NVIDIA Rubin: 50 PFLOPS AI Platform With 5x Blackwell Uplift

NVIDIA unveiled its most advanced AI platform at CES 2026—the Vera Rubin system featuring six custom chips delivering unprecedented performance leaps. With 50 PFLOPS inference capability and 88-core Vera CPUs, the Rubin platform achieves 5x inference performance gains over Blackwell while slashing inference token costs by up to 10x through extreme hardware-software codesign.

Six-Chip Architecture Breakthrough

The Rubin ecosystem integrates six purpose-built chips: the Rubin GPU with third-generation Transformer Engine, Vera CPU powered by custom Olympus Arm cores, NVLink 6 Switch enabling 3.6 TB/s bidirectional bandwidth, ConnectX-9 SuperNIC for networking, BlueField-4 DPU managing data infrastructure, and Spectrum-6 Ethernet Switch with co-packaged photonics achieving 5x better power efficiency than traditional switches.

Rubin GPU and Vera CPU Specifications

Vera Rubin NVL72 Supercomputer

The flagship NVL72 rack configuration packs 72 Rubin GPUs and 36 Vera CPUs connected through NVLink 6, delivering 3.6 EFLOPS of inference compute and 2.5 EFLOPS training performance. The system features 20.7TB of HBM4 memory, 54TB of LPDDR5x capacity, and aggregate 1.6 PB/s HBM bandwidth—260 TB/s of total scale-up bandwidth enabling seamless all-to-all GPU communication.

The redesigned rack uses cable-free modular trays enabling 18x faster assembly and servicing compared to Blackwell systems, addressing deployment challenges experienced with previous generations.

Cost and Efficiency Revolution

Rubin targets massive efficiency improvements for mixture-of-experts models, requiring only 25% of the GPUs needed for MoE training versus Blackwell while cutting inference token costs by 10x. These advances stem from HBM4’s dramatically higher bandwidth feeding compute resources, plus hardware-accelerated adaptive compression in the Transformer Engine.

The platform achieves 8x better performance-per-watt for inference compared to Blackwell, critical as data centers face power and cooling constraints at unprecedented scale.

Production and Deployment Timeline

NVIDIA confirmed all six chips returned from fabrication with satisfactory performance in Q1 2026. Major cloud providers including AWS, Google Cloud, Microsoft Azure, and Oracle Cloud Infrastructure will deploy Vera Rubin instances in the second half of 2026, alongside NVIDIA Cloud Partners like CoreWeave and Lambda.

Microsoft’s Fairwater AI superfactories will scale to hundreds of thousands of Vera Rubin Superchips, representing the largest known planned deployment.

For comprehensive AI hardware coverage, visit TechnoSports. Learn more at NVIDIA’s official blog.

FAQs