Maximizing AI Factory Performance per Watt with NVIDIA DSX MaxLPS

NVIDIA Targets Power Constraints with DSX MaxLPS for AI Factories

NVIDIA has detailed DSX MaxLPS, a system-level suite of technologies designed to maximize AI compute output within the fixed power budgets of modern data centers. The initiative directly addresses the industry's shift from measuring capacity by GPU count to evaluating efficiency by application-level performance per watt. By reclaiming power capacity that is typically stranded by outdated static provisioning methods, DSX MaxLPS aims to increase GPU density and overall AI factory productivity without requiring additional grid power.

Reclaiming Stranded Power with a Software-Defined Approach

The core of the system is a combination of dynamic power allocation, software optimizations, and advanced thermal design. Traditional data centers allocate power to each rack to handle worst-case peak demand, leaving significant headroom unused during normal operation. DSX MaxLPS replaces this model with a dynamic approach, using its Dynamic Power Software (DPS) to continuously monitor and redistribute unused power to GPUs that can use it for active computation. This turns isolated power islands into a coordinated, fungible power pool.

• Dynamic Power Allocation: Continuously monitors and reallocates unused power headroom across GPUs and racks using policy-driven controls.
• Performance-per-Watt Techniques: Employs optimized workload profiles (WPPS) to align GPU power, memory, and frequency settings with specific job types like inference or training.
• 45°C Thermal Design: Leverages warm-water liquid cooling to improve Power Usage Effectiveness (PUE), dedicating a larger portion of the facility's power budget directly to compute.

The Impact on Data Center Design and Economics

The practical implications for AI infrastructure operators are substantial. Validated on NVIDIA Vera Rubin NVL72 and GB200 NVL72 systems, DSX MaxLPS was shown to enable up to 40% more GPU capacity within the same facility envelope. This translates to a 1.3x to 1.5x improvement in performance per watt for representative inference workloads. This capability allows data center architects to design for a site's full lifecycle capacity from day one, accommodating future workload shifts from power-intensive training to more efficient inference without costly facility retrofits.

Strategic Takeaway: NVIDIA is repositioning power management from a static infrastructure problem to a dynamic, software-defined optimization layer, treating megawatts as a fungible resource to be managed in real-time for maximum computational output and economic return.