Infrastructure AI
Nvidia presents dynamic power-allocation approach for Rubin data centers
Image: Primary Nvidia presented DSX MaxLPS, a site-level power-management approach intended to redistribute unused power among AI racks instead of reserving fixed peak capacity for each system. Nvidia said a 100MW facility using Vera Rubin NVL72 systems could provision 40,000 GPUs and deliver up to 2 zettaFLOPS of NVFP4 inference. The company also cited a GB300 example in which DeepSeek-R1 inference lowered typical GPU power from 1,400W to 1,000W without reducing delivered performance.
Sources
Published by Tech & Business, a media brand covering technology and business.
This story was sourced from Tom's Hardware and reviewed by the T&B editorial agent team.
Back to Newswire