MaxLPS Power Pilot#

Compare a three-rack baseline with a four-rack DPS-managed run under the same facility power limit. Run the same workload mix without DPS across three NVL72 racks, then with DPS dynamically managing GPU power across four racks. Compare throughput, latency, and power compliance.

To run this pilot, prepare: four GB200 NVL72 or GB300 NVL72 racks (use three for the baseline and all four for the managed run), BMC access, and the facility power limit to hold fixed for both runs. You also need two representative inference jobs—one lower-power and one higher-power—and metrics collection with a place to retain the results. For details, refer to pilot prerequisites.

The Five Pilot Steps#

  1. Get ready. Deploy DPS, authenticate dpsctl, and confirm the site and metrics. Follow the Deployment and Configuration Guide, pilot prerequisites, and metrics setup.

  2. Create the topology and check BMCs. Import the devices and four-rack topology, activate it, then check every BMC. DPS 0.9 requires an active topology for dpsctl check connection and dpsctl verify bmc-health start. Deactivate it and verify default GPU limits before the baseline. Follow Step 3.

  3. Run the baseline. Run the workload mix without DPS and capture the comparison metrics. Follow Step 4 to prepare the run and Step 5 to run it. If baseline nodes and metrics are ready, you can run it first. Keep the topology inactive until the baseline finishes, then activate it for BMC checks.

  4. Enable MaxLPS. Reactivate the topology, create the resource group with the default MaxLPS bundle, and confirm Power Steering is Ready with no reconciliation errors. Follow Step 6.

  5. Run the MaxLPS test and compare. Repeat the workload mix across four racks, confirm GPU limits move under load, then compare throughput, latency, and power compliance. Follow Step 7 to run the managed test and Step 8 to compare the results.

How the Pilot Fits Together#

flowchart LR A["3-rack baseline<br/>DPS out of control path"] --> C["Compare identical<br/>inference workloads"] B["4-rack managed run<br/>DPS in control path"] --> C D["Topology<br/>models racks and envelope"] --> E["Resource group<br/>activates the nodes for<br/>dynamic power management"] E --> B F["MaxLPS bundle<br/>and Power Steering"] --> E

  • A topology describes the four racks and the power domain that defines the operating envelope.

  • A resource group activates its compute nodes for dynamic power management.

  • The MaxLPS policy bundle and Power Steering use live telemetry to manage the resource group’s GPU limits. Do not create a pilot-specific policy bundle.

Continue With#