Nvidia debuts Rubin chip with 336B transistors and 50 petaflops AI performance.

Nvidia Corp. today announced Rubin, a new flagship graphics processing unit that provides five times the inference performance of Blackwell. The GPU made its debut at CES alongside five other data center chips. Rubin includes 336 billion transistors that provide 50 petaflops of performance when processing NVFP4 data, compared to Blackwell's up to 10 petaflops. Rubin's training speed is 250% faster at 35 petaflops. The chip's computing power includes a module called the Transformer Engine, which also shipped with Blackwell. According to Nvidia, Rubin's Transformer Engine is based on a newer design with a performance-boosting feature called hardware-accelerated adaptive compression, which reduces the number of bits a file contains and decreases the amount of data AI models have to process. "Rubin arrives at exactly the right moment, as AI computing demand for both training and inference is going through the roof," said Nvidia Chief Executive Officer Jensen Huang. "With our annual cadence of delivering a new generation of AI supercomputers — and extreme codesign across six new chips — Rubin takes a giant leap toward the next frontier of AI."

Nvidia plans to ship Rubin as part of an appliance called the Vera Rubin NVL72, which combines 72 Rubin chips with 36 of the company's new Vera central processing units, also debuted at CES. Vera includes 88 cores based on a custom design called Olympus and is compatible with Armv9.2, a widely-used version of Arm Holdings plc's CPU instruction set architecture. The Vera Rubin NVL72 keeps its chips in modules called trays with a cable-free design that cuts assembly and servicing times by a factor of up to 18 compared with Blackwell-based appliances. The RAS Engine, a subsystem that automates maintenance tasks, has been upgraded to provide fault tolerance features and perform real-time health checks to verify hardware is working as expected. The Vera Rubin NVL72 provides 260 terabits per second bandwidth, which is more than the entire internet, with 220 trillion transistors across the system.

Three new networking chips facilitate data exchange within and across systems. NVLink 6 Switch enables multiple GPUs inside a Vera Rubin NVL72 rack to exchange data simultaneously to coordinate distributed AI models. The Spectrum-6 is an Ethernet switch that facilitates connections between GPUs installed in different racks. The ConnectX-9 is a SuperNIC, a hardware interface that allows servers to access the data center network while performing networking tasks historically carried out by a server's CPU, leaving more processing capacity for AI workloads. Additionally, the BlueField-4, a data processing unit, offloads work from a server's main processor across networking, cybersecurity, and storage management operations and powers the new Inference Context Memory Storage Platform designed to optimize large language models' key-value cache.

The Vera Rubin NVL72 will ship alongside a smaller appliance called the DGX Rubin NVL8 that includes eight Rubin GPUs instead of 72. Both systems form the basis of the DGX SuperPOD, a reference architecture for building AI clusters that combines Nvidia's latest chips with Mission Control, a software platform for managing AI infrastructure. Rubin-powered systems will start shipping in the second half of 2026.