NVIDIA teases next-gen Kyber rack-scale tech: up to 576 NVIDIA Rubin Ultra GPUs in 2027

NVIDIA has teased more details of its next-gen Kyber rack-scale AI server, which will house a huge 576 NVIDIA Rubin Ultra AI GPUs in 2027.

2

At the OCP Global Summit recently, NVIDIA showcased some of what to expect in the future from AI factories, with the company unveiling multiple interesting new developments. One of those is the work NVIDIA has put into its new Kyber rack-scale technology, replacing Oberon and would see the company scaling up to its NVL72 configuration, using 576 NVIDIA Rubin Ultra AI GPUs.

NVIDIA explains: "The OCP ecosystem is also preparing for NVIDIA Kyber, featuring innovations in 800 VDC power delivery, liquid cooling and mechanical design. These innovations will support the move to rack server generation NVIDIA Kyber - the successor to NVIDIA Oberon - which will house a high-density platform of 576 NVIDIA Rubin Ultra GPUs by 2027".

• Read more: NVIDIA could change cooling solution for Rubin Ultra AI GPUs for huge 2300W thermal concerns

NVIDIA's current-gen rack configuration is dubbed Oberon, while its next-gen rack configuration is the recently-teased Kyber deployment. It features new innovations like 800 VDC power delivery, liquid cooling and mechanical design, which the company says are just a couple of the parts that will make up Kyber.

2

Oberon has been NVIDIA's main design for the Blackwell GPU generation with GB200 and GB300, but with Rubin Ultra, NVIDIA plans to transition into the new Kyber generation, which has multiple big upgrades. One of the big improvements is the rack architecture, with NVIDIA moving to stacking compute trays, mounted vertically -- as Wccftech puts it "similar to books" -- which are called vertical blades, as shown in the picture above.

Kyber will allow for higher GPU density with more efficient networking, with built-in NVLink switch blades inside of the Kyber rack enclosure, enabling higher scalability and easier maintenance. On the performance side of things, another benefit of Kyber is that NVIDIA can switch to using an 800 VDC facility-to-rack power delivery method, which is a huge upgrade from the 412 or 480 VAC three-phase systems used in Oberon.

• Read more: AI servers of the future: 'rack density' of 1000kW+ with NVIDIA's next-gen Rubin Ultra AI GPUs

NVIDIA will enjoy the benefits of massively improved power efficiency from Kyber, as well as delivering a 150% upgrade in power transferred through the same copper wire, which NVIDIA says will save millions of dollars through far less copper needed to deploy large-scale clusters.