Is a cold plate blocked? Compare pressure drop with GPU temperature

Is the cold plate blocked, or is the GPU simply hot?

A hydraulic restriction changes the pressure-flow relationship; GPU temperature supplies its thermal consequence. Match inlet temperature and workload power to separate warmer supply, increased load and deteriorated contact from a flow fault.

In Anvilfield's interpretation of OCP Cold Plate Cooling Loop Requirements Rev 2, added branch resistance raises pressure drop at fixed flow and reduces branch flow at fixed manifold differential pressure. Use this hydraulic relationship to interpret measurements; the installed assembly's accepted envelope sets its limits.

This investigation covers pumped, single-phase water and inhibited water-glycol systems. Identify series versus parallel paths using the direct-to-chip architecture guide: series plates share flow; parallel passages can conceal maldistribution. Two-phase and passive loop-heat-pipe assemblies require different models.

Why normal manifold pressure does not clear a branch

A common-header pressure controller can maintain its setpoint while one restricted branch loses flow. Total CDU flow can conceal that loss, as the shared-manifold discussion in OCP Loop Requirements Rev 2, section 8, explains. At fixed pump speed, both pressure and flow can move with the network operating point; record the actual control mode.

Raising pump speed until an alarm clears does not establish recovery. Retest against the accepted pressure budget and the unaffected branches' operating limits.

What must you measure before touching the branch?

Map CDU supply, row supply, supply quick disconnect, hose, plate or series plates, return connector and row return. Locate pressure taps, branch flowmeter, inlet/outlet probes and GPU UUID together. Retrieve the CDU manifold baseline from CDU commissioning.

Record fluid product, batch, concentration and temperature, component revisions, flow direction, valve positions and pump control mode. Use a calibrated local meter within its supported tube/fluid/range combination, a suitably ranged differential-pressure sensor and matched temperature sensors. Purge sensing lines, check zero and tap elevation, and retain installation and calibration effects in the measurement uncertainty record.

How long should the test run?

Plan a five-minute stable analysis window at each point after temperatures settle, with one-second synchronized records where supported: these are Anvilfield-selected test settings. OCP's qualification white paper, section 1.3.2.1, requires constant processor power until case and inlet temperatures stabilize, rather than a prescribed fixed dwell. Allow stabilization time in addition to the recording window.

For this worked protocol, Anvilfield selects inlet-temperature span no greater than 0.5 °C, power within ±2% of target and no monotonic thermal rise across five minutes. Stop a load ramp at the actual equipment limit or approved protective stop, whichever comes first. Retain each signal's update rate, averaging period and timestamp offset: one-second polling does not create fresh one-second data from a slower sensor. Resolve stale values with the telemetry calibration procedure.

How do you build and compare branch resistance curves?

Measure several approved flows at matched fluid conditions, retaining pressure readings, repeatability and fit residuals. Plot clean, suspect and replacement points together: OCP Loop Requirements Rev 2 calls for pressure-flow and thermal-resistance-flow data, and the qualification white paper tests multiple flows.

Method note: Anvilfield estimates pressure by multiplying the selected resistance coefficient by flow squared: Δp = Kq². Divide measured pressure drop by squared flow to compare this model's coefficient. At 20 kPa, take the square root of 20/K for the operating flow. The restricted operating flow is 29% below the clean value.

Keep the diagnostic pressure boundary identical. Place any temporary flow-control valve outside those taps or subtract its independently characterized loss; otherwise its changing position manufactures a false curve. Real curves can contain linear and quadratic terms: OCP Universal Quick Disconnect Specification v2.0.0 (September 22, 2026), section 9.1.4, uses a best-fit relationship Δp(Q) = AQ² + BQ. A quadratic-only fit should stay within its demonstrated fluid and flow range.

What changes when the coolant temperature or glycol concentration changes?

OCP's qualification white paper, section 1.2.1.3, shows temperature-dependent flow for 25% propylene glycol. Use the exact supplier's matched-fluid curve or a measured baseline at the same temperature; a water-only connector coefficient needs the alternative-fluid correction or test data addressed by OCP UQD v2.0.0, footnote 7.

What if there is no branch flowmeter or clean baseline?

Obtain local flow with a qualified meter or characterized bench; estimating flow from the old curve assumes the restriction away. Without a commissioning curve, use an exact supplier curve or accepted identical assembly, retaining configuration and uncertainty. Otherwise report suspected restriction and proceed to component testing. Pressure ratios cannot reveal percentage of channels blocked.

Is the restriction in the plate, hose, valve or quick disconnect?

Inspect kinks, valve position, flow direction, connector seating and both QD halves against the approved assembly. A familiar UQD size does not establish its coefficient; use the QD selection guide for interface and pressure-budget selection.

OCP's undated Cold Plate Development and Qualification white paper, section 1.3.2.2, determines plate pressure drop by subtracting bypass-loop loss from plate-installed-loop loss at the same flow. Apply that bench method through approved service access and a clean characterized fixture. Re-characterize ancillary loss whenever hose, valve, connector or bypass geometry changes.

The QD-fault row settles an expensive ambiguity: high branch loss can coexist with a normal plate contribution. Pressure-integrity testing answers a separate question. A connector's working or burst pressure is neither a blockage threshold nor authorization to pressure-test an installed GPU assembly at that value.

Worked manufacturer check: how much loss can the connectors add?

Anvilfield estimates connector loss from the minimum Cv values published in Gates UQD specification sheet (©2025): 0.54 for UQD02 and 1.53 for UQD04. Select water at specific gravity 1.00 and flow 1.00 L/min. Following OCP UQD v2.0.0's coefficient method, divide flow in US gpm by Cv, square the result and multiply by specific gravity to obtain psi.

Method note: hold published minimum Cv constant at the selected flow; use 3.785411784 L per US gallon and 6.894757 kPa per psi, then round results. Two connected sets mean complete mated supply and return paths. Verify actual curves for low-flow/viscous-fluid service. Use the supplied connector revision for its compliance record.

How do you calculate cold plate thermal resistance?

Subtract plate inlet temperature from component case temperature, then divide by applied component power at a stated stabilized flow. OCP's undated qualification white paper, section 1.3.2.1, defines R = (Tc − Tin)/P, in °C/W, for the assembly mounted with its recommended TIM2 thermal-interface material. That case-to-inlet result includes the specified interface.

Preserve the measurement endpoints. Boyd Hi-Contact datasheet BC 06.2022 plots mounting-surface-to-inlet-water resistance, while Boyd Liquid Cooled Cold Plates datasheet (©2020) gives 0.02 °C/W as a typical product-family value. Neither is a universal GPU acceptance limit. In series paths, use the plate's local inlet; rack supply temperature can miss heat added by upstream processors.

What can GPU telemetry tell you about thermal resistance?

Call a ratio using GPU sensor temperature and reported GPU/board power apparent GPU-to-inlet resistance. Its boundary includes package, interface, spreading, plate and coolant effects. Preserve the exact power-field definition: reported board power is different from module power, memory-subsystem power and applied heater power. TDP is not a measured denominator.

Anvilfield estimates an apparent GPU-to-inlet resistance of 60 K/kW for a selected 65 °C GPU, 35 °C inlet and 500 W reported board power, rising to 90 K/kW when the GPU reaches 80 °C at the same inlet temperature and power. Divide the temperature difference by power in kW to reproduce the 50% increase. Kelvin and Celsius temperature differences are numerically equal.

Method note: Anvilfield estimates uncertainty using NIST Technical Note 1297 (1994), with selected independent standard uncertainties of 0.5 K for GPU temperature, 0.2 K for inlet and 10 W for power. Root-sum-square propagation gives standard uncertainties of 1.61 and 2.10 K/kW; selected coverage factor k = 2 gives expanded uncertainties of 3.22 and 4.20 K/kW. Use actual sensor uncertainty, repeatability and covariance for acceptance; the selected inputs are not NVIDIA accuracy specifications.

Thermal slowdown can reduce power enough to hide deterioration behind a temperature plateau. Compare a controlled lower load both assemblies sustain or the same pre-slowdown interval, with workload throughput retained.

Does coolant temperature rise prove a blockage?

Coolant temperature rise reflects heat pickup relative to mass flow. At steady conditions, multiply coolant mass flow by fluid-specific heat and outlet-minus-inlet temperature to calculate heat carried across that boundary, using supplier properties at measured concentration and temperature. OCP Loop Requirements Rev 2 describes heat split and fluid-property boundaries. Greater rise can follow less flow or more heat; throttling can reduce both power and rise. Board power is not automatically the heat reaching one plate's coolant.

Does the GPU telemetry show thermal throttling?

Capture UUID, model, driver, utilization, SM/memory clocks, measured power and enforced limit, supported temperatures, throughput, event reasons and thermal limits. Tie readings to the physical plate; clock reduction alone does not identify a cooling fault.

NVIDIA's nvidia-smi documentation (live edition checked November 5, 2026) distinguishes SW Thermal Slowdown and HW Thermal Slowdown from SW Power Capping and HW Power Brake. HW Slowdown alone combines thermal and power-brake causes. Use supported GPU Max Operating Temp and Slowdown Temp values from the actual device; a blanket 80 °C or 85 °C threshold would misstate model-specific behavior. N/A means unavailable evidence.

NVIDIA DCGM 3.0 Field Identifiers defines DCGM_FI_DEV_THERMAL_VIOLATION as thermal violation time in microseconds. Plot its increment over the test interval, preserving collector units and marking resets or restarts as discontinuities. An unavailable counter is not a zero-event interval.

How do you capture the GPU evidence?

Verify query fields with the installed driver's help, then replace GPU-UUID with the affected physical UUID for this read-only capture.

nvidia-smi -L
nvidia-smi --help-query-gpu
nvidia-smi -i GPU-UUID -q -x > gpu-before.xml
nvidia-smi -i GPU-UUID --query-gpu=timestamp,uuid,temperature.gpu,power.draw,clocks.sm,utilization.gpu --format=csv -l 1 > gpu-trend.csv

Stop recording after the window and capture gpu-after.xml with the same full query. Add supported events/counters and memory temperatures from DCGM/NVML or the collector; retain software versions and power-field definition. NVIDIA defines Average Power Draw as preceding-second board power on Ampere (except GA100) or newer devices. Record whether the selected field is averaged or instantaneous, and align hydraulic averages to that interval.

Align the hydraulics and workload on one timeline

Plot hydraulics, temperatures, power/throughput and thermal events on synchronized axes. Annotate workload/pump/service changes and averaging periods. Use the GPU burn-in workload guidance for repeatable loading. A neighbor is a valid comparison only when hardware, firmware, workload, series order and supply conditions match.

How do you tell a contact fault from a flow fault?

Resolve mapping and uncertainty, then use the matrix to separate hydraulic restriction from contact, deposits and maldistribution.

Normal bulk flow does not establish clean heat-transfer surfaces. A thin insulating deposit can degrade heat transfer before its hydraulic effect is measurable. Partial closure of parallel microchannels can redirect flow away from a hot region while total branch flow remains near baseline. Compare matched-flow thermal behavior, then use approved component examination to distinguish those paths from mounting loss.

State the narrowest supported finding. “Branch restriction” follows a changed assembly curve; “restriction localized to the plate” requires component evidence. “Particle blockage confirmed” needs recovered particles, inspection or supplier analysis. A temperature correlation alone cannot identify the obstructing material.

Where does trapped gas fit?

OCP's qualification white paper requires a bubble-free loop before thermal testing. De-air under the loop commissioning procedure, then repeat the stable comparison. Persistent hydraulic penalty after correct de-airing keeps solid restriction in the investigation. Intermittent behavior alone does not distinguish gas, debris or cavitation; record the response to the approved procedure.

What should the contamination register record?

Preserve the as-found condition before flushing or teardown redistributes the evidence. Link fluid and deposit samples, removed filters and the plate serial number to the same incident. Photograph recovered material before alteration; use “not analyzed” where composition is unknown. Dark deposits require analysis before being called copper oxide, scale or biofilm.

For every entry record incident/time, loop/branch/GPU, fluid/batch/concentration, source location, sample/filter/part ID, collection procedure/organization, lab-specified handling, chain of custody, method/revision, result/units, product limit, report ID, corrective action and retest link.

OCP's PG guideline Rev 2.0/Version 1.0.0, section 11.2, calls for a representative well-mixed circulating-fluid sample and discarding stagnant sampling-port fluid before collection. Label deliberate low-point or deposit samples separately: they answer different questions from circulating-fluid health. Select microscopy, elemental/chemical analysis or matrix-suitable microbiology to resolve the suspected material.

Correct the contamination source before reconnecting the replacement. Vertiv Guidelines for Deployment of Single Phase Technology Cooling System, SL-80056_REVA_02-25 (©2025), sections 6.1 and 6.4, addresses cleaning third-party equipment and establishing cleanliness before IT connection. Connect findings to coolant chemistry and treatment and wetted-material/leachables qualification.

Which filtration and chemistry numbers actually apply?

Vertiv SL-80056_REVA_02-25, section 6.4, specifies strainer rating at 10% to 50% of the narrowest channel and describes full-flow filtration typically at 25% to 40%. For a selected 100 µm channel, Anvilfield estimates 10 to 50 µm and 25 to 40 µm respectively by multiplying channel width by those fractions. Record nominal/absolute efficiency, location and clean/dirty pressure loss alongside the micron rating.

OCP Guidelines for Using Water-Based Transfer Fluids (undated final edition) recommends side-stream filtration below 5 µm and describes component-approved inline examples such as 50 µm. Distinct locations and duties explain those different values.

OCP PG Rev 2.0/Version 1.0.0 gives new-fluid pH in Table 1 and different in-service OAT/hybrid guidance in Table 14. OAT means organic acid technology; IAT means inorganic acid technology. Use the fluid supplier's maintenance range for the installed formulation when judging a circulating sample.

Resolve incompatible instructions for the exact assembly and coolant. Choosing the numerically narrower pH range or a stronger cleaner without checking inhibitor chemistry does not establish compatibility. Use the approved product program rather than applying an untreated-DI-water conductivity target to inhibited operational coolant.

Should you clean the plate or replace it?

Replace an assembly that leaks, is sealed and non-serviceable, or cannot be restored by approved cleaning to accepted hydraulic and thermal performance. This is Anvilfield's service decision logic; the purchased assembly's manual determines serviceability. Preserve samples and the removed part before cleaning changes the evidence.

Remove the affected assembly from load under its approved service procedure when flow, leakage, thermal limits or protective alarms require it. Use the concurrent-maintenance isolation plan to preserve cooling service and retain leak detection and intervention results.

Vertiv's 2025 deployment guide, section 6.1, restricts chemical cleaners/detergents except with new-fluid-supplier guidance; OCP's PG guideline, section 10, describes controlled cleaning and rinsing. Resolve one compatible component/fluid procedure before work. A diagnosis does not supply a generic acid dose, backflush pressure or ultrasonic duty.

Vertiv Rev A, section 1, requires de-energized maintenance and safe component temperatures. For servicing covered by OSHA 29 CFR 1910.147(d)(5) and (d)(6), checked November 5, 2026, control stored energy and verify isolation before work. Use the assembly procedure to isolate cooling and relieve trapped hydraulic pressure before opening it.

OCP UQD v2.0.0, section 9.1.5, calls for progressive filling and draining within ratings to prevent seal washout.

Record failed and replacement serials, revisions, lots, connector changes, mounting hardware and interface materials. Use the processor/assembly specification for torque and compression. The OEM acceptance and responsibility guide defines supported configuration and service authority.

How do you retest the replacement cold plate?

Repeat the accepted pressure boundary and flow points with the same fluid product, concentration and temperature, inlet reference, power definition, workload and firmware. Complete approved leak checks and de-airing before loading. Retest once at accepted header pressure to prove restored service, and again at matched flow/power to separate remaining thermal degradation from reduced-flow effects.

Tables 2 and 5 return from 0.71 to 1.00 L/min at 20 kPa and from 90 to 60 K/kW: Anvilfield estimates of paired recovery. Itemize every changed hose, connector, plate and interface; simultaneous changes leave component cause unresolved.

Before testing, enter the model-specific minimum flow, pressure range, inlet-temperature range, device thermal limits and baseline comparison tolerance, each with source/revision. Compare corrected readings with the accepted supplier curve or clean baseline using the uncertainty and decision-rule method.

Anvilfield selects a conservative guard-band rule: measured flow minus expanded uncertainty must meet the specified minimum flow; measured pressure drop plus expanded uncertainty must stay at or below the specified maximum drop. A result overlapping the limit is held for better measurement or corrective work. For a selected minimum of 0.90 L/min and measured 0.94 L/min with expanded uncertainty 0.06 L/min, Anvilfield estimates a lower bound of 0.88 L/min: hold release. NIST TN1297 supplies uncertainty methods, not this acceptance rule.

Release requires expected power and throughput, no new thermal slowdown during the matched workload, stable temperature, acceptable flow, no leak and correct protective alarm response. A cooler idle GPU leaves the delivered-load requirement unresolved.

How often should you look again?

Anvilfield selects post-repair reviews at the stabilized retest, 24 hours, 7 days, 30 days and first comparable high-load event. For each checkpoint retain UTC window, component IDs, fluid condition, inlet/flow/pressure/GPU/power readings, apparent resistance, event increment/reset flag, throughput, criterion/revision, pass/hold and raw-file link. Repeat invasive curve sweeps when the approved program or changed trend calls for them.

OCP's PG guideline Rev 2.0/Version 1.0.0, section 11.2, recommends at least annual qualified-laboratory analysis generally. Its low-risk schedule is commissioning baseline then annual; medium risk uses baseline then six-monthly laboratory testing for 18 months, reducing to annual only if stable; high risk uses baseline then quarterly for 18 months, reducing to six-monthly only if stable. Medium risk separately includes six-monthly on-site checks.

Those scenarios retain their assumptions: low risk includes inline pH/refractive monitoring, acceptable materials and a familiar user; medium lacks inline monitoring and has an unfamiliar user; high additionally describes materials as “mostly acceptable.” That sampling description does not approve an incompatible material.

For loops below 100 gallons, section 11.2 separately recommends startup pH and glycol-concentration/refractive-index verification at every rack, then qualified-lab analysis of a statistically significant rotating subset at least every six months where regular every-rack testing is impractical. Similar racks and wetted materials are prerequisites. Document population, selection and coverage; no universal sample count is supplied. Take incident samples promptly, and reopen shared-fluid/filtration/material investigations if restriction recurs.

What belongs in the service scope and cost estimate?

Separate workload coordination, instrument setup, curve testing, sample handling/analysis, OEM service, consumables, leak/de-air checks, witnessed replacement retesting and follow-up reviews. Keep plate cost, laboratory quote, lead time and downtime value separate. Use the commissioning subcontract and retest guide to define attendance and repeat visits.

A public cooling scope should separate branch diagnosis, coolant analysis and witnessed replacement retests; PorchRocket's Bid Control finds public HVAC and refrigeration bids and early capital-plan signals where contractors price that work.

Anvilfield estimates eight technician-hours for a selected two-person, four-hour diagnosis/retest visit, plus four technician-hours for a separate two-hour repeat visit. Method: multiply crew size by elapsed time. Add actual stabilization, service, laboratory turnaround, delivery and outage allowances to the work order; instrument access and workload coordination control the schedule.

Who decides whether the branch is ready for service?

The server/plate supplier governs assembly limits, the coolant supplier governs its fluid's treatment constraints, and the owner accepts the restoration evidence against approved project criteria. Name who supplies workload, isolates cooling, witnesses results, retains the failed part and funds repeat tests. Record “repair scope changed” separately from “same plate retested.”

Field checklist and closeout record

Package raw trends, curves, uncertainty, samples, photographs, part changes, limits, exceptions and signed retest results together. Enter results and disposition beside each item; assign unresolved work.

1. Identify GPU UUID, plate, branch, loop and part revisions; correct unmapped telemetry. Result: ______ Disposition: ______
2. Retrieve clean hydraulic/thermal curves and operating limits; resolve missing release criteria. Result: ______ Disposition: ______
3. Draw pressure boundaries, including hoses, valves and connected QDs between taps. Result: ______ Disposition: ______
4. Confirm zeros, sensing-line fill/elevation, meter range, mapping and calibration; reject invalid readings. Result: ______ Disposition: ______
5. Record fluid product/batch/concentration/temperature; match baseline conditions. Result: ______ Disposition: ______
6. Capture header pressure, filter loss and neighbor behavior; investigate shared faults first. Result: ______ Disposition: ______
7. Synchronize supported temperatures, power, limits, clocks, utilization, throughput and event reasons. Result: ______ Disposition: ______
8. Stabilize supported load; retain analysis window and protective stop conditions. Result: ______ Disposition: ______
9. Compare loss at matched flow or flow at matched pressure; identify controlled condition. Result: ______ Disposition: ______
10. Inspect routing, valves and QD seating/identity; record changed parts. Result: ______ Disposition: ______
11. De-air by approved procedure; repeat unstable measurements before diagnosing solid blockage. Result: ______ Disposition: ______
12. Localize plate loss with characterized bench/bypass; protect IT from dirty test equipment. Result: ______ Disposition: ______
13. Retain fluid, debris and removed part with lab-agreed handling; distinguish unconfirmed identities. Result: ______ Disposition: ______
14. Use approved cleaning/replacement/mounting; record assembly and interface material. Result: ______ Disposition: ______
15. Retest matched conditions, expected workload, protection behavior and leak/alarm results. Result: ______ Disposition: ______
16. Sign restoration decision; assign mechanism follow-up, operational reviews and risk-based fluid analysis. Result: ______ Disposition: ______

Close restoration and mechanism findings separately. The replacement can pass its service test while deposit identity remains pending; retain the removed part and assign that analysis explicitly. Link the closeout to the branch curves, fault matrix, contamination register and replacement decision.

Sources checked and reviewed by Anvilfield: 2026-11-05. Update log: 2026-11-05: reviewed source editions and calculations; distinguished new-fluid and in-service pH, clarified thermal events and service controls, and added a flow-acceptance example.

The primary-source register below identifies editions and application boundaries. The OCP Cold Plate project index is the discovery index; use the linked document edition for technical claims. Retain queried device, driver and collector versions with each service record. Cite the specific table fragment when reusing a calculation or decision register.

How to cite this page: Anvilfield, Is a cold plate blocked? Compare pressure drop with GPU temperature, 2026, https://anvilfield.com/field-guides/datacenter/cold-plate-blockage-pressure-drop-diagnosis/

Related tools

FAQ

People also ask

• What is direct-to-chip liquid cooling?
• What is a CDU in a data center?
• What coolant is used in data center liquid cooling?
• What is direct-to-chip liquid cooling?
• How do I select OCP liquid-cooling quick disconnects for pressure and compatibility?

Sources checked

• S1. Open Compute Project, ACS Liquid Cooling Cold Plate Requirements, Rev 1.0, October 9, 2019. Component selection, geometry-specific filtration and filter pressure loss. Checked 2026-11-05.
• S2. Open Compute Project, Cold Plate Cooling Loop Requirements, Rev 2, cover undated. Single-phase pressure-flow and thermal-resistance-flow data; section 8 covers shared manifolds. Checked 2026-11-05.
• S3. Open Compute Project, White Paper: Cold Plate Development and Qualification, undated public edition. Sections 1.2.1.3, 1.2.2, 1.3.2.1, 1.3.2.2 and 1.3.3.2: fluid properties, mounting/TIM, stabilized testing, bypass subtraction and analysis. Checked 2026-11-05.
• S4. Open Compute Project, Guidelines for Using Water-Based Transfer Fluids in Single-Phase Cold Plate-Based Liquid-Cooled Racks, undated final public PDF. Water-program pH, side-stream filtration and component-approved inline filtration. Checked 2026-11-05.
• S5. Open Compute Project, Guidelines for Using Propylene Glycol-Based Heat Transfer Fluids in Liquid Cooled Racks, Revision 2.0, Version 1.0.0, September 15, 2026. Tables 1 and 14 distinguish new-fluid and in-service pH. Sections 10 and 11 cover cleaning and analysis; pages 61 to 64 cover sampling and risk schedules. Checked 2026-11-05.
• S6. Open Compute Project, Cold Plate subproject, live index checked 2026-11-05. Live document-discovery index, accessed November 5, 2026. Checked 2026-11-05.
• S7. NVIDIA, NVIDIA System Management Interface documentation, live HTML edition checked 2026-11-05. Live manual: temperature limits, sensor availability, event reasons and board/module power definitions. Checked 2026-11-05.
• S8. NVIDIA, DCGM 3.0, Field Identifiers. Versioned field definitions: temperatures, power, device limits and thermal-violation counter units. Checked 2026-11-05.
• S9. Boyd, Hi-Contact 2-Pass Cold Plate Technical Datasheet, BC 06.2022. Mounting-surface-to-inlet-water thermal endpoint and paired hydraulic/thermal curves. Checked 2026-11-05.
• S10. Boyd, Liquid Cooled Cold Plates Technical Datasheet, ©2020. Product-family typical thermal resistance and coolant/material configurations. Checked 2026-11-05.
• S11. NIST, Technical Note 1297, Guidelines for Evaluating and Expressing the Uncertainty of NIST Measurement Results, 1994 edition. Uncertainty propagation and expanded uncertainty. Checked 2026-11-05.
• S12. Open Compute Project, Universal Quick Disconnect Specification, v2.0.0, effective September 22, 2026. Section 9.1.4: water-relative Cv, test conditions, fixture subtraction and polynomial curves; footnote 7 covers alternative fluids. Checked 2026-11-05.
• S13. Gates, Universal Quick Disconnect technical specification sheet, ©2025. Published minimum Cv, working pressure and temperature envelope. Checked 2026-11-05.
• S14. Danfoss, Hansen Universal Quick Disconnect fact sheet, printed AI423953312825en-000106, September 2026. Printed pages 2 and 4: assembly temperature and size/state-specific working pressures. The printed document identifier differs from the URL suffix. Checked 2026-11-05.
• S15. Vertiv, Guidelines for Deployment of Single Phase Technology Cooling System, SL-80056_REVA_02-25, ©2025. Sections 6.1, 6.4 and 7: cleaning, filtration/channel ratios, filter monitoring and material controls. Checked 2026-11-05.
• S16. OSHA, 29 CFR 1910.147, The control of hazardous energy, current text checked 2026-11-05. Paragraphs (a), (d)(5) and (d)(6): covered servicing, stored energy and verification of isolation. Checked 2026-11-05.