NVIDIA DSX MaxLPS Overview
MaxLPS Overview
Facilities and Site Design
Dynamic Power Software
Workload Perf Optimization
FAQ
Next Steps
Glossary
Design and operate AI factories for more throughput per megawatt
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.
This shows the target architecture. MaxLPS reclaims stranded power inside the site. Flex and Exchange integration is not a prerequisite for deploying MaxLPS; it works on its own. In a greenfield site, plan for all three together.
Select a component or connection to view its details.
Grid / utility
What it does
Supplies the site’s power and sets the rules for using it, including programs that ask large loads to flex when the grid is constrained.
What comes in
Status from the site, through Flex, showing whether it responded as committed.
What goes out
Grid signals and utility commitments, such as a request to reduce load for a set time.
What the customer gets
Not a customer itself, but a site that can flex reliably is easier for a utility to approve and connect.
DSX Flex: sets the power target
What it does
Interprets grid signals and the site’s utility commitments, and turns them into a time-bound load target for the site.
What comes in
Grid signals and utility programs, usually through a grid partner (a grid ISV).
What goes out
A load target onto DSX Exchange, and status back to the grid partner showing how the site responded.
What the customer gets
A credible way to commit to grid flexibility. Depending on the jurisdiction, that can mean more approved megawatts, or power sooner.
DSX Exchange: carries the signals
What it does
The communication layer between IT and facility (OT) systems: an event bus plus shared schemas (the Flex power schema and the BMS schema), plus an agent gateway for AI agent tools.
What comes in
Load targets from Flex, and power, cooling, and energy readings from facility systems.
What goes out
The same signals, in a shared format, to DPS and to partner software.
What the customer gets
Integrate IT and facility systems once instead of building custom connections for every vendor. Partners plug in through the schemas.
DSX MaxLPS: keeps GPUs productive
What it does
Keeps total GPU power inside the approved envelope, while moving unused headroom to the GPUs that need it. DPS (Dynamic Power Software) runs this loop every few seconds.
What comes in
The site power budget and targets, facility signals through Exchange, and live power telemetry from the GPU racks.
What goes out
Per-GPU power limits, applied at each server’s BMC (its management chip) over Redfish.
What the customer gets
More GPUs inside the same fixed power, and more tokens per watt. The site can also respond to a grid request without stranding compute.
GPU racks
What it does
Run the customer’s AI workloads. Each server’s BMC accepts power limits and reports power use.
What comes in
Power limits from DPS.
What goes out
Power telemetry back to DPS, so it can confirm the fleet stays in budget.
What the customer gets
Workloads keep running on their existing scheduler. DPS manages power, not job placement.
Facility systems
What it does
Cooling, building management systems (BMS), electrical monitoring, and onsite energy like batteries (BESS) and generation.
What comes in
Not shown here: their own control systems keep running the building.
What goes out
Power, cooling, and energy readings onto Exchange through the BMS schema.
What the customer gets
Facility conditions, like cooling headroom, can shape how much power the GPUs get.
Grid signals: Grid to Flex
Requests and commitments from the utility, such as reduce load by a set amount for a set time. Usually delivered through a grid partner.
Status: Flex to grid
Reported site response, so the grid partner can see whether the site did what it committed to. Site metering and reconciliation decide program compliance.
Load target: Flex to Exchange
A time-bound power target for the site, published in the Flex power schema.
Facility readings: facility to Exchange
Power, cooling, and onsite energy readings, published in the BMS schema. When partners follow the schema, integrations work across the stack.
Signals: Exchange to MaxLPS
Targets and facility conditions delivered to DPS, which decides how to allocate power across GPU groups within the budget.
Power limits: MaxLPS to GPUs
Per-GPU power limits, set at each server’s BMC over Redfish, on a loop of seconds. Millisecond spikes are handled by hardware and electrical protection, not DPS.
Telemetry: GPUs to MaxLPS
Live GPU and rack power use, so DPS can find unused headroom and confirm the fleet stays inside the approved budget.
Up to 40% more GPUs within the same site-power envelope
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.
Deliver more inference throughput from a fixed power budget
Compare the MaxP, MaxQ, and MaxLPS reference modes. Selecting a mode updates the GPU capacity and token throughput relative to MaxP for the illustrative one megawatt inference scenario.
Use the left and right arrow keys to move between modes. Home and End move to the first and last mode.MaxLPS
GPU capacity per fixed 1 MW power basis400 GPUsIllustrative Vera Rubin NVL72 inference scenario from Sizing DSX AI Factory Infrastructure for MaxLPS (NVOnline #1161311). Values assume PUE 1.1. Vera Rubin testing remains in progress.
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.
