The Rubin platform spreads workloads across new chips, including the Vera CPU, Rubin GPU, NVLink 6 switch, ConnectX-9 SperNIC, BlueField, and Spectrum-6 Ethernet switch. The combined firepower will cut training time and inference token costs, Nvidia said. Rubin’s run costs will be one-tenth of its Blackwell platform by Nvidia’s estimates.
Nvidia says Rubin will use 72 GPUs in one system – just one-fourth of the GPUs used in the previous Blackwell generation. Featuring the Vera Rubin NVL72 rack-scale solution and the HGX Rubin NVL8 system, the Rubin platform will offer performance gains to better handle agentic AI and reasoning models.
Juggling Power, Performance, and Efficiency
Nvidia is pitching Rubin as an efficiency breakthrough, as much as a performance upgrade. By using fewer GPUs per system and better balancing compute and networking workloads, Rubin could help data center operators reduce power density, cooling demands, and AI infrastructure costs – if Nvidia’s lab benchmarks hold true in real-world applications.
