NVIDIA Introduces ‘Kyber’ Rack-Scale Architecture: Scaling Up to 576 Rubin Ultra GPUs by 2027 for Enhanced AI Performance

NVIDIA recently unveiled several key advancements in its AI computing portfolio during the OCP Global Summit, particularly focusing on the evolution of its Kyber rack-scale generation—a major leap forward in technology.

NVIDIA’s Kyber Rack-Scale Generation: Elevating GPU Density and Power Efficiency

In a landscape where the demand for enhanced computing power continues to escalate, NVIDIA is increasingly setting the pace for innovation. At the OCP Global Summit, the company showcased an exciting vision for the future of ‘AI factories’, highlighting transformative developments in AI compute solutions. Among these, the introduction of the new-gen Kyber rack-scale technology stands out, poised to succeed the Oberon design and potentially enable a remarkable NVL576 configuration.

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.– NVIDIA

For those less familiar, Kyber and Oberon are terms that denote different generations of rack architecture, with a focus on chip configurations, power capacities, and cooling methodologies. Historically, Oberon has been NVIDIA’s primary influence for Blackwell generations (GB200/GB300), but with the forthcoming Rubin Ultra, the transition to Kyber marks a significant evolution with several key enhancements.

NVIDIA computer server with multiple unit slots visible in a server rack. — NVIDIA’s Kyber Compute Nodes | Image Credits: ServeTheHome

One of the most notable advancements includes a revamped rack architecture featuring vertical stacking of compute trays, akin to books on a shelf—known as vertical blades. This design not only maximizes GPU density but also enhances network efficiency.

Another substantial upgrade involves the integration of NVLink switch blades directly within the rack enclosure, which promises improved scalability and simplified maintenance. In terms of power performance, one of Kyber’s standout features is the transition to an 800 VDC facility-to-rack power delivery model, a notable upgrade from the former 415 or 480 VAC three-phase systems.

This new power model is expected to yield remarkable improvements in energy efficiency, allowing for a 150% increase in power transmission through existing copper wiring. This innovation is projected to lead to significant cost savings—”millions saved” by minimizing the amount of copper needed for extensive cluster deployment.

Ultimately, the advancements brought by the Kyber generation will set the stage for NVIDIA’s ambitious Rubin Ultra NVL576 systems, unlocking unprecedented computational power and reinforcing NVIDIA’s leadership in the race for sophisticated AI infrastructure.

Source & Images