NVIDIA DSX Ready Qualifies the CDU. Who Checks the Loop?
If branch flow falls after a filter service, the operator needs a usable before-and-after comparison. The CDU qualification record and the site acceptance test are valuable, but neither identifies whether a restriction, pump setting, changed load, or measurement problem explains that reading.
The useful next step is to compare that branch with its accepted operating condition. That requires a record of the fluid, the measurement points, the pump state, and what changed during service. Product qualification becomes much more useful when those site records travel with it.
NVIDIA launched DSX Ready on September 21, 2026, initially covering battery energy storage systems and cooling distribution units (CDUs). For CDU providers, NVIDIA describes a self-qualification suite for determining whether a specific offering meets applicable functional requirements.
For the team buying and operating a CDU, that leaves three distinct questions: What offering is qualified? What did our installed system demonstrate? What has changed since that demonstration? Keep the answers together, with a named owner for each.
Put the exact offering in the purchase record
The qualification applies to a specific offering. A supplier name alone does not identify the model and configuration covered by the evidence. Put that detail beside the purchase specification.
Ask the supplier for the covered model and configuration, applicable conditions, and supporting CDU documentation. The launch also includes batteries, but their qualification criteria should not be used to explain a CDU. Keep the evidence specific to the category and offering being purchased.
The project engineer still needs to check the proposed fluid, temperatures, pressures, loads, controls, and maintenance arrangement against the project requirements. Record open questions before ordering. If a model or configuration changes later, review the substitution against the same record.
Name the two circuits before comparing readings
In NVIDIA's DSX facilities reference design, facility water cools a separate secondary loop serving GPU cold plates. Heat crosses the CDU exchanger; the two liquids normally remain separate.
That distinction matters during a handoff. Facility inlet temperature and equipment supply temperature describe different points. So do pressure differences across a filter, a rack branch, and the whole secondary circuit. A trend labelled only water temperature or differential pressure leaves the next person guessing.
Trace the heat. Keep the fluid records separate.
Heat travels from the chips into the secondary coolant, across the CDU plates, then into the facility circuit. The fluids normally remain separate.
Connect each important trend to a sensor ID, circuit, physical location, unit, and drawing reference. Keep the fluid specification attached to the correct circuit. A clear point map lets an operator follow a reading back to the pipework before deciding what it means.
The same CDU can serve very different branch conditions
The US Department of Energy explains the relationship between pump performance and system resistance in its pumping-system sourcebook. The surrounding system matters: pipework, fittings, branch geometry, and restrictions participate in the operating result.
Consider the simple model below. At a fixed 60 kPa pressure difference and 10 kW of heat, doubling branch resistance reduces flow from 20.0 to 14.1 L/min. Bulk coolant temperature rise increases from 7.2 to 10.1 degrees C. The CDU identity has not changed; the condition at the load has.
Same pressure. Less flow through a higher resistance.
The selected branch needs twice the pressure difference to carry the same flow.
Each branch: 60 kPa available pressure / 10 kW heat input
Reference
- Branch flow
- 20.0 L/min
- Coolant temperature rise
- 7.2 °C
- Return temperature
- 32.2 °C
Selected branch
- Branch flow
- 14.1 L/min
- Coolant temperature rise
- 10.1 °C
- Return temperature
- 35.1 °C
29% less flow carries the same heat. Coolant rise increases by 3.0 degrees C. Changing supply temperature shifts the return temperature; it leaves the modeled flow unchanged.
Model, units, and limits
Pressure loss = R x Q squared; Q = square root of (60 / R). Q is in L/min and R is in kPa per (L/min) squared. Reference R = 0.15; selected R = 0.30.
Temperature rise = P / (mass flow x heat capacity). P = 10,000 W; density = 1,000 kg/m3; heat capacity = 4,180 J/(kg K); mass flow = Q x density / 60,000. Return temperature = supply temperature + rise.
An ideal CDU holds 60 kPa across each branch. Two independent parallel branches have the same heat input. Fixed water-like properties, turbulent quadratic resistance approximation, no phase change, no heat loss, and no pump/control limit are assumed. This does not model a real CDU, glycol formulation, heat-exchanger approach temperature, transient behavior, or a loaded-filter curve.
Engineering background: DOE pumping-system sourcebookThose are calculated examples, not a prediction for a particular installation. Real pumps and controls respond, fluid properties vary, and shared pipework couples branches. The practical lesson is to retain the accepted branch condition with its load and control state. A CDU nameplate cannot supply that missing comparison.
When a rack is added or pipework is altered, document the change and compare the result with the approved project model. Keep the earlier reference. Stable readings alone are not a reason to overwrite it.
Our earlier article The Loop Is Running. Is It Ready for the GPUs? covers fluid, air, surface preparation, and pressure readiness before connection. The acceptance evidence from that work should remain available to the operating team.
After service, separate the observation from the explanation
Return to the lower-flow rack. The filter was serviced, but supply temperature and workload also changed. Makeup fluid may have entered the loop. The timing makes the service relevant; it does not identify the cause. First establish which conditions can be compared.
A service event is a clue. Build the comparison.
Keep a comparison the next shift can use
The installed branch has an accepted operating record.
- On record
- Flow, temperatures and pressure taps are recorded with pump mode, workload and fluid identity.
- Still missing
- There is no post-service comparison yet. Preserve this run as the reference.
- Next check
- Link the trends to the branch and sensor IDs on the as-built drawing.
Compare the same pressure taps and account for flow and pump state. Check the service and makeup record, then use fluid evidence from an appropriate sampling point. A pressure difference measured at a different flow is not, by itself, a filter condition report. Keep unknowns visible until the evidence resolves them.
NVIDIA describes point metadata, units, and control boundaries in its DSX Exchange BMS integration documentation. Even where a site uses a different integration, the engineering question is practical: can the next person identify what a point means and compare it with the right condition?
The next check belongs inside the approved site procedure, with equipment limits and return-to-service authority intact. The goal is a comparison that supports a decision, rather than a setpoint change made merely to quiet a trend.
For isolation and restart planning, see Liquid Cooling Reliability Is Now an AI Compute-Uptime Problem. Keep that plan connected to the operating record.
Try this on one CDU before the next handoff
Ask the person receiving the system to locate three records without help from the person who assembled them:
- The qualified offering: exact model, configuration, supporting evidence, and unresolved supplier questions.
- The accepted installation: as-built drawing, fluid identity, point map, test conditions, results, and signoff.
- The operating comparison: accepted branch trends, pump and load context, and the most recent service or makeup event.
If one is missing, name its owner and close that gap. If all three exist but cannot be connected by equipment and point IDs, fix that connection. This is a practical way to turn a procurement decision into evidence the next shift can use.
Reliability Engine focuses on connecting coolant condition with loop and thermal behavior. That work complements product qualification and site acceptance by helping operators understand what changed and what to investigate next.
For help bringing those operating records together, talk with Reliability Engine.
Subscribe to updates
Get the latest engineering perspectives sent straight to your inbox.