CDU commissioning field guide for AI data center liquid cooling

The CDU, and why it sits between two loops

A coolant distribution unit, the CDU, is the package of pumps, a heat exchanger, filtration, and controls that conditions and isolates the clean coolant loop running to the rack cold plates. It is the heart of a direct-liquid-cooling deployment. The chips never touch facility water. They touch the coolant the CDU makes, at the temperature, flow, and pressure the CDU holds, and the whole reason the unit exists is to keep the dirty, warmer, building-pressure facility water on one side of a heat exchanger and the clean, conditioned rack coolant on the other.

Strip away the marketing and a CDU does four jobs. It moves heat from the secondary loop into the facility water across the heat exchanger. It pumps the secondary coolant out to the rack manifolds and back. It holds the supply temperature above the room dew point and the flow and pressure the cold plates need. And it watches itself for leaks, fouling, and pump trouble and tells the building when something is wrong. Commissioning is proving all four under load, not reading the nameplate.

The loop on the far side of the CDU, the piping, the manifolds, the cold plates, the coolant chemistry, and the flush that has to happen before any plate is connected, is its own scope. The liquid cooling loop commissioning guide covers the loop end to end. This guide stays on the box itself: the CDU, what it has to prove, and the sequence that accepts it.

What is the difference between a liquid-to-liquid and a liquid-to-air CDU?

A liquid-to-liquid (L2L) CDU rejects the rack heat into facility water through a plate heat exchanger. A liquid-to-air (L2A) CDU rejects it straight into the room air across a fan coil and skips the facility water loop entirely. That one difference decides almost everything else about the unit, its capacity, and what you commission.

The L2L is the workhorse of a built-out hall. Facility water is already plumbed to the row, the plate heat exchanger passes the heat into it, and the two loops stay hydraulically isolated, so the clean secondary side never sees what is in the building loop. L2L units carry the big numbers, with row-scale and perimeter units rejecting into the high hundreds of kilowatts and past a megawatt per unit, because a water-to-water exchanger moves far more heat in a given footprint than a coil dumping into room air. The catch is that an L2L is only as good as the facility water behind it. Commission the facility water flow and temperature alongside the CDU, because a CDU starved of facility water cannot make its capacity no matter how good its pumps are.

The L2A earns its place where there is no facility water loop to reject into: an edge site, a retrofit hall, a proof-of-concept row, a deployment that has to stand up before the chilled water plant is ready. It takes the rack heat into the room air, which means the room air handling has to absorb it, so an L2A does not reduce the building heat load, it relocates it. L2A units top out lower than L2L for the same footprint, and they put the heat back in the space the air-side cooling already has to handle. Know which one is in front of you before you write the test plan, because the L2L test proves the facility-water interface and the L2A test proves the room can take the rejected heat.

Why does a CDU have two loops?

A CDU has two loops because the water the building makes is not clean enough or controlled enough to run through a cold plate, and the cold plate's micro-channels are too fragile to forgive it. The primary loop is the facility water system, the building chilled water from the plant. The secondary loop is the technology cooling system, the conditioned coolant that runs from the CDU out to the racks. The heat exchanger in the CDU passes heat from the secondary to the primary without passing the water, and that isolation is the entire point.

Keeping the loops separate buys three things at once. The facility water can stay at building pressure, building temperature, and building water quality, while the secondary loop runs at the controlled pressure and the warm, stable, above-dew-point temperature the chips need. Construction debris, scale, and biology in the facility water never reach a cold plate. And a leak on the secondary side is a small, contained volume of known, treated coolant, not the whole building loop draining onto a cabinet. The naming, primary versus secondary, technology cooling system versus facility water system, is worth getting right on day one, because the test plan, the chemistry, and the flush all reference one side or the other and a crew that mixes them up flushes the wrong loop.

The secondary loop coolant, its glycol concentration, its inhibitor package, and the cleanliness it has to meet before cold plates go on, is covered in the liquid cooling loop guide. For the CDU itself, the loop count is what to fix in your head: the unit is the wall between the two, and most of what you commission is whether that wall does its job under load.

The heat exchanger and the approach temperature

The heat exchanger is where the secondary loop hands its heat to the facility water, and on an L2L CDU it is almost always a brazed-plate or gasketed-plate exchanger: a stack of thin corrugated plates that put the two fluids on opposite faces with a lot of surface in a small box. Plate exchangers move heat efficiently and pack a large capacity into the CDU cabinet, which is why they dominate. The number that defines one is the capacity it rejects, the kilowatts it can pass at the design flows and the design temperatures on both sides.

The reading that tells you how the exchanger is doing is the approach temperature, the gap between the facility water coming in and the secondary coolant leaving on the cold side. A tight approach means the exchanger is transferring heat well. A widening approach over time means it is fouling or scaling, losing capacity, and it shows up first as the secondary supply temperature creeping up even though the facility water is on spec. Record the design approach at commissioning. It is the single baseline that lets operations catch a fouling exchanger before a chip throttles, and an exchanger with no baseline approach is one nobody can tell is drifting.

The trap at commissioning is proving capacity on a cool afternoon with light load. An exchanger sized right looks fine at 20 percent load no matter what. You only see whether it makes its rated rejection, at its rated approach, when both sides are at design flow and the secondary side is carrying the heat it was built for. That is why the capacity proof lives in the load test, not in the static checks.

The pumps and the redundancy inside the CDU

The pumps are what push the secondary coolant out to the racks and back, and they are the most common point of redundancy in the unit. The usual arrangement is N+1: one more pump than the load needs, so a pump can fail or be pulled for service and the racks keep their flow. Some units run one pump with a standby, some load-share across both and ride at part speed; either way, N+1 means a single pump can drop without the racks losing coolant. The redundant pump is worth nothing until you prove it carries the load, which is the failover test, not the data sheet.

The pumps are commonly on variable-frequency drives so the CDU can modulate flow to hold its control target instead of running flat out. A VFD lets the unit turn down at part load to save pump energy and ramp up when a pump drops or the load climbs. Common practice is to specify the pump with head margin above the calculated design point, often in the range of 15 to 25 percent, and a turndown that reaches well below rated flow, so the unit has room to push harder when one pump is gone and to throttle back when the hall is quiet. The exact margin and turndown are the manufacturer's and the design's call, so confirm them against the submittal.

Two failures hide in the pump section. A pump that holds flow on its own at full speed but cannot hold it once its partner drops, because the survivor runs out of head, is N+1 on paper and N in practice. And a VFD set up to chase the wrong signal hunts, ramping and backing off while the flow swings. Prove the survivor holds the per-rack flow under load, and watch the drive hold steady as the load steps, before you sign the pumps off.

How the CDU controls supply temperature, flow, and pressure

The CDU controller is what makes the unit more than a pump and a heat exchanger in a box. It modulates the pump speed and the facility-water control valve to hold three things at once: the secondary supply temperature, above the room dew point and at the design setpoint, and the flow or the differential pressure to the rack manifolds. Most units control to a differential pressure across the supply and return manifolds, so the dP stays constant as racks draw more or less, and the per-rack flow holds even as the load shifts around the row.

Differential pressure control is the right instinct for a manifold-fed loop. The pump develops head, the manifolds and cold plates consume it, and if the controller holds the dP across the manifold headers steady, every branch keeps the pressure it needs to drive its flow. Let the dP sag and the racks farthest from the pump go short first, the same way the top of a tall riser starves on a hydronic system. So the controller's job is to ride out a load swing or a facility-water upset without letting the supply temperature wander out of band or the manifold dP collapse.

Test the control loop the way it will actually be challenged, not at one quiet setpoint. Step the load. Drop the facility water temperature or flow. Pull a pump. Watch whether the CDU holds the secondary supply temperature and the manifold dP inside the design limits through each of those, and how fast it recovers. A controller tuned soft drifts out of band under a step and takes too long to come back; a controller tuned tight hunts. The acceptance is stable control through the transients, recorded, against the sequence of operations the design wrote.

Why must the secondary water stay above the dew point?

The secondary coolant has to stay above the room dew point because the moment a cold surface drops below the dew point of the air around it, water condenses on it, and a CDU and a rack full of GPUs is the last place you want water forming out of the air. Condensation on the secondary piping, the manifolds, or inside the cabinet drips onto powered electronics exactly like a leak, except this water comes from the room, not the loop, so no amount of leak-tight piping prevents it. The only defense is keeping the cold surfaces warm enough that the air cannot condense on them.

This is why liquid cooling runs warm coolant and why warm-water cooling pays off. A common control rule is to hold the secondary supply temperature a margin above the measured room dew point, often cited around 2 C above dew point, so the coldest surface in the loop stays dry and the cooling stays fully sensible, removing heat without ever wringing water out of the air. The CDU reads the room dew point, or is given it, and floors its supply setpoint above it. The chip cools fine on coolant in the 30s of degrees C because the junction-to-coolant temperature difference is still large, so running warm costs the chip nothing and keeps the loop dry.

The blunt version: set the secondary supply below the room dew point and you have built a condensation machine sitting on top of the most expensive hardware in the building. The room dew point, not the relative humidity, is the number that governs, and it moves with the room conditions, so the CDU has to track it, not hold a fixed setpoint and hope. The water-class framework and the chip-side temperature limits behind that setpoint are covered in the liquid cooling loop guide.

Filtration and secondary water quality

The CDU is where the secondary loop gets filtered, because everything downstream of it ends in a cold plate with channels a fraction of a millimeter wide. Most units carry a supply-side filter, commonly around 50 microns, in the main flow path to catch particles before they reach the manifolds, and many add a side-stream filter, a slipstream off the main flow run through a much finer element, sometimes down toward 0.2 microns, that polishes the whole loop volume over time without choking the main flow. Confirm both are installed, the right rating, and that the filter differential-pressure alarm is set and reports, because a loaded filter starves the loop the same as a closed valve.

Filtration is only half of clean coolant. The secondary loop also has to meet a water-quality spec, the chemistry that keeps it from corroding, scaling, or growing biology in the warm, dark, low-flow loop. Manufacturers commonly call out limits on chlorides, sulfates, and total suspended solids, with figures often cited in the low single-digit to low-tens of parts per million, alongside a glycol concentration and a corrosion-inhibitor package matched to the metals in the loop. Those numbers vary by manufacturer and by the metallurgy of the system, so the equipment specification governs, not a generic table.

The coolant chemistry, the glycol, the inhibitor, the biocide, and the sampling baseline you set at turnover, is covered in depth in the liquid cooling loop guide, because the chemistry lives in the loop, not in the CDU. What belongs to the CDU commissioning is narrower and concrete: the filters are the specified rating, the filter dP alarm works, the side-stream is flowing, and the loop was filled with coolant that meets the spec, not whatever was in the yard. A CDU pushing dirty or wrong-chemistry coolant is a CDU fouling cold plates on a schedule.

CDU leak detection, the drip tray, and auto-isolation

Leak detection on a CDU is layered, and commissioning it means proving each layer fires, not confirming it was installed. The unit sits in a drip tray or has an internal containment pan, because the CDU itself is full of wet connections, the pumps, the exchanger, the valves, the filter housings, and a weep inside the cabinet has to be caught before it reaches the floor. A point or float sensor in that pan reports liquid where there should be none. Rope, or linear, leak-sensing cable runs along the secondary piping and reports both that a leak happened and roughly where.

Many CDUs are wired to do more than alarm. On a detected leak, the unit can run a programmed response: raise the alarm to the building management system, and where the design calls for it, isolate the affected loop by closing valves or shutting a pump so the leak stops feeding. That auto-isolation is the difference between a contained drip and a loop emptying onto a cabinet, and it only counts if it was tested. Trip the pan sensor and confirm the alarm reaches the BMS and the CDU does what its sequence says, whether that is an alarm only, a pump action, or a valve isolation.

The reflex on the floor is to treat leak detection as a checkbox: sensors installed, item closed. The better instinct is to assume a leak someday and prove that when it comes, it is detected, contained in the tray, and isolated before it reaches power. Leak detection across the whole liquid loop, the rope and point sensors out at the manifolds and quick-disconnects, is detailed in the liquid cooling loop guide. At the CDU, the pan sensor and the auto-isolation are the ones to wet and watch, because the CDU is the densest cluster of wet connections in the system.

The connection to the rack manifold

The CDU feeds the racks through supply and return manifolds, and the handoff from the unit to the manifold is where the CDU's job meets the rack's. The CDU pushes conditioned coolant out a supply header, it runs to the rack manifolds, the vertical distribution headers that split flow to each node, and the warmed return comes back to the CDU to be cooled again. The manifold dP the CDU controls to is measured across these headers, so the unit and the manifolds are one hydraulic system, not two.

What the CDU commissioning has to confirm at this interface is that the unit delivers the design flow and pressure at the manifold connections, and that the secondary side was made up clean. Quick-disconnects join nodes to the manifold for live service, and every coupling is a sealing surface and a potential leak and contamination path. The in-rack distribution, the manifold, the quick-disconnects, and the per-node balancing are detailed in the liquid cooling loop guide, because that work lives in the rack.

The line to hold is this: the CDU is responsible for delivering the right flow and pressure to the manifold connection, and the loop downstream is responsible for getting it to each node. They get balanced together, but the CDU has to put the flow and the dP at the manifold inlet before the rack-side balancing can do anything with it. A CDU that cannot hold dP at the manifold leaves the rack-side balancing chasing a moving target.

How do you flush, fill, and de-air the CDU secondary loop?

You flush the secondary loop clean, fill it with the specified coolant, and bleed the air out, in that order, before the unit ever carries real load. New piping, manifolds, and the CDU's own internals come full of construction debris, weld slag, scale, cutting oil, and grit, and sending that into a cold plate chokes a micro-channel on the one chip that can least afford it. The flush circulates a procedure fluid through the loop at a velocity high enough to scour the walls, through staged filtration from coarse to fine, until the loop meets a measured cleanliness target. Connect cold plates only after the cleanliness result passes. The staged flush and the cleanliness targets are detailed in the liquid cooling loop guide; what the CDU contributes is the pump that drives the flush and the filtration that catches what it mobilizes.

Filling is not just adding fluid. You fill with the coolant the spec calls for, the right glycol concentration and the right inhibitor package, not plain water and not a top-off of something convenient, and you fill in a way that does not trap air. Air in a liquid loop is a real problem: a pocket at a high point or in a cold plate blocks flow, an air-bound pump loses prime and cavitates, and entrained air degrades the heat transfer. The CDU usually has an air separator and bleed points, and the fill procedure works the air up to them and vents it.

De-airing is the step crews shortcut and pay for later. You run the pumps, work the air to the high points and the separator, bleed it, top off, and repeat until the loop holds steady flow and pressure with no air signature, no fluttering flow, no pressure swing, no low-NPSH or air-ingress alarm from the pump. Then the loop gets its pressure test, held above working pressure long enough for temperature to stabilize so you can tell a real leak from the test fluid cooling and contracting. The pressure test parameters and medium come from the project spec and the applicable piping code, covered in the liquid cooling loop guide. A loop that was filled fast and never properly de-aired runs with a pump that cavitates and a cold plate that air-locks, and both look like a mystery until someone bleeds the loop.

The CDU commissioning sequence

CDU commissioning runs the same level structure as the rest of the mechanical plant, building from the installed unit up to the racks under load, and nothing functional gets tested until the static checks are signed. The order is not arbitrary. Each step proves something the next step depends on, and skipping ahead means testing on top of an unproven foundation.

Component verification first: the CDU is the unit specified, set in place, with the secondary piping and the facility-water connections made, power on, and the controls points checked end to end. Then the loop is flushed to the cleanliness target and the cold plates are connected only after it passes. Fill with the specified coolant, de-air the loop, and pressure-test it above working pressure while trending pressure and temperature to separate a leak from a thermal settle. Start the pumps and establish flow, then bring the facility water on and confirm the heat exchanger is passing heat at the design approach. Bring up the controls and confirm the unit holds the secondary supply temperature above dew point and the manifold differential pressure at setpoint.

Then the tests that decide acceptance. Trip the leak detection and confirm the alarm and the auto-isolation fire. Fail a pump under load and confirm the N+1 standby holds the per-rack flow. And run the load and heat verification, with the racks at design load on IT load or load banks, to prove the CDU rejects its rated capacity at its design temperatures and flows while holding every control target. That last one is the keystone, and on a real facility it runs alongside the electrical integrated test, because the CDU has to ride through a power event the power side has proven, and the chilled water plant behind an L2L unit has to deliver facility water through the same event.

How do you load-test a CDU and prove its capacity?

You load-test a CDU by putting heat into the secondary loop, IT load or load banks, and confirming the unit rejects it at the design temperatures and flows while holding its control targets. Capacity is the kilowatts the CDU moves from the secondary loop into the facility water, or into the room on an L2A, and you cannot see it on a cool, lightly loaded afternoon. The exchanger and the pumps look fine at 20 percent load no matter what. Full load is the only test that matters, because that is the condition the owner bought the unit for.

The numbers you take are the heat rejected and the delta-T on both sides. The secondary delta-T, the temperature rise from supply to return, against the secondary flow, gives the heat the loop is carrying. The primary delta-T against the facility-water flow gives the heat the exchanger is passing through. The two should agree within reason, and the heat exchanger approach should sit at its design value. A weak secondary delta-T at full flow means the racks are not handing the loop the heat the test expected, often a load-placement or rack-side issue rather than a CDU one; a wide approach at full load means the exchanger is short of capacity or already fouling. Record all of it as the baseline.

Hold the unit at design load long enough to stabilize, then watch the control targets through it: the secondary supply temperature stays above dew point and on setpoint, the manifold dP holds, and the pumps run inside their range with margin to spare. A CDU that makes its supply temperature only by running its pumps flat out has no margin for the day a pump drops or the load climbs, and that is a failed test wearing a passing number. Acceptance is the rated heat rejected at the design temperatures and flows, with the control targets held and pump headroom left.

How do you test CDU redundancy and failover?

You test CDU redundancy by taking out a redundant component under load and confirming the racks never lose their flow or their supply temperature. For the pumps, that means failing one with the racks at design load, by command, by killing its power, or by the method the test plan specifies, and watching the standby pick up and the per-rack flow and the secondary supply temperature hold through the handoff. You time how fast they recover and confirm they never cross the limit. A CDU that alarms or lets the supply temperature climb out of band when a pump drops is not N+1 in practice, whatever the nameplate says.

Where the design uses redundant CDUs feeding a common loop, the test goes up a level: drop a whole CDU and confirm the remaining units carry the load and hold the loop. That is a different proof from a single unit's pump failover, and it depends on how the units are headered and controlled together, so the test plan has to match the topology. A pair of CDUs that share a loop but were never proven to carry it on one unit is a single point of failure nobody tested.

The handoff is where redundancy is won or lost, not the steady state. A loop has thermal mass but not much, and a rack at high density heats fast when flow drops even for a moment, so the question is whether the standby ramps and the controls catch the loop before the chips throttle. Pull the pump, or the unit, watch the flow and the supply temperature, and record the recovery. That record is the redundancy. Everything else is a number on a data sheet.

Monitoring and alarms to the BMS and DCIM

A CDU generates a stream of data, and the commissioning job is to confirm all of it reaches the building management system and the DCIM platform and reads correctly, because an alarm that lives only on the unit's local screen is an alarm nobody sees at 3 a.m. The points that matter are the secondary supply and return temperature, the secondary flow and the manifold differential pressure, the facility-water temperatures and flow, the heat exchanger approach, the pump status and speed, the filter differential pressure, the coolant level and conductivity, and every leak and fault alarm the unit can raise.

Point-to-point verify the alarms end to end, not by trusting the integration. Trip each critical alarm at the unit, a leak, a high filter dP, a high supply temperature, a low flow, a pump fault, and confirm it shows up at the BMS and the DCIM with the right label and the right value, and that the critical ones drive the response the design specifies. An alarm mislabeled or mapped to the wrong point is worse than no alarm, because it sends operations to the wrong unit while the real one fails.

The monitoring is also what turns the commissioning baseline into something operations can use. Capture the as-accepted supply temperature, approach, flows, and dP as the numbers a future reading gets compared against, so a warm chip next year can be checked against how the CDU was accepted instead of guessed at. A CDU handed over with alarms that reach the BMS but no baseline to trend against is a unit that can shout but cannot tell operations it is slowly drifting.

The maintenance the owner takes on

A CDU is a pumped, filtered, chemistry-dependent machine running every hour for years, and naming the maintenance it carries in the commissioning deliverable is what keeps it from degrading quietly. The filters are first. The supply and side-stream filters load up, the differential pressure climbs, and a fouled filter starves the loop the same as a closed valve, so the filter-change interval and the dP alarm setpoint belong in the turnover package, not in operations' guesswork.

The coolant chemistry is the slow one. The glycol concentration drifts, the corrosion inhibitor depletes, and biology will grow in a warm loop if the chemistry lets it, all of which foul the exchanger and narrow the cold-plate channels. A coolant-sampling schedule and the acceptance limits to sample against, set from the baseline taken at commissioning, are part of the handover. The deeper coolant-chemistry program lives in the liquid cooling loop guide, because the chemistry is a loop property, but the CDU is where the sample is usually drawn and where the side-stream filter does its work.

Then the mechanical parts. The pumps and their VFDs wear, the heat exchanger approach drifts up as it fouls, and the leak detection and auto-isolation have to keep working, which means they get exercised, not assumed. A CDU turned over without a filter interval, a sampling schedule, a pump and exchanger baseline, and a periodic leak-detection test is a unit that will run fine until it does not, and nobody will see it coming. The maintenance plan is a commissioning deliverable, not an afterthought for operations to invent later.

How do you size and place a CDU for high-density AI racks?

You size a CDU to the heat the racks it serves actually produce, plus the redundancy and the margin the design calls for, and you place it in the form factor that fits the deployment. AI and GPU racks now pull 40, 80, and past 100 kW, more heat than any volume of air can carry through a cabinet, which is the whole reason the CDU is there. The unit has to reject that heat at the design temperatures with one pump out, so the sizing is the rack load times the rack count, divided across the units, with the N+1 condition still making capacity, not the nameplate at full pump complement.

The form factor follows the density and the deployment. An in-rack CDU sits inside the rack and serves that rack or a small cluster, commonly into the low hundreds of kilowatts, and suits a contained, self-standing high-density rack. An in-row or end-of-row CDU sits in the row and feeds a line of racks, scaling to far larger numbers, with row-class units reaching well past a megawatt. A sidecar CDU bolts onto a rack as a dedicated companion. Perimeter and large standalone L2L units serve a whole zone. The capacity figures are vendor-and-design numbers, not code, so size against the actual product and the project load, not a rule of thumb.

The sizing trap is the average and the redundancy. A hall can have enough total CDU capacity on paper and still cook a row, because an AI training cluster at one end pulls far more than a half-empty row at the other, and capacity averaged across the floor starves the dense zone. Size to the zone, not the hall. And size so the N+1 condition holds the load, because a unit that makes capacity only with every pump running has no redundancy at all. The rack-side architecture, cold plates, manifolds, and the hybrid air-and-liquid floor, is in the liquid cooling loop guide and the in-row cooling guide; the CDU's part is making the heat go away at the design temperature with a pump in reserve.

What to document

A CDU that was proven but never documented hands operations a unit with no baseline, and the first warm chip becomes a guessing game. The record is what tells the next engineer whether a reading is a new problem or how the unit has always run, and what it was accepted at. Capture the unit identity and type, the rated and as-tested capacity, the primary and secondary supply and return temperatures and flows, the heat exchanger approach, the manifold differential pressure, the dew-point margin held, the filtration and coolant chemistry baseline, the leak-detection and auto-isolation test result, and the redundancy and failover result.

Two records carry the most weight later. The load-test capacity at the design temperatures and flows proves the unit does the job it was bought for, which you cannot re-prove once the hall is in production, and the failover result proves the redundancy was actually exercised. A turnover package missing either leaves the owner trusting that the two most consequential things got done.

Common mistakes

• Setting the secondary supply temperature below the room dew point and condensing water onto the manifolds and the cabinets.
• Connecting cold plates to a secondary loop the CDU never flushed to the cleanliness target, then choking a micro-channel.
• Filling and starting the loop without de-airing it, so the pump cavitates and a cold plate air-locks.
• Running with the wrong coolant, the wrong glycol concentration, or no inhibitor, and fouling the exchanger and the cold plates.
• Skipping or undersizing the filtration, or leaving the filter differential-pressure alarm unset, so a loaded filter starves the loop unseen.
• Installing leak detection and auto-isolation but never tripping a sensor to prove the alarm and the isolation actually fire.
• Calling the pumps N+1 without failing one under load to prove the standby holds the per-rack flow and supply temperature.
• Sizing the CDU on the hall average instead of the dense zone, and sizing without the N+1 condition still making capacity.
• Commissioning an L2L unit without proving the facility-water flow and temperature behind it, so the CDU is starved at full load.
• Accepting the unit on a light-load afternoon instead of proving the rated capacity at design temperatures and flows.

Field checklist

Standards and references

The thermal framework comes from ASHRAE Technical Committee 9.9, whose thermal guidelines and liquid-cooling guidance set the facility water temperature classes, the supply-temperature bands, and the water-quality and material-compatibility guidance the secondary loop is built around. The Open Compute Project publishes the liquid-cooling and cold-plate requirements many AI deployments design to, including coolant, cleanliness, and CDU-related guidance. Name the OCP document by topic and confirm the current revision, since the documents revise on their own cycle, and the dew-point margin, the filtration ratings, and the water-quality limits are governed first by the equipment manufacturer's specification, which overrides any general guideline where it is stricter.

The capacity, the flows, the coolant, the redundancy scheme, and the sequence of operations are the manufacturer's and the project's call. The CDU rejects what the manufacturer rates it to reject at the design conditions, and the project specification sets the test pressures, the cleanliness targets, the acceptance criteria, and the witnessed demonstrations. The pressure test of the piping falls under the applicable pressure-piping code, commonly an ASME B31 section by topic, together with the project spec that sets the test pressure, medium, and hold; confirm which section applies before setting a number.

The commissioning process framework draws on the ASHRAE commissioning guidance, commonly Guideline 0, and where a facility is chasing a tier, the Uptime Institute standards drive the witnessed redundancy and integrated-test demonstrations. Edition numbers, the water-class values, and the OCP revision levels move between cycles, so confirm them against the published documents and the equipment manufacturer before citing them on a submittal. ASHRAE TC 9.9 and OCP give the framework; the manufacturer's specification and the project documents control the limits you commission to.

Units, terms, and acronyms

CDU commissioning carries vocabulary from HVAC piping, from the IT side, and from the chip vendors, and the same idea reads differently across a CDU submittal, a cold-plate datasheet, and an ASHRAE guideline. The terms below travel across the whole CDU scope.

CDU

Coolant distribution unit, the pumps, heat exchanger, filtration, and controls that isolate and condition the secondary loop

L2L / L2A

Liquid-to-liquid CDU rejecting to facility water through a heat exchanger, or liquid-to-air rejecting to room air through a coil

Primary / secondary loop

Facility water (FWS) on the building side, technology cooling system (TCS) coolant on the clean rack side, separated by the CDU heat exchanger

Approach temperature

The gap between facility water in and coolant out at the heat exchanger; a widening approach signals fouling

Differential pressure (dP)

The pressure difference the CDU holds across the manifold headers to keep each rack branch fed

Dew point

The temperature at which room air condenses; the secondary supply is held above it, commonly with a margin, to stay dry

Sensible cooling

Cooling that removes heat without condensing water, the mode the loop runs in when the coolant stays above dew point

N+1

One more pump or unit than the load needs, so a single one can drop and the racks keep their flow

VFD

Variable-frequency drive on the pump, letting the CDU modulate flow to hold its control target and ramp on a pump loss

Side-stream filtration

A slipstream off the main flow run through a fine filter that polishes the whole loop volume over time

kW rejected

The heat the CDU moves from the secondary loop to the facility water or room air, its capacity at design conditions

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 in-row cooling?
• Why do data centers need liquid-cooling leak detection?
• What is a rear-door heat exchanger?
• What is direct-to-chip liquid cooling?