NVIDIA Debuts DSX MaxLPS Power Scheduler for AI Factories, Lifting Token Output per Megawatt by Up to 40%

At its AI Infrastructure Summit on September 16, NVIDIA demonstrated the DSX MaxLPS power scheduling system, designed for AI factories. The company said the system can increase token throughput per megawatt by up to 40% without changing the power budget.

Token throughput per megawatt refers to how many tokens an AI system can generate or process per second for each megawatt of power consumed. The metric accounts for power usage, making it suitable for measuring output in power-constrained AI data centers.

NVIDIA noted that the DSX MaxLPS technology has been integrated by Emerald AI into the Conductor platform for data center load orchestration under grid conditions. The Conductor AI platform's DSX Flex can dynamically adjust data center power demand based on grid status. The platform has received 200 grid status signals, and none of the related operations required user intervention.

Cloud providers such as Lambda have tested the system. Lambda, a GPU cloud provider with more than 10,000 customers, ran a test cluster with 5 racks and 19 nodes. Within the same power budget as a 16-node cluster running at full power, the 19-node cluster's throughput rose from about 4 million tokens per second to 5 million tokens per second, a 24% increase, with a 23% gain in performance per watt.