NVIDIA Unveils Next-Gen Rosa CPU Targeting AI Agent Workloads

NVIDIA (NVDA-US) has recently unveiled further technical details of its next-generation data center CPU—Rosa—for the first time.

This processor will launch alongside the next-generation Feynman GPU, targeting rapidly growing AI agent workloads by further enhancing CPU single-thread performance to deliver stronger computing power for AI inference and AI agent execution.

According to NVIDIA, Rosa will adopt the new Rigel CPU core, the next-generation in-house Arm CPU microarchitecture following Olympus, also based on the Arm v9.2 instruction set.

Compared to the Olympus core used in Vera, Rigel’s focus is not on increasing core area, but on improving per-core performance within the same silicon footprint.

NVIDIA states that key improvements in Rigel include more efficient instruction delivery, larger L2 cache, more efficient memory access mechanisms, and stronger single-core performance.

In particular, the larger secondary cache and more efficient data access capabilities help reduce memory access frequency, improving execution efficiency for numerous small tasks, frequent branching, and data scheduling during AI inference.

The Vera CPU, announced at GTC 2026, has already entered mass production and is being delivered with the Vera Rubin AI system.

Vera uses NVIDIA’s in-house Olympus core, offering approximately a 50% IPC improvement over Grace’s Arm Neoverse V2, increasing core count from 72 to 88, and introducing spatial multithreading technology for the first time.

Rosa continues this evolutionary path. NVIDIA notes that as AI agents become increasingly sensitive to CPU single-thread response speed, Rosa will further enhance single-core performance, not just core count. The official has not yet disclosed Rosa’s core count, so whether core scaling will continue remains to be confirmed.

From currently available information, the development direction of Grace, Vera, and Rosa is clear. Grace positioned NVIDIA’s formal entry into the CPU market; Vera marked the full adoption of in-house CPU cores; Rosa further optimizes core architecture, boosting single-core performance while maintaining roughly the same die area.

CPU is becoming an increasingly important component of NVIDIA’s AI platform.

Over the past few years, NVIDIA has built its AI computing advantage primarily on GPUs, but CPUs are now becoming a key part of its AI platform strategy.

Both the shared memory mechanism between CPU and GPU and the rising importance of CPUs in the AI agent era are key reasons why NVIDIA continues to invest in next-generation CPU development.

Grace, Vera, and future Rosa will not only be used in data centers but also extended to AI PC platforms. NVIDIA has previously confirmed that the future RTX Spark chip will also adopt the same CPU core architecture, enabling data center and end-user products to share the same CPU technology roadmap.

At the same time, NVIDIA continues to adopt a monolithic chip design rather than a chiplet architecture, aiming to reduce inter-die communication latency and better meet the low-latency and high-bandwidth demands of AI inference.

According to NVIDIA’s current product roadmap, Vera CPU will enter mass production in 2026 with the Vera Rubin system; Rosa CPU will launch alongside the Feynman GPU in 2028; the Rosa data center platform is expected to enter the market around 2029; and the RTX Spark AI PC chip based on the Rosa architecture is expected to debut around 2030.