NVIDIA Rosa CPU with Rigel Core targets Agentic AI with higher per-core performance

NVIDIA Reveals Rosa CPU Architecture Details for Agentic AI Workloads

NVIDIA has disclosed the first substantial details about its next-generation Rosa CPU, which will feature a brand new core architecture called the Rigel core 1

1

2

2

. The announcement comes shortly after the company's Vera CPU with Olympus cores began shipping in late May, signaling NVIDIA's aggressive push into the data center processor space. Rosa will launch alongside the company's Feynman lineup and is specifically optimized for Agentic AI workloads, which continue to demand increased processing capabilities as AI agents become more sophisticated and autonomous.

Rigel Core Delivers Performance Gains Within Same Silicon Footprint

The Rigel core represents NVIDIA's next step in Arm v9.2-based design, delivering higher per-core performance than the Olympus cores found in Vera while maintaining the same silicon footprint 1

1

. This achievement is particularly noteworthy as it demonstrates NVIDIA's ability to extract more performance without increasing chip size or power consumption. The NVIDIA Rosa CPU will achieve even more max single-threaded CPU performance at scale than its predecessor, building on Vera's already impressive capabilities. While the current NVIDIA Vera CPU features 88 Olympus cores—up from 72 cores on Grace—NVIDIA has not yet disclosed whether Rosa will include a core count increase 2

2

.

Key Architectural Improvements Target Memory and Instruction Efficiency

NVIDIA highlighted three major improvements in the Rosa architecture: better instruction delivery, a larger L2 cache, and more efficient memory handling 1

1

2

2

. These enhancements address critical bottlenecks in AI processing workloads where memory bandwidth and latency can significantly impact overall system performance. The larger L2 cache will enable the processor to keep more frequently accessed data closer to the cores, reducing the need to fetch information from slower main memory. Combined with improved instruction delivery mechanisms, these changes should translate into measurable performance gains for the complex, multi-step reasoning tasks that characterize agentic AI applications.

Competitive Positioning Against AMD and Intel in Server CPU Markets

The Rosa announcement positions NVIDIA to compete more directly with established server CPU markets leaders, particularly AMD EPYC Venice and Intel Xeon Diamond Rapids 1

1

. NVIDIA's strategy of pairing its CPUs with Blackwell GPUs and future GPU architectures creates tightly integrated systems optimized for AI workloads. The company has made significant progress since Grace, with Vera already in full production and shipping in Vera Rubin configurations to major AI firms globally 2

2

. Rosa is expected to enter data centers by 2029, followed by PC-specific variants in Rosa Feynman Spark solutions by 2030, extending NVIDIA's CPU strategy beyond enterprise into consumer markets. The same core architectures will power next-generation RTX Spark chips, with the first Grace-based versions expected this fall and Vera Rubin combinations arriving in 2028 2

2

.