Kyber Rack: Nvidia Shows Rubin Ultra Boards

Nvidia's next AI accelerator, Rubin Ultra, gets a new server platform without cables. Up to 144 fit into one system.

The heart of Nvidia's next top server model. The slots where Rubin Ultra GPUs will be located in the future are still covered here. Presumably, there are no finished samples yet.

(Image: Nvidia)

at 9:16 am CET

2 min. read

By

Mark Mantel

At Nvidia's in-house trade fair, GTC 2026, CEO Jensen Huang pulled a wealth of new chips and server platforms out of his hat. At the very top of the portfolio, starting in the second half of 2027, is Rubin Ultra. As with the transition from Blackwell to Blackwell Ultra, Nvidia is improving an existing architecture for AI data centers.

In the case of Rubin Ultra, the company doubles the silicon usage compared to the normal Rubin version: a GPU then consists of four compute dies instead of two, each measuring 800 mm². Consequently, the available computing power is expected to roughly double. This is complemented by 16 High-Bandwidth Memory (HBM4e) stacks, which together have a capacity of one terabyte.

Presentation on Rubin Ultra with Jensen Huang on stage — Still from GTC 2025: Nvidia's announcement of Rubin Ultra with initial specifications.

(Image: Nvidia)

Kyber Rack with 144 Giant GPUs

At GTC 2025 a year ago, Huang already showed a concept of 144 Rubin Ultra GPUs with 72 self-designed Vera processors. At the time, the system was still called Rubin Ultra NVL576, based on the 144 times four compute dies. However, it is questionable whether the final system will be called that.

Nvidia has devised a new structure for this platform: instead of mounting CPUs and GPUs on a horizontal server drawer, several vertical levels are used. The redesign is likely also due to the increasing electrical power consumption. The announcement can be seen in the keynote video from minute 1:53:08.

Jensen Huang during a presentation — The three levels of a Kyber Rack: on the far left, a board with CPUs and GPUs, to its right a midplane, and on the right the backplane.

(Image: Nvidia)

On the foremost level sits the actual computing hardware, including four Rubin Ultra GPUs and two Vera CPUs. Behind it is a so-called midplane, through which the power supply and data connections run. At the very back, for several mainboards, there is an NVLink backplane with network switches that connect all chips together.

Nvidia adds the suffix Kyber to the midplanes, backplane, and the complete rack. The construction has a decisive advantage: it relies on fixed power and data lines, so no complex cabling is necessary. This is said to significantly shorten installation time.

Videos by heise

A single Rubin Ultra is said to achieve 100 petaflops in the most compact data format FP4 with four-bit floating-point numbers, i.e., 100 quadrillion calculations per second. A complete system with 144 accelerators reaches 15 FP4 exaflops or, alternatively, 5 FP8 exaflops.

Empfohlener redaktioneller Inhalt

Mit Ihrer Zustimmung wird hier ein externer Preisvergleich (heise Preisvergleich) geladen.

Preisvergleiche immer laden

Ich bin damit einverstanden, dass mir externe Inhalte angezeigt werden. Damit können personenbezogene Daten an Drittplattformen (heise Preisvergleich) übermittelt werden. Mehr dazu in unserer Datenschutzerklärung.