DSX MaxLPS
A power management technology for AI data centers introduced by NVIDIA. By applying dynamic scheduling and statistical multiplexing principles, it shifts GPU cluster power provisioning from static peak reservation to elastic resource allocation, enabling more GPU compute capacity to be accommodated within a given power budget and improving the overall energy efficiency of AI factories.
Timeline (last 90 days)
NVIDIA announced DSX MaxLPS (Maximum Load Per Site) technology.
DSX MaxLPS is designed to address the waste caused by power redundancy in hyperscale data centers.
MaxLPS leverages the statistical property that different tasks at different stages exhibit offset peak loads, enabling more GPUs to be accommodated within a given power capacity through intelligent power management and scheduling.
MaxLPS introduces the statistical multiplexing concept of power oversubscription into the power supply layer of GPU clusters.
NVIDIA positions MaxLPS as a key component for maximizing AI factory throughput and efficiency.
The information disclosed by NVIDIA this time leans toward concepts and value propositions; the specific implementation details, performance data, and applicable hardware platforms of MaxLPS have yet to be further disclosed.
All Facts (5)
DSX MaxLPS is designed to address the waste caused by power redundancy in hyperscale data centers.
50%UnverifiedMaxLPS leverages the statistical property that different tasks at different stages exhibit offset peak loads, enabling more GPUs to be accommodated within a given power capacity through intelligent power management and scheduling.
50%UnverifiedMaxLPS introduces the statistical multiplexing concept of power oversubscription into the power supply layer of GPU clusters.
50%UnverifiedNVIDIA positions MaxLPS as a key component for maximizing AI factory throughput and efficiency.
50%UnverifiedNVIDIA announced DSX MaxLPS (Maximum Load Per Site) technology.
50%