for more details on Nvidia's VR NVL72 Oberon and future roadmap, check out our article from February:
SemiAnalysis Twitter · SemiAnalysis (@SemiAnalysis_) · 2026-05-31
SemiAnalysis publishes a deep technical breakdown of Nvidia's Vera Rubin VR NVL72 platform, detailing its six silicon products, extreme co-design rack architecture, and total cost of ownership analysis.
Appears in
Extraction
Topics: nvidia-hardwareai-acceleratorsdata-center-infrastructuregpu-architecturehbm-memory
Claims
- Rubin's dense FP4 and FP8 FLOPs increase by roughly 3.5x versus GB200, while FP16 FLOPs rise by only about 1.6x.
- HBM bandwidth in Rubin scales approximately 2.8x over GB300, while HBM capacity remains flat.
- Nvidia's competitive advantage stems from being the only vendor offering best-in-class or near-best-in-class silicon across all major components in an AI server system.
- VR NVL72 adopts a more modular approach compared to Grace Blackwell for improved integration efficiency and throughput.
- The Vera Rubin platform uses a 3nm process and disaggregates I/O into chiplets while maintaining two reticle-sized dies with eight HBM stacks.
Key quotes
Nvidia's competitiveness strengthens with its extreme co-design supremacy. It is the only player with the best in class or close to the best in class silicon product offerings for all the major silicon contents in an Nvidia trail-blazed AI server system design.
VR NVL72 has a much more holistic design with a modular approach compared to Grace Blackwell for the purpose of integration efficiency and throughput.
Rubin's architecture prioritizes bandwidth and low-precision compute.