Data center liquid-cooling leak detection field guide

Leak detection, and why detecting is only half the job

Data center liquid-cooling leak detection is the system of sensors, alarms, and isolation that finds coolant where it should not be and acts on it before it reaches the IT. As AI and high-density racks pushed the industry to direct-to-chip cold plates and immersion, coolant now runs inches from energized electronics worth millions, so leak detection became a protect-the-load system rather than a facilities nicety. A leak that reaches a powered board is not a maintenance ticket. It is hardware loss and downtime.

Here is the part that gets missed: a detector that only alarms is not enough. By the time a person reads the alarm, walks the floor, and finds the source, the coolant has already spread. A loop under pressure does not wait for a technician. So the system has to detect and isolate, closing a valve to stop the flow at the source, on its own, in seconds, before the leak finds a live component.

Three things make that work, and the rest of this guide walks each one. Build sensor coverage at every place coolant can escape, the fittings and the trays and the low points. Tie that coverage to fast automatic isolation and the building controls so the loop shuts itself down and the room finds out. Then prove the whole chain in commissioning with a simulated leak, because a leak-detection system nobody tested is a system nobody can trust. The direct-to-chip cooling guide and the CDU commissioning guide cover the loop and the unit those sensors live on.

Detect and isolate: the rule the whole system serves

The single truth to carry off this page is that detection alone does not protect anything. You have to detect the leak and automatically isolate it, shutting the valve and stopping the flow, before the coolant reaches the IT. An alarm that lights a panel while coolant keeps pumping onto a server is a record of the failure, not a defense against it.

The arithmetic is what makes this blunt. A direct-to-chip loop under pressure moves real volume per minute, and a dense rack of GPUs sits directly below the manifolds and quick-disconnects feeding it. Studies of facilities with integrated detection put average human response at something like eight to twelve minutes from alarm to action. An automatic isolation valve acts in seconds. The gap between those two numbers is the gap between a contained drip and a flooded cabinet, and it is why the act half of the system matters as much as the find half.

So treat detection and isolation as one system with two jobs, not two projects. The sensor finds the leak and locates it. The valve stops it. The controls tell the room. Design any one of those in isolation and you have built something that looks finished and protects nothing. ASHRAE's liquid-cooling resiliency guidance, the cooling OEM, and the commissioning agent all frame the requirement the same way: find it and stop it, automatically, before it spreads.

Why does liquid cooling make leaks a primary risk?

Liquid cooling makes leaks a primary risk because the coolant now runs inches from energized electronics that air cooling kept dry. For decades the water in a data center stayed in the perimeter units and the pipes overhead, far from the boards. Direct-to-chip put a coolant loop on top of the processor itself, and immersion put the whole server in fluid, so the separation that used to protect the hardware is gone.

AI and high-density compute drove the change. A modern accelerator concentrates hundreds to over a thousand watts into a package the size of a deck of cards, and racks pack dozens of them into 80, 100, or past 120 kW. Air cannot pull that heat flux off the die, so the liquid is not optional, it is the only way the chips run. The direct-to-chip cooling guide covers why air ran out of room. The consequence for this guide is that the protection against leaks has to scale up with the value sitting under the loop.

The stakes are not abstract. A single GPU tray can be worth more than the truck the technician drove to site, and a slow weep onto it over a weekend, with nobody watching, can take out a row. Coolant exposure causes immediate shorts, corroded board traces, or slow oxidation that degrades parts over months. The damage starts the moment liquid touches a live trace, which is why the response has to be measured in seconds, not in the minutes a person needs to get there.

What coolants run in the loop, and which ones conduct?

Three coolant families show up in a liquid-cooled hall, and the leak risk depends on which one is in the pipe. The facility water loop is the building chilled or condenser water on the primary side, and it conducts. The technical or treated coolant is the clean secondary fluid that touches the cold plates, commonly a water-glycol mix like PG25, roughly 75 percent water to 25 percent propylene glycol, and it also conducts. The dielectric is the non-conductive fluid used in immersion and some two-phase loops, chosen so a leak onto live electronics is less catastrophic.

The conductivity is the whole story for leak risk, and it surprises people. The treated secondary coolant is not a dielectric. Its corrosion-inhibitor package, the chemistry that protects the mixed metals in the loop, raises its electrical conductivity by design, often well above 2,000 micro-mhos per centimeter. So a leak of PG25 onto a powered board behaves like a leak of water, not like an inert fluid. Propylene glycol is chosen over ethylene glycol because it is far less toxic around a populated building, not because it is any safer electrically.

Match the detection technology to the fluid. Resistive sensors find a conductive leak by the change in resistance when fluid bridges the sensing element, and they work well for facility water and for the glycol mixes most loops actually run, which sit in the 25 to 50 percent glycol range for direct-to-chip. Push the glycol concentration high enough, past roughly 75 percent, and conductive sensing stops working because the fluid no longer conducts, so a dielectric or a very high-glycol loop needs a different sensing method. Confirm the coolant and its concentration with the OEM before you pick the sensor.

Why conductivity decides how bad a leak is

A leak of conductive coolant onto a live board is a short circuit waiting for a path, and that is the difference between an inconvenience and a destroyed component. Water conducts. The treated glycol coolant conducts because of its inhibitor chemistry. When either bridges two points at different potential on an energized board, current flows where it was never meant to, and the result is an immediate fault, a scorched trace, or a slow corrosion that shows up as flaky hardware weeks later.

Dielectric fluid changes that calculation. Because it does not conduct, a drip onto a powered component is far less likely to cause an instant short, which is part of why immersion tanks tolerate the fluid touching everything. That does not make a dielectric leak harmless. You still lose containment, you still have fluid where it should not be, and you still have to find it and stop it. The dielectric buys time and lowers the odds of a catastrophic short. It does not remove the need for detection and isolation.

The speed of damage is the reason the response has to be automatic. A conductive leak can fault a board in the time it takes to spread across it, well inside any human response window. So the protection is layered by how fast each part acts: containment catches the first drips, detection sees them, and isolation stops the source, all before the fluid completes a circuit on something energized.

Where do leaks actually happen?

Leaks happen at connections far more than in the middle of a pipe or hose, and the connection count in a liquid-cooled rack is high. The quick-disconnects between the server and the manifold are the number one source, because they are the most-handled part of the system and the one that gets opened and closed every time a node is serviced. After the QDs come the fittings and the manifold joints, then the hoses where vibration and routing work a connection loose over months, then the cold plates themselves, then the valves and the CDU internals.

The pattern holds because joints are where the system is assembled by hand and where the seal depends on a person seating it right. A length of hose rarely fails in the middle. A coupling that was cross-threaded, not fully latched, or fouled with a trapped particle is a leak waiting for a vibration cycle to find it. This is why the sensor coverage and the maintenance discipline both concentrate on the connections, not the pipe runs.

Map the leak points before you place a single sensor. Walk the loop from the CDU out to the cold plates and list every connection, every low point where escaped fluid will collect, and every place a drip would land on something energized. That map is what the detection layout serves. Cover every fitting and every tray, because the leak you did not cover is the one that finds the GPU.

The quick-disconnects: the number one leak and service point

The quick-disconnect, the QD, is the dripless blind-mate coupling that joins a server to the rack manifold, and it is both the most useful and the most leak-prone part of the loop. It earns its place because it lets a technician pull a node for service without draining the loop or puddling the cabinet. A dry-break QD seals both halves as they part, so a good one loses a drop or less on disconnect. The direct-to-chip cooling guide covers the dripless design and the in-rack manifold in depth.

It is the number one leak point precisely because it is the number one service point. Every time a node comes out and goes back, a person seats that coupling, and a coupling that is cross-threaded, unlatched, worn, or disconnected under pressure turns a routine swap into a leak event on a powered rack. The hardware is only as good as the procedure. A worn or cheap QD weeps under vibration, and a weep onto a live server is exactly the failure the whole detection system exists to catch.

So the QDs get two kinds of attention. They get sensor coverage, because they are where fluid most often escapes, and they get a service discipline, because the connection is made by hand. Confirm the couplings are dry-break, accessible without fighting the rack, and latched fully on every reconnect. Treat the QD as the part most likely to be the source when coolant shows up where it should not be.

The detection layers: distributed coverage plus targeted spots

Good leak detection is two layers working together, not one device. The distributed layer is leak-sensing rope or cable, a continuous element run along the routes the coolant would travel, under the racks, along the pipe and manifold runs, and beneath the CDU. The targeted layer is spot or point sensors placed at the specific places fluid collects, the individual fittings, the drip trays, the manifold low points, and the CDU pan. The rope covers the length. The spots cover the points.

Each layer answers a different question. The rope tells you a leak happened somewhere along its run and gives you a location to within a usable distance, which is what you need when the leak could be anywhere across a long route. The spot sensor tells you a known high-risk place has fluid in it, which is what you need at a fitting or a tray where a leak is both likely and catchable early. The best designs combine them: spots at the critical connections and trays, rope for the broad runs in between.

Coverage is the discipline that makes the layers count. A leak that lands between the sensors is a leak the system never sees, so the layout follows the leak map from the section above. Every fitting, every low point, every place a drip would reach something energized gets a sensor on it or a rope passing through it. Detection you did not route to the leak is detection that does not protect the load.

Rope and cable sensors: catching the leak anywhere along the run

Leak-detection rope, also called sensing cable or linear leak detection, is a continuous element that senses fluid anywhere along its length and reports roughly where. You run it where the coolant would go if it escaped: under the racks, along the supply and return pipe routes, at the manifold low points, and around the base of the CDU. The moment fluid bridges the sensing element, the cable alarms, and a locating system reports the distance along the run so the responder goes straight to the leak instead of searching.

The value of the rope is its reach. A spot sensor only sees fluid that reaches the spot. A long pipe run, a cable tray full of hoses, or the floor under a row of racks is too much area to cover with point sensors alone, and the rope covers all of it as one continuous line. Route it along the low edge of the area it protects, because escaped fluid runs downhill and collects at the bottom, and that is where the cable should be waiting for it.

Routing is where rope systems are won or lost. Run it where fluid will actually travel, keep it off surfaces that condense, and confirm the locating calibration so the reported distance lands on the real leak. A rope strung along the ceiling protects the ceiling. The leak is on the floor under the manifold, and that is where the cable has to be.

Spot sensors: the targeted catch at the fitting and the tray

Spot or point sensors are the targeted layer, placed at a specific high-risk location where even a few drops should raise an alarm. They go at the individual fittings and quick-disconnects, in the drip trays, at the manifold low points, and in the CDU pan, the places where a leak is both likely to start and likely to be caught early. A spot sensor inside a CDU cabinet or under a rack manifold sees the first weep before it becomes a spread.

The spot sensor trades reach for certainty. It only sees fluid that reaches it, so it does nothing for the long pipe run between racks, but where it sits it gives an early, unambiguous signal that this connection or this tray has fluid in it. That makes spots the right tool at the densest clusters of wet connections, which is why the CDU commissioning guide treats the pan sensor and its auto-isolation as a separate proof: the CDU is the tightest cluster of pumps, valves, and joints in the whole system.

Place spots at the points your leak map flagged as most likely and most consequential. A drip tray under the manifold with a sensor in it catches the QD weep at its source. A pan sensor under the CDU catches an internal leak before it reaches the floor. The spot is the catch at the place you already know is risky.

Automatic isolation: the act half of the system

Automatic isolation is the part that stops the leak, and it is what separates a real protect-the-load system from an alarm panel. On a detected leak, the system drives an actuated valve closed to stop the flow, and depending on the design it can also stop the pump and isolate the affected loop or the affected rack so the coolant stops feeding the leak. The whole point is to act before the fluid spreads, which means the action has to be automatic and fast, not waiting on a person to interpret an alarm and walk the floor.

Be blunt about why this cannot be manual. A loop under pressure keeps pushing coolant out of the breach for as long as it stays energized and open. Human response in a well-instrumented facility runs minutes. An actuated isolation valve closes in seconds. A conductive leak can fault a board inside that gap, so a design that detects in seconds and then waits minutes to act has thrown away its own head start. The valve closing on its own is the difference between a contained event and a destroyed row.

The design intent is to isolate the leaking loop without dropping the racks that are fine. Isolation valves at the rack or row level let the system shut the one zone with the leak while the rest of the hall keeps cooling, the same way you would drain a single rack for service without taking the pod offline. ASHRAE's cold-plate resiliency guidance, the CDU OEM, and the commissioning agent all push the same point: detect and isolate automatically, and prove the isolation actually fires.

Fast shutoff: closing the valve in seconds, not minutes

Fast shutoff is the mechanics of the isolation: an automated valve that closes on detection and stops the flow at the source. The sensor sees the leak, the controller commands the valve, and the actuator drives it shut, all without a person in the loop. A common approach is a bypass arrangement that shuts down or diverts the fluid to the cooling loop with the leak, so the leaking branch loses its supply while the healthy branches keep running.

Speed is the spec that matters here. The reason to automate the shutoff at all is that the coolant spreads faster than a technician can reach it, so the valve has to act in seconds to beat the spread. Size and select the valve and actuator for that closing time, and confirm it during commissioning, because a valve that takes a slow minute to close has given the leak a minute to find a board. The valve closing time is a number worth writing down and testing against.

The design discipline is to isolate the leak without collapsing the cooling for everything else. Zone the isolation so a single leak shuts its own loop or rack, not the whole hall, and so the remaining racks ride through on their own loops. A shutoff that drops the entire room on one leak trades a coolant problem for a thermal one, and a dense rack overheats in seconds when it loses flow. Isolate the leak, keep the rest cooling, and prove the sequence does exactly that.

Drip trays, containment, and drainage: the physical defense

Before detection and isolation ever fire, the physical defense is containment: drip trays and pans that catch escaped coolant and drainage that carries it away from the IT. The tray under a manifold or a CDU holds the first drips so they collect in one place instead of spreading across the floor, and a sloped or drained tray sends that fluid to a safe point rather than letting it pool where it can reach energized gear. Containment is the layer that buys the detection and isolation time to work.

The containment and the detection are designed together, not separately. The spot sensor goes in the tray, because the tray is where escaped fluid collects, so the sensor sees the leak at the exact place the containment caught it. A tray with no sensor catches the fluid but never tells anyone. A sensor with no tray sees scattered drops too late. Put them together: the tray catches it, the sensor in the tray sees it, and the fluid drains away from the load.

Drainage is the part that gets value-engineered out and missed later. A tray that catches coolant but has nowhere to send it overflows onto the floor once it fills, so the secondary containment needs a path to a floor drain or a safe collection point, with enough slope to actually move the fluid. Catch the leak, hold it off the live gear, and drain it somewhere it does no harm. That physical defense is the first line, before a single electronic sensor reports anything.

Integration: tying detection to the BMS, the alarm, and the shutdown

Leak detection that lives only on a local panel is detection nobody sees at three in the morning, so it has to integrate with the building management system and the DCIM platform. The detection ties into the BMS or DCIM so an alarm reaches the network operations center the moment a sensor fires, with the location, the escalation path, and the shutdown sequence all driven from there. The point of the integration is that a leak shuts its zone and tells the room, not that it blinks a light on a box in an empty hall.

The integration carries both the alarm and the action. On detection, the system raises the alarm to the NOC with the leak's location and runs the programmed response: close the isolation valve, stop the affected pump, and where the design calls for it, escalate to a controlled shutdown of the hardware in the affected zone. There is a relationship with the EPO, the emergency power-off, in that both are last-resort protective actions, but a leak response is meant to be a surgical isolation of the affected loop, not a hall-wide power kill. Keep the leak shutdown sequence and the EPO distinct and documented so a leak isolates a loop rather than dropping a room.

The facility runs this monitored around the clock, which is the whole reason the alarm has to reach the NOC. A sensor and a valve handle the seconds. The NOC handles the response after, the verification, the dispatch, the assessment, and the cleanup. Tie the detection to the controls so both halves work: the automatic action in the first seconds and the human response in the minutes that follow.

BMS and DCIM: getting the alarm to someone who can act

The monitoring side of the system is what turns a tripped sensor into a response, and the job at commissioning is to confirm every leak alarm reaches the BMS and the DCIM with the right label, the right location, and the right escalation. An alarm mapped to the wrong point is worse than no alarm, because it sends the responder to the wrong rack while the real leak keeps spreading. Point-to-point verify each leak alarm end to end, from the sensor to the screen the NOC actually watches.

The location is what makes the alarm useful. A rope system reports the distance along its run, and a spot sensor reports its fixed position, so the BMS should show the responder where the leak is, not just that one exists somewhere in the hall. On a floor with thousands of connections, an alarm that says only leak detected sends a person searching while the clock runs. An alarm that says leak at this manifold, this rack, sends them straight to it.

The escalation is the rest of the monitoring job. Set the alarm thresholds, the notification path, and the defined incident response, who gets called, what the first action is, and when the shutdown triggers, so the alarm drives a known sequence rather than a scramble. That configuration is what converts a detection investment into an operational capability, and it is verified at commissioning alongside the sensors and the valves.

How do you commission leak detection and isolation?

You commission leak detection by proving the detection and the isolation both work before the hall goes live, not by confirming the parts were installed. The acceptance test is a simulated leak: introduce fluid at a sensor, then confirm the sensor alarms, the location reports correctly, the isolation valve closes, the alarm reaches the BMS and the NOC, and the whole chain completes inside the design response time. Detection that was never tripped and isolation that never fired are unproven, and unproven protection is the kind that fails the first time it is needed.

Run it as part of the larger commissioning sequence, alongside the loop and the CDU. The CDU commissioning guide covers the unit's own pan sensor and auto-isolation, and the direct-to-chip cooling guide covers the loop the sensors protect. The leak-detection acceptance ties into both: the loop has to be flushed, filled, and pressure-tested first, and the CDU's response sequence has to be proven, so the simulated leak is tested on a system that is otherwise ready. Test it before coolant is carrying real load, because a leak found at commissioning is a punch-list item and a leak found after the GPUs are in is a disaster.

Write the script before you test. Define every sensor to be tripped, the expected alarm, the expected location, the expected isolation action, and the maximum acceptable response time, then run each one and record the result against the script. ASHRAE's commissioning guidance frames the process, and the project specification and the OEM set the acceptance criteria. The deliverable is a documented test that shows each sensor was tripped, each valve closed, and each alarm landed, with the times. That record is the proof the protection is real.

Testing: simulating the leak and verifying the chain

The test that proves the system is a simulated or injected leak at the sensor, run to verify the full chain: the sensor sees it, the location is right, the valve closes, the alarm fires, and the time is inside the limit. You apply a small, controlled amount of the actual fluid, or a test fluid the sensor responds to the same way, at each sensing point, and you watch the response from the sensor through the controls to the valve and the BMS. Anything you do not trip is something you have not proven.

Verify each link, because the chain is only as good as its weakest one. A sensor that alarms but whose signal never reaches the valve has detection without isolation. A valve that closes but whose alarm never reaches the NOC isolates the leak but leaves the room blind. A rope that alarms but reports the wrong location sends the responder to the wrong place. Test the sensor, the location, the valve, the alarm, and the response time as separate confirmations, and record each one.

Document the result and plan to re-test. A leak-detection system degrades quietly, sensors drift, valves stick, and integrations break when someone changes a BMS point, so the simulated-leak test belongs in the periodic maintenance plan, not just in the one-time commissioning. The commissioning test sets the baseline. The periodic re-test confirms the protection still works years into operation, which is exactly when a complacent operator assumes it does without checking.

Why does condensation cause false alarms, and how do you stop them?

Condensation causes false leak alarms because a leak sensor cannot tell, on its own, whether the water on it came from a breached pipe or from the air. When a cold surface in the loop drops below the room dew point, water condenses on it out of the air and drips exactly where a real leak would, onto the piping, into the tray, onto the sensor. The CDU commissioning guide covers holding the secondary supply above dew point for this reason. A loop running below dew point makes its own water, and the leak detector dutifully reports it.

Nuisance alarms are not a harmless annoyance, they are a hazard, because an operator who gets false leak alarms learns to ignore leak alarms. The day the alarm is real, it reads like the dozen false ones before it, and the response that should have taken seconds does not happen. A leak-detection system that cries wolf is worse than a quieter one, because it trains the people watching it to disregard the one signal that matters.

You stop the false alarms with placement, dew-point control, and tuning, in that order. Keep the cold surfaces above dew point so they do not condense in the first place, insulate the piping and exchangers that run cold so any surface below dew point cannot drip, and place the sensors away from the surfaces that condense. Then tune the sensor sensitivity to the coolant blend and the environment so it still catches a small real leak without tripping on humidity. The goal is a system the operators trust, because a trusted alarm is the one they act on.

The response runbook: what happens after the alarm

The automatic isolation handles the first seconds. The response runbook handles everything after, and it is what turns a contained leak into a closed incident instead of a lingering one. The runbook is the written sequence the NOC and the floor team follow: the alarm comes in, they verify it is real and locate it, they confirm the automatic isolation fired and apply manual isolation if it did not, they assess the damage and the affected hardware, and they clean up and restore the loop. Each step is defined so nobody improvises during an event.

The verify step matters because not every alarm is a flood and not every alarm is false. The responder confirms whether this is a real leak, a condensation nuisance, or a sensor fault, and they confirm the automatic isolation actually stopped the flow. Where the auto-isolation did not fire, or where the leak is upstream of it, manual isolation is the backup, so the runbook names which valves to close by hand and where they are. Detection and automatic isolation are the primary defense. Manual isolation is the catch behind it.

A runbook is only as good as the people trained on it, so the team drills it. Walk the sequence, run a simulated leak as a live exercise, and confirm the responders know the valves, the isolation points, and the escalation. Keep the spares on hand for the cleanup and the QD or fitting replacement, because the leak that just got isolated still has to be repaired before the loop comes back. A trained NOC running a drilled runbook is what closes the gap between the valve and the recovery.

The maintenance the system needs to keep working

A leak-detection system is a protective system that sits idle until the day it is needed, which is exactly the kind of system that degrades unnoticed. Naming its maintenance in the turnover package is what keeps it alive. The sensors get tested on a schedule, not assumed, because a rope or spot sensor that has drifted or failed is a blind spot nobody knows about until a leak lands on it. The periodic simulated-leak test from the commissioning section is the core of the program.

The connections and the coolant get their own attention. The quick-disconnects and fittings, the number one leak points, get inspected for weeping, seating, and wear, because catching a loosening coupling on a PM beats catching it as an alarm. The coolant chemistry gets sampled, because the inhibitor that keeps the fluid conductive enough for the sensors to detect, and that protects the metals, depletes over time, and an off-spec fluid changes both the corrosion picture and the detection behavior. The deeper coolant program lives in the direct-to-chip and CDU guides; here the point is that the chemistry and the detection are linked.

Then the mechanical parts of the response. The isolation valves get exercised, because a valve that never moves can seize and a seized valve does not close on a leak. The trays get cleaned so debris does not block drainage or fool a sensor. A leak-detection system handed over without a sensor-test interval, a QD inspection, a coolant-sampling schedule, a valve-exercise routine, and a tray cleaning plan is a system that will work fine until the day it quietly does not.

What to document

A leak-detection system that was installed but never documented hands operations a black box, and the first alarm becomes a guessing game about what is supposed to happen. Capture the detection map, the isolation scheme, the commissioning and simulated-leak results, the response runbook, and the alarm configuration, so the people who inherit the hall know what protects it and how to maintain it. A field tool like FieldOS keeps the sensor map, the test records, and the maintenance log in one place tied to the asset, so the next simulated-leak test compares against the last one instead of starting blind.

Two records carry the most weight. The simulated-leak commissioning result proves the detection and isolation actually fired and met the response time, which you cannot re-prove once the hall is in production without staging another test. The detection map and isolation scheme tell the responder, during a real event, where every sensor is, what zone each valve isolates, and which connection an alarm points to. A turnover package missing either leaves the owner trusting the protection without a way to verify or run it.

Common mistakes

• Detection that only alarms, with no automatic isolation to stop the flow before the coolant spreads.
• No sensor coverage at the quick-disconnects and fittings, the number one leak points.
• No drip trays or containment, so escaped coolant spreads to the floor and the live gear before anything catches it.
• Leak detection not integrated to the BMS and the shutdown sequence, so the alarm lives on a local panel nobody watches.
• Never simulating a leak in commissioning, so the detection and isolation are assumed to work rather than proven.
• Nuisance alarms from condensation that get ignored, training the operators to disregard the one alarm that is real.
• Picking a conductive sensor for a high-glycol or dielectric loop it cannot detect, leaving a blind spot in the coverage.
• Isolation that drops the whole hall on one leak instead of isolating the affected loop or rack.

Field checklist

Standards and references

The thermal and resiliency framework comes from ASHRAE Technical Committee 9.9, whose liquid-cooling guidance and cold-plate resiliency material address leak response, CDU redundancy, and coolant-quality management, and whose thermal guidelines set the supply-temperature and dew-point context that keeps a loop from condensing. Treat the specific water classes, dew-point margins, and quality limits as the published ranges and confirm them against the current edition, because the guidance is revised as the hardware moves. ASHRAE gives the framework. It does not replace the equipment specification.

The cooling OEM, the CDU and cold-plate and manifold manufacturers, specify the coolant, the wetted materials, the sensor types their equipment supports, and the auto-isolation behavior built into the unit. The leak-detection manufacturer specifies the sensor selection for the fluid and concentration, the rope locating accuracy, and the response characteristics. Where an OEM specification is stricter than a general guideline, the OEM specification governs the equipment it covers, and the coolant the OEM approves drives which sensing method can detect it.

The design engineer sets the leak-detection layout, the isolation zoning, and the integration to the BMS and the shutdown sequence, and the commissioning agent proves all of it with the simulated-leak test against the project specification and the sequence of operations. ASHRAE commissioning guidance frames that process, and where a facility chases a tier the Uptime Institute drives the witnessed demonstrations. Hedge the detection method, the isolation scheme, and the coolant to ASHRAE, the OEM, and the commissioning agent, and confirm the editions and revisions against the published documents before citing them on a submittal.

Units and terms

Liquid-cooling leak detection borrows vocabulary from controls, from the chip vendors, and from the fluids side, and the same idea reads differently across a sensor datasheet, a CDU submittal, and an ASHRAE guideline. The terms below are the ones that cause the most confusion on a first job.

Leak detection

The sensors, alarms, and isolation that find coolant where it should not be and act on it before it reaches the IT

Detect vs isolate

Detection finds and locates the leak; isolation stops it by closing a valve. Both are required, not just the first

Rope / cable sensor

A continuous linear element that senses fluid anywhere along its length and reports the location, for distributed coverage

Spot / point sensor

A fixed sensor at one high-risk place, a fitting, tray, low point, or CDU pan, for an early targeted catch

Quick-disconnect (QD)

A dripless blind-mate coupling joining a server to the manifold for live service, the number one leak point

Automatic isolation valve

An actuated valve that closes on detection to stop the flow and isolate the affected loop or rack in seconds

Drip tray / containment

A pan that catches escaped coolant and drains it to a safe point, the physical defense before the sensor fires

Dielectric vs conductive coolant

Dielectric fluid does not conduct; water and treated glycol do, so a conductive leak on a live board can short it

BMS / DCIM integration

Tying the detection to the building and data center management systems so a leak alarms, locates, and drives a shutdown

Related tools

Calculators and readiness checks for this work

Compare your options

FAQ

People also ask

• What is direct-to-chip liquid cooling?
• What is a CDU in a data center?
• What is direct-to-chip liquid cooling?
• What is a leak detection system in a data center?
• What is a rear-door heat exchanger?