During GTC 2025, Nvidia presented more details about its future high -performance artificial intelligence infrastructure generation: Rubin Ultra and the rack system NVL576 Kyber.
Expected to reach the market from 2027this technology aims to bring a significant leap in computational capacity, with servers that can consume Up to 600 kW per rack.
The company had already introduced the line Blackwell B200which can demand about 120 kw por rack. With the evolution for the Vera Rubinem 2026and later for the Rubin Ultrathe expectation is for an exponential increase in efficiency and server processing density.
According to Jensen Huang, CEO of Nvidia, Datacers of the future may require megawatts by rack to meet the growing demands of the Generative Ia and Deep Learning.
The new Rubin Ultra architecture and the Kyber infrastructure
O Rubin Ultra will be based on Nvidia’s new architecture and will bring innovations as o NVL576 Kybera rack system designed to support the very high demand for energy and cooling of these servers.
The system will have four “pods” per rack, each equipped with 18 blades. Each blade may contain eight Rubin Ultra GPus and two CPUS Veraresulting in 176 GPUS by Pod and 576 GPUS by Rack.
NVIDIA also confirmed that the next generation of NVLinkrequired to interconnect GPUS and CPUS, will receive significant improvements. The new units will have Three next generation NVLINK connectionsagainst two of the current model, providing greater bandwidth and communication efficiency between chips.
Data centers will need much more energy than we imagined a year ago. This growing demand will require more robust architectures, such as Rubin Ultra
Jensen Huang, NVIDIA CEO
Scaling the performance: From Blackwell Ultra to Rubin Ultra
The line Blackwell Ultra GB300announced for 2025, already represents an advance in relation to the previous generation, with 288 GB of HBM3E Memory e 20 petaflops de performance FP4 by gpu. However, the big leap will come with the Vera Rubinscheduled for 2026, which will have 50 petaflops de fp4 By chip.
O Rubin Ultrascheduled for 2027, will double this number, reaching 100 petaflops de fp4 by chip and 1 tb of memory. A complete rack NVL576 Rubin Ultra can deliver 15 FP4 Inference Examflows e 5 FP8 training examsrepresenting an increase in 14 times in performance compared to the racks Blackwell Ultra NVL72 of 2025.
Energy Challenges and Impact on Data Centers
The exponential growth of computational demand raises questions about energy efficiency and infrastructure. Current servers already require huge amounts of electricity, and with the projection of racks consuming 600 kW or butData Centers will need advanced cooling and energy distribution solutions.
In addition, experts discuss the environmental impacts of this advance. Increased computational capacity is required for advanced AI applications, but energy consumption should be managed sustainably.
Also read:
The Future: Towards Architecture Feynman
Beyond Rubin UltraNvidia has already announced that its next great architecture, called Feynmanis scheduled for 2028. The name honors Richard Feynman, renowned theoretical physicist, as well as Vera Rubin, an astronomer who named the generation of 2026.
Nvidia continues to lead the AI sector with aggressive advances in hardware and infrastructure. The challenge now is to balance computational capacity, energy efficiency and sustainabilityensuring that the evolution of AI is supported by increasingly robust and innovative date centers.
Source: NVIDIA
Join the Adrenaline offers group
Check out the main offers of hardware, components and other electronics we find over the internet. Video card, motherboard, RAM and everything you need to set up your PC. By participating in our group, you receive daily promotions and have early access to discount coupons.
Enter the group and enjoy the promotions
Source: https://www.adrenaline.com.br/nvidia/nvidia-infraestrutura-rubin-ultra-600-kwatts-por-rack/
