Nvidia's power management tech promises more compute per megawatt in data centers
Nvidia highlighted its DSX MaxLPS power management suite during a Hot Chips presentation on the Rubin GPU. The company claims that using these technologies allows operators to fit 40,000 Rubin GPUs within a 100MW power budget, delivering up to 2 zettaFLOPS for inference and 1.4 zettaFLOPS for training. These figures are estimated rather than measured, according to Tom's Hardware's analysis.
The DSX MaxLIPS suite addresses a fundamental constraint in AI infrastructure: fixed facility power budgets. Traditional static provisioning assigns worst-case power peaks per rack, which can strand unused capacity while other racks hit their limits. Nvidia's approach instead applies real-time monitoring and dynamic power reallocation across racks based on actual application load.
The company's claims center on Vera Rubin NVL72 systems, with 40,000 GPUs across roughly 40 DGX SuperPODs operating within 100MW. However, Tom's Hardware notes these performance figures are estimates derived from public specs, not measured results from real workloads, which would likely fall short of the stated zettaFLOPS targets.
This technology could reshape how data center operators plan capacity, potentially lowering the cost barrier for large-scale AI deployments by extracting more compute from existing power infrastructure. Utilities and grid operators may see more predictable demand patterns, while smaller operators could benefit from tools that maximize efficiency without massive capital investment. However, if performance claims prove optimistic, operators may face shortfalls in expected capacity, affecting AI service reliability and pricing.