Up to 40% more GPUs within the same site-power envelope
NVIDIA DSX MaxLPS is an AI factory design and operating framework that coordinates facilities and site design, NVIDIA Dynamic Power Software (DPS), and advanced performance-per-watt techniques to help NVIDIA Cloud Partners (NCPs) and other AI factory operators maximize performance per watt within a fixed power budget. Planning AI factories for MaxLPS can also make more GPU capacity available to tenants and increase throughput per megawatt.
Select one or more MaxLPS capabilities to reveal where they apply in the stack, then hover over or select a highlighted layer capability to view related resources.
MaxP and MaxQ are static GPU-level power settings: MaxP is the highest GPU power setting, while MaxQ targets the best application performance per watt. MaxLPS extends optimization across the AI factory by combining facilities and site design, DPS for dynamic power allocation, and advanced performance-per-watt techniques within a fixed site power envelope.
Select MaxP, MaxQ, or MaxLPS to update the GPU capacity and AI factory outcome within the same fixed power budget.
The capacity comparison above shows the potential AI factory outcome. Realizing this potential requires planning for the full MaxLPS target, including a facility design that supports 45°C cooling, even when day-one workloads with higher GPU-power requirements, such as training, offer less headroom than inference. Planning the AI factory to support the power, cooling, and network capacity of that target gives operators flexibility to expand quickly as the workload mix changes over time.
The next three tabs show how MaxLPS applies these connected optimizations at the site, rack, and workload levels: facilities and site-design innovations help make more of the fixed site-power envelope available to IT; Dynamic Power Software manages available IT power, including reclaiming unused allocation; and workload performance-per-watt optimization makes managed capacity more productive.
#### Facilities and Site DesignRealizing the full MaxLPS benefit—including up to 40% more GPU capacity within the same site-power envelope—requires a facility design that supports 45°C cooling. Where site conditions support efficient 45°C operation, more of the site-power envelope can be available to IT. That gives the project a basis to plan the full MaxLPS design point—power, cooling, East-West network capacity, and rack positions—before infrastructure choices are finalized, even when the MaxLPS GPU count will not be deployed on day one.
In many AI factory designs, site-power allocation is largely fixed. That allocation must cover IT, power distribution, cooling, and other facility overhead; when the cooling system maintains the thermal design point with less power, more of the fixed site-power envelope is available to IT, and the reverse is also true.
The 45°C Technology Cooling System (TCS) design point can widen the opportunity for dry cooling and reduce cooling overhead when site and ambient conditions permit. Conditions differ by location and time of year; for example, a cooler location such as Ireland may need trim support only during the hottest part of the year, while an Arizona site may need trim support for much of the year and mechanical cooling during the hottest conditions. These operating patterns must be validated during facilities planning.
At the rack, the TCS is designed for a 45°C supply design point. A coolant distribution unit (CDU) transfers that rack-level heat to a separate facility cooling loop, which carries it to on-site heat-rejection equipment.
The diagram also shows a separate, lower-temperature air-cooling loop for CRAHs and air-cooled support equipment. This loop is distinct from the liquid TCS loop serving the liquid-cooled racks.
Plan the full MaxLPS design point before infrastructure choices are finalized, then deploy the GPU count needed for the day-one workload and expand within the planned site capacity as workload needs evolve.
NVIDIA Dynamic Power Software (DPS) reallocates unused allocation across managed GPU and rack groups within configured policies and the validated topology. This helps operators make deployed capacity more productive as demand changes and, where the site was planned for it, supports later population of additional GPU rack positions.
To protect the facility, static configurations allocate power capacity at individual racks for credible peaks, failures, and other operating extremes. Actual demand is less uniform: tenant activity, workload mix, and operating conditions mean that not every GPU and rack needs its maximum power at the same time. The gap between allocated capacity and observed use can strand power that could otherwise support productive AI compute. Across a fleet, workload-specific demand can sit below static rack allocations, creating headroom that MaxLPS can manage within the same fixed site-power envelope.
Using an approved Vera Rubin 227 kW MaxP Thermal Design Power (TDP) as the fixed per-rack allocation basis, the rack comparison below shows how workload demand below static allocation bands creates stranded rack power and how MaxLPS can manage that headroom within the same planned rack-power budget.
NVIDIA DSX MaxLPS uses DPS to monitor GPU and rack-level consumption and manage available power across configured GPU and rack groups within the fixed AI factory power budget, reallocating capacity within defined budgets as conditions change.
Validate the managed power topology, device model, and policy boundaries before using DPS. It can improve use of GPU rack positions populated on day one and support growth into space the site was designed to accommodate as power is recovered when the workload mix evolves.
MaxLPS-managed capacity creates the opportunity to improve useful inference output from the AI factory GPUs. Workload and serving choices determine how productively that capacity is used for the intended service objective.
Once NVIDIA DSX MaxLPS manages available power within configured policies, workload and serving choices determine how productively inference services use that capacity. Model behavior, traffic patterns, serving topology, and performance targets can affect GPU power demand. Power-aware orchestration can also coordinate inference services with MaxLPS-managed capacity—for example, with NVIDIA Dynamo. For supported inference workloads, GPU power profiles can align configured power to the workload’s power-performance objective.
Apply the checks below to validate workload fit and serving behavior before committing capacity to the intended workload.
AI factory operators should plan for MaxLPS from the start. Site power is constrained, and recovering stranded power can improve AI factory performance per watt within the available power envelope.
Current support includes Blackwell HGX platforms (B200 and B300), NVL72 platforms (GB200 and GB300), and Vera Rubin NVL72. NVIDIA plans to add support for Rubin HGX and additional platforms over time.
MaxLPS can be evaluated for both existing and planned AI factories. New-site planning can build in capacity optionality early. For existing sites that have stranded power, an assessment of cooling, networking, and space is required.
MaxLPS should be used for site-level system planning, including rack population, the operating power envelope, and peak concurrency. Local electrical distribution and rack-side cooling capacity should be sized for MaxP, while the facility water loop should be sized for MaxQ. This layered approach provides the local peak capacity needed to allocate and cap power dynamically while keeping total site demand within the planned power envelope.
NVIDIA DSX MaxLPS uses DPS to manage available power across configured GPU and rack groups within defined policies and budgets. Facility controls and electrical protections continue to operate and protect the physical power and cooling infrastructure; MaxLPS does not replace those systems.
Review the DPS documentation to understand the dynamic power-allocation model and deployment requirements. DPS can also be installed as a developer preview. For site-specific deployment and support guidance, contact your NVIDIA representative.
A MaxLPS pilot or a DPS developer-preview pilot begins with a bounded scope and a validated deployment plan. This includes confirming the expected inference-focused workload mix and operating objective to gauge fit, as well as a supported management environment, a validated power topology and device model, usable telemetry, and operator-approved policies, targets, and recovery steps.
#### Next StepsReady to evaluate MaxLPS for your AI factory? Follow these steps and work with your NVIDIA representative to access NVOnline resources and plan your site-specific deployment.
Engage NVIDIA early to work through the planning sequence and the reference-design resources below.
Use these resources to scope the power-management approach for a MaxLPS deployment or pilot.
NVIDIA Dynamic Power Software Learn the dynamic power-allocation model through the SDK documentation and use the DPS guidance and the available developer preview to scope a site pilot for GPU and rack power management.Use workload-aware power profiles as a starting point for advanced performance-per-watt tuning.
Workload Power Profiles See how workload-aware power profiles can be a starting point for advanced performance-per-watt tuning. #### GlossaryPlain-language definitions for terms used across the MaxLPS overview.
| Term | Definition |
|---|---|
| Power Usage Effectiveness (PUE) | A measure of how much facility power reaches IT equipment compared with the total power used by the data center. |
| MaxP | MaxP is the highest GPU power setting, applied through static power allocation. It is used to design rack- and row-level power and cooling infrastructure. In practice, use it when maximum throughput matters and the site’s power envelope allows it. |
| MaxQ | MaxQ is the GPU setting that provides the best application performance per watt, applied through static power allocation. It caps GPU power below MaxP but, by itself, does not coordinate power across the AI factory. |
| NVIDIA DSX MaxLPS | An AI factory framework that combines facilities and site design, DPS, and advanced performance-per-watt techniques to maximize AI factory performance per watt and throughput per megawatt, improving efficiency and utilization within fixed AI factory power envelopes. Unlike static GPU-level settings, MaxLPS coordinates power dynamically across the AI factory. |
| Thermal Design Power (TDP) | A specified power level used for thermal and infrastructure planning for a product or rack operating configuration. It is not a prediction of observed workload power. |
| Technology Cooling System (TCS) | The rack-level cooling loop that carries heat from liquid-cooled IT equipment to the CDU. |
| Coolant distribution unit (CDU) | A heat-exchange system that transfers heat from the rack-level TCS loop to the facility cooling loop. |
| Dry cooling | Heat rejection to ambient air through equipment such as dry coolers, reducing reliance on mechanical cooling when site and ambient conditions permit. |
| Mechanical cooling | Cooling that uses equipment such as chillers when dry cooling alone cannot meet site conditions. |
| Dynamic Power Software (DPS) | NVIDIA software that monitors power consumption and manages available power across configured GPU and rack groups within defined policies and budgets. |
| GPU power telemetry | Power-consumption data reported by GPUs and used by DPS to verify allocations and manage available capacity. |
| Workload power profiles | Pre-tuned GPU configurations selected to better match supported workloads and their power-performance objectives. |
| East-West network | The GPU Compute network that carries traffic between GPUs and racks within the AI factory. |