory power consumption is under 40W, depending on the configuration.
Furthermore, Vera uses a single compute die combined with the second-generation NVIDIA Scalable Coherency Fabric. According to the official page, 88 cores, cache, memory, I/O, and NVLink-C2C are connected on a single die, achieving 3.4TB/s of bisection bandwidth. Rather than cutting costs through chiplet design, this is an architecture built to avoid slowdowns caused by cores competing for memory and cache.
CPU-Side Processing That Keeps the GPU Waiting
The reason NVIDIA emphasizes the CPU to this degree is that AI agent work consists of a continuous series of small waiting periods. A delay in a single tool call, a single code execution, a single test, or a single search causes a delay in the next model call as well. GPUs are expensive assets, and waiting on the CPU side directly translates into lower GPU utilization.
NVIDIA states that Vera delivers 1.8x sustained per-core performance compared to x86 on CPU workloads representative of agentic execution. In Perplexity's real-world workflows, cloning a repository and running a test suite in a sandbox was about 1.5x faster, and launching concurrent sandboxes was up to 1.9x faster, according to the company. Other cited measurements include a 3x improvement in Starburst's large-scale SQL analytics and up to a sixth of the latency in Redpanda's real-time streaming.
This claim also connects to the rack configuration of Vera Rubin. The NVIDIA Vera Rubin NVL72 combines 72 Rubin GPUs with 36 Vera CPUs. A separate line of Vera CPU racks stacks up to 256 Vera CPUs, supporting over 22,500 concurrent sandbox environments. As the unit of the AI factory expands from GPU cards to racks and further to PODs, the CPU becomes not a supporting component but the foundation that keeps an agent's work queue from clogging.
Specification Gaps That Remain for Rosa
Regarding Rosa, the confirmed official information at this point is limited to the direction of the Rigel core. Core count and thread count have not been disclosed. L2 capacity and memory bandwidth also remain unknown. NVIDIA has not revealed the memory type, die configuration, or power consumption, nor has it explained the server configuration. NVIDIA's statement about "the same silicon area" also refers specifically to a comparison between the Rigel and Olympus cores, not a claim that the entire Rosa CPU will match Vera's die area.
Tom's Hardware has reported, as part of the GTC 2026 roadmap, a pairing of the 2028-generation Feynman GPU with the Rosa CPU. However, the new information added by today's official blog post concerns the CPU core itself, not the rack specifications for the Feynman generation. What NVIDIA needs to reveal next is how many Rigel cores will be included, what memory it will be paired with, and how much agent processing it will support in which rack.
The shift from Grace to Vera signaled the end of the era in which NVIDIA treated the CPU as a supporting role to the GPU. Rigel in Rosa pushes that direction one generation further. A faster CPU raises the value of the GPU. In the design of AI factories, this paradox is only growing stronger.