16 min read

What Is a CDU? Facility Water, In Row, and In Rack Explained

A practical guide to coolant distribution units: the two loops they separate, the three form factors, how approach temperature sets the plant, and how to size flow from rack heat load.

ByAndré Ribeiro· Founder, Obelinf
What Is a CDU? Facility Water, In Row, and In Rack Explained
What Is a CDU? Facility Water, In Row, and In Rack Explained · November 12, 2026
On this page

A coolant distribution unit is the piece of equipment that makes liquid cooling a facility system instead of a rack experiment. It separates the water that leaves the building plant from the water that enters the servers, and in doing so it decides the temperatures, the pressures, the filtration, and the failure domain of the whole loop. Most of the confusion around CDUs comes from treating them as one product category. They are not. An in rack unit and a 2 MW in row unit share a name and a function, and almost nothing else.

This guide follows the numbers an operator actually has to choose and check: the two loops, the approach temperature, the flow that a rack heat load requires, and the redundancy that keeps a pump failure from becoming a rack outage. The math is the same heat balance that governs every cold plate loop, so it can be checked with a calculator before a purchase order is signed.

At a Glance: CDU Form Factors

In rack In row Centralized
Location Inside the cabinet Floor standing, end of row Plant room or gallery
Racks served One Several in a row Many rows or a hall
Typical capacity 60 to 250 kW 300 kW to 2.5 MW Several MW
Failure domain One rack One row Every rack on the header
Redundancy Inside the unit Inside the unit plus N+1 units N+1 units and dual headers
Rack space cost 2U to 4U per unit None, uses floor space None at the row
Best fit Retrofits and mixed halls Dense AI rows New builds at scale

The CDU Is the Boundary Between Two Loops

Liquid cooling has two loops with different owners, different water quality, and often different budgets. The facility water system, usually shortened to FWS, is the building side: chillers, cooling towers, dry coolers, pumps, and the headers that run through the hall. The technology cooling system, or TCS, is the IT side: the CDU, the row manifolds, the rack manifolds, the hoses, the quick disconnects, and the cold plates. The CDU is the heat exchanger and pump set that joins them.

That boundary is why the two loops have separate water quality targets. The facility side tolerates more chloride, more hardness, and more bacteria because its passages are larger and its heat exchanger is robust. The technology side feeds microchannel cold plates, so it needs tighter chemistry, finer filtration, and a fluid package that protects mixed metals. Applying the facility water specification to the TCS is a common and expensive mistake, and applying the TCS specification to the building plant is often impossible without a water treatment plant.

The facility water loop and technology cooling loop joined by a CDU TWO LOOPS, ONE HEAT EXCHANGER Facility plant chiller or tower FWS up to W32 or W45 CDU heat exchanger TCS PG25 or water Cold plates 32 C supply 4 K approach 36 C supply 45 C return The CDU never cools the chip directly. It moves heat from the TCS to the FWS and holds the TCS supply temperature close to the facility supply plus the approach. Water quality, filtration and ownership change at this boundary, which is why it has its own record.

What Is Inside a CDU

A liquid to liquid CDU is a small pump station wrapped around a plate heat exchanger. The facility water passes through one side of the exchanger and leaves warmer. The technology coolant passes through the other side, gives up its heat, and returns to the racks at a controlled temperature. Around that core sit the parts that make the loop operable.

  • Pumps move the TCS coolant. Most units use two pumps in a 1+1 arrangement, with one running and one on standby, or N+1 for larger units.
  • A plate heat exchanger transfers heat to the facility water. Its size and approach determine how cold the TCS supply can be for a given facility supply.
  • Filters sit on the TCS side, commonly 25 to 50 micrometers, and protect the cold plate microchannels from debris.
  • An expansion tank and makeup system absorb thermal expansion and replace small fluid losses without letting air into the loop.
  • Sensors and controls read supply and return temperature, flow, differential pressure, and pump speed, and drive the pumps to hold a setpoint.
  • Leak detection and a drain pan catch the failure that matters most, which is a seal or fitting that starts to weep inside the unit.
  • Valves and isolation points allow service without draining the whole row, and a bypass keeps flow available during maintenance where the design calls for it.

A liquid to air CDU replaces the plate heat exchanger with a coil and fans. It is a useful transition product when a hall has liquid servers but no facility water, because it lets a direct to chip rack run without a water connection. The heat still ends up in the room air, which means the room cooling has to remove it, and the density ceiling is lower than a liquid to liquid design.

The main components inside a liquid to liquid CDU INSIDE A LIQUID TO LIQUID CDU Heat exchanger FWS to TCS Pumps 1+1 or N+1 Filter 25 to 50 um Controls flow and temperature Expansion tank and makeup Leak sensor and drain pan Facility water passes through one side of the exchanger. The TCS pump loop serves the racks. The exchanger and pumps set the capacity. The filter, controls and sensors decide how long that capacity holds.

In Rack, In Row, or Centralized

The form factor is a decision about failure domain and space, not about cooling technology. An in rack CDU is a 2U to 4U unit inside the cabinet. It serves one rack, usually up to about 250 kW in current designs, and it takes its facility water from a small branch. Some vendors use a sidecar beside the rack instead, which trades rack units for floor space. Its failure domain is one rack, which is attractive during a phased rollout: you can convert a rack at a time and keep the rest of the hall unchanged. The cost is rack space, more units to maintain, and a shorter TCS loop with less thermal inertia.

An in row CDU is a floor standing unit at the end of a row. It feeds several racks through supply and return headers and commonly ranges from 300 kW to 2.5 MW. It uses no rack units, which matters when every U is revenue, and one unit can serve a dense row without multiplying pump sets. Its failure domain is the row, so production designs usually add N+1 units and sometimes dual headers. A centralized CDU or a gallery design takes the same idea to a hall: fewer, larger units, longer pipes, and a single point of failure that has to be engineered out deliberately.

Typical cooling capacity ranges for in rack, in row and centralized CDUs TYPICAL CAPACITY BY FORM FACTOR In rack In row Centralized 60 to 250 kW 300 kW to 2.5 MW 2 to 10 MW The solid part is the common range and the faded part is the upper end of current product lines. Capacity is always quoted at a set of conditions, never as a single number.

A simple rule of thumb follows from the capacity ranges. If a row of ten 130 kW racks needs about 1.3 MW of liquid cooling, a single 1.5 MW in row unit covers the load at rated conditions, and N+1 means two units. Ten in rack units would cover the same row with ten failure domains and ten filter changes. Past roughly ten racks in a row, the in row design usually wins on unit count and service effort.

Approach Temperature Decides the Plant

Approach temperature is the difference between the facility water supply and the TCS supply. It is usually 2 to 5 K, and it is the number that connects the CDU selection to the building plant. A CDU rated at 1.5 MW with a 4.5 K approach on 32 C facility water delivers coolant at about 36.5 C. The same unit on 27 C water delivers about 31.5 C, and it may also deliver more capacity, because a colder facility side improves heat transfer.

This is where the ASHRAE water classes enter. The classes are named after the maximum facility water supply temperature, so W17 means up to 17 C, W27 up to 27 C, W32 up to 32 C, W40 up to 40 C, W45 up to 45 C, and W+ means above 45 C. Every class has a lower limit of 2 C. The class describes the facility side, not the coolant at the cold plate, which is why the approach has to be added before you compare the number to a server inlet specification.

A worked example makes the relationship concrete. A GPU vendor specifies a maximum coolant inlet of 45 C. The facility is designed to W32, so the facility water can reach 32 C on a design day. A CDU with a 4 K approach supplies about 36 C, which is inside the 45 C limit with 9 K of margin. The same vendor on a W45 plant would need an approach of zero to stay under 45 C, which is not a practical design, so W45 facility water is only usable with equipment rated above 45 C. The class and the approach are a pair.

The other constraint is dew point. If the TCS supply drops below the room dew point, moisture condenses on cold plates, hoses, and manifolds. Design guidance is to keep the coolant at least 2 C above the room dew point, and a warm water loop makes that easier to guarantee. A very cold loop is not a safety feature. It is a condensation risk and a chiller bill.

Facility water classes, CDU approach and the coolant temperature at the cold plate FROM FACILITY WATER TO COLD PLATE 2 C 17 C 27 C 32 C 40 C 45 C W17 chiller W27 warm water W32 tower W40 to W+ heat reuse 4 K A W32 plant plus a 4 K CDU approach gives about 36 C coolant. The chip inlet limit decides whether that fits. The 4 K bar is the approach, and it is the part of the chain a CDU vendor controls.

Size It From Heat Load, Flow, and Delta T

A CDU rating is meaningless without its conditions. Two units can both say 1.5 MW while one is quoted at a 3 K approach and the other at a 5 K approach, with different facility water temperatures, secondary flows, and filter states. Always ask for a selection at your actual facility supply temperature, your actual rack heat load, and your actual design flow. Then check the result against the heat balance yourself.

The heat balance for a liquid loop is the same one used for any sensible heat transfer:

kW = L/min x 0.0697 x delta T      (water)
kW = L/min x 0.0667 x delta T      (PG25)

The inverse is more useful during planning. Water needs about 1.44 liters per minute per kilowatt at a 10 C rise, and PG25 needs about 1.5. The Open Compute Project guidance for cold plate loops uses a 7.5 to 12 C rise, which maps to roughly 1.25 to 2.0 liters per minute per kilowatt, with 10 C and 1.5 as the typical design point.

Heat load Delta T Water flow PG25 flow Note
20 kW 10 K 29 L/min 30 L/min Small CPU loop
60 kW 10 K 86 L/min 90 L/min Single dense server
130 kW 10 K 187 L/min 195 L/min GB200 class rack
130 kW 12 K 156 L/min 162 L/min Wider rise, warmer return
130 kW 7.5 K 250 L/min 260 L/min Tighter rise, more pump energy

A 130 kW rack is a useful reference because it is close to the current GB200 NVL72 class, where roughly 115 kW goes to liquid and the rest to air. At a 10 C rise that rack needs about 190 liters per minute, which is why a rack manifold has many parallel branches and why a single quick disconnect is never the whole answer. The flow is split across cold plates, and the pressure drop budget has to be checked branch by branch.

for kw in 20 60 130; do
  for dt in 7.5 10 12; do
    awk -v q="$kw" -v d="$dt" 'BEGIN { printf "%d kW at %g K -> %.0f L/min water, %.0f L/min PG25\n", q, d, q*14.4/d, q*15/d }'
  done
done
20 kW at 7.5 K -> 38 L/min water, 40 L/min PG25
20 kW at 10 K -> 29 L/min water, 30 L/min PG25
20 kW at 12 K -> 24 L/min water, 25 L/min PG25
...
130 kW at 10 K -> 187 L/min water, 195 L/min PG25
130 kW at 12 K -> 156 L/min water, 162 L/min PG25

The delta T choice is a trade. A wider rise cuts flow, pipe size, and pump energy, and it returns warmer water that a heat reuse plant can use. It also leaves less margin between the supply temperature and the chip limit, and it can push the last cold plate in a series path too warm. A narrower rise costs flow and pump energy but keeps every plate closer to the inlet temperature.

Pressure is the other half of the selection. A CDU has a pump curve, and the loop has a pressure drop made of headers, manifolds, hoses, quick disconnects, and cold plates. A dense rack manifold can drop 1.5 to 2.5 bar at design flow, and the CDU must deliver that head at the required flow, not just the flow at zero head. If the datasheet gives only a capacity and a flow, ask for the pump curve and the available head at your operating point.

Redundancy Is Four Separate Decisions

A CDU with redundant pumps is not a redundant cooling system. Availability has to be specified at four levels, and the weakest level sets the outcome.

  1. Inside the unit. Redundant pumps, redundant power supplies, redundant sensors, and an automatic transfer switch where the design calls for it. A 1+1 pump set means either pump can carry the load alone.
  2. Across units. N+1 CDUs on a shared header, so any single unit can be isolated for service while the rest carry the row. This is where in row designs usually add a unit, and where in rack designs may add a second unit to the same rack.
  3. The distribution path. Dual headers, A and B branches, or a ring main. Two CDUs on one pipe is still one pipe.
  4. The plant. Dual facility water paths, dual cooling towers, or a chiller and a tower that can each carry the design load. The CDU cannot create redundancy that the plant does not provide.

The failure domain follows the form factor. An in rack unit that fails takes one rack out of thermal compliance, and the rack may throttle or shut down depending on its inlet limit and thermal mass. An in row unit that fails takes the whole row, unless N+1 units and a header design allow the remaining units to pick up the load. Centralized units concentrate the risk further, which is why they are usually deployed in a redundant pair or a ring.

Cold plates hold very little water, so the thermal ride through time after a pump stops is short. That is why CDU pumps and controls are commonly on UPS power even when the IT load is on a different UPS. A pump that stops during a transfer is not protected by a generator that starts 30 seconds later.

Document the Loop Before the Racks Go In

The documentation that matters for a CDU is the loop, not the brochure. Record the facility source, the design conditions, the rack feed, the CDU as a device, and the ports that connect the TCS side to the row. In Obelinf, that means a cooling source for the plant, a cooling feed for each loop that reaches a rack, and intake and outflow records on the CDU device itself.

The rack summary is the fastest way to check that the documentation matches the design. It reports heat load against cooling capacity, which is exactly the relationship the CDU has to satisfy.

curl --fail --silent --show-error \
  -H "Authorization: Bearer ${OBELINF_API_KEY}" \
  "${API}/cooling/summary/rack/${RACK_ID}" \
  | jq '{rack_name, cooling_capability, cooling_capacity_w, feed_capacity_w, heat_load_w, utilization_pct}'

If the rack shows 130 kW of heat load and 80 kW of feed capacity, the CDU selection and the rack record disagree. One of them is stale, and the discrepancy is easier to resolve before the racks are installed than after the first training run.

Commission With Numbers, Not Opinions

Commissioning is where a CDU selection is proved or exposed. The checks below are the minimum set for a liquid loop that serves IT equipment. The exact acceptance criteria come from the CDU vendor, the server vendor, and the project specification, so treat these as the shape of the test rather than the final values.

Check What to record What a pass looks like
Pressure test Test pressure, hold time, drop No loss beyond the specification, all joints dry
Flow balance Flow at each rack and each branch Within the design tolerance, often plus or minus 5 percent
Approach at load FWS supply, TCS supply, heat load Approach at or below the selected value at rated load
Delta T at load TCS supply and return, total flow Close to the design rise at the measured heat load
Leak detection Sensor zone and alarm test Every zone reports, no bypassed sensors
Sensor calibration Temperature, flow, pressure readings Within the vendor tolerance against a reference
Filtration Filter differential pressure when clean Baseline recorded for later comparison
As builts Valve positions, port map, feed names Matches the documentation and the rack record

The flow balance is the check most often skipped and the one that causes the most mysterious hot chips. A rack manifold with twelve parallel branches does not distribute flow evenly just because the total is correct. Measure each branch, compare it with the design, and adjust the balancing valves before the first workload runs.

Start With One Rack and One Calculation

The first step is not a vendor comparison. Pick the densest rack you plan to install, write down its liquid heat load, choose a design delta T, and calculate the flow. Then ask every candidate CDU for a selection at your facility water temperature, your flow, and your required head. A unit that cannot state its approach and its available head at those conditions is not a candidate yet.

That single calculation will tell you more than a brochure, and it gives you the numbers to record in the rack and feed records before the hardware arrives. The rack management and data center management pages are the place to keep those numbers next to the rest of the inventory.

Frequently Asked Questions

What does a CDU do in a data center?
A coolant distribution unit transfers heat from the technology cooling loop that serves the cold plates to the facility water loop that rejects it. It also filters and circulates the technology loop, controls supply temperature and flow, and provides the pressure and redundancy that the IT equipment needs.
What is the difference between an in rack and an in row CDU?
An in rack CDU sits inside the cabinet and serves one rack, usually from a few kilowatts up to about 250 kW. An in row CDU is a floor standing unit that feeds several racks through row manifolds, with capacities from several hundred kilowatts to multiple megawatts. The in rack unit limits a failure to one rack, while the in row unit can affect the whole row.
What is approach temperature on a CDU?
Approach temperature is the difference between the facility water supply and the technology loop supply. A 4 K approach on 32 C facility water produces 36 C coolant at the cold plates. A smaller approach needs a larger heat exchanger and lets you run the facility water warmer, which improves free cooling hours.
How much coolant flow does a 100 kW rack need?
At a 10 C rise, water needs about 1.44 liters per minute per kilowatt and PG25 needs about 1.5, so a 100 kW rack needs roughly 145 to 150 liters per minute. At a 12 C rise the same load needs about 120 to 125 liters per minute. The rack and cold plate vendor flow schedule overrides the rule of thumb.
Do I need a chiller for direct to chip cooling?
Not always. A CDU designed for ASHRAE W32 or W40 facility water can run on a cooling tower or dry cooler for most of the year. The chip inlet limit decides the class, and the CDU approach sits between the facility water temperature and the coolant temperature at the cold plate.
How is a CDU documented in Obelinf?
Record the facility plant as a cooling source, the loop to each rack as a cooling feed, and the CDU itself as a device with intake and outflow ports. The rack feeding the CDU can be linked to the device ports, and the rack summary then reports heat load against cooling capacity.

Stop reaching for a spreadsheet

Obelinf keeps every subnet, device, circuit, and rack in one live source of truth, with audit logs and a topology view. Free for personal use.

Related Articles