NVIDIA has announced several important updates to its AI compute portfolio at the OCP, including development around its Kyber rack-scale generation, a significant advancement.
NVIDIA's Kyber Rack-Scale Generation Brings Massive Increase In Per Rack GPU Density & Power Figures
As the world increasingly demands more computing capabilities, NVIDIA has been making advances across its entire tech stack, leaving competitors with little room to catch up in the race. Now, at the OCP Global Summit, Team Green has showcased the future of 'AI factories' and while the firm reveals several interesting developments, one of the most optimistic ones includes the work around next-gen Kyber rack-scale technology, which will replace Oberon and potentially allow the firm to scale up to a massive 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
If you are still unfamiliar with the topic being discussed, Kyber and Oberon refer to rack generations, where the primary aspects addressed are chip arrangements, power ratings, and cooling solutions. Oberon has been NVIDIA's dominant design for the Blackwell generations (GB200/GB300), but with Rubin Ultra, Team Green plans to transition to the Kyber generation, which will bring several upgrades. One central improvement area is rack architecture, where NVIDIA will switch to stacking compute trays, mounted vertically, similar to books, called vertical blades.
This technique will bring in higher GPU density and more efficient networking. Another interesting upgrade is the inclusion of built-in NVLink switch blades inside the same rack enclosure, which will enable higher scalability and easier rack maintenance. In terms of performance, one of the key benefits of Kyber is that NVIDIA will switch to an 800 VDC facility-to-rack power-delivery model, which is also an upgrade from the previous 415 or 480 VAC three-phase system.
This will enable significantly improved power efficiency figures, and more importantly, a 150% increase in power transferred through the same copper wire, leading to 'millions saved' by reducing the amount of copper needed to deploy large-scale clusters. Overall, Kyber will pave the way for NVIDIA's Rubin Ultra NVL576 systems, bringing unprecedented computing power onboard and driving the race for cutting-edge AI infrastructure.
Follow Wccftech on Google to get more of our news coverage in your feeds.
