CDU Control: Temperature, Pressure & PID for AI Cooling

October 28, 2026
Latest company blog about CDU Control: Temperature, Pressure & PID for AI Cooling

A CDU control system holds secondary-loop coolant temperature by modulating flow with a control valve or variable-speed pump on a PID loop, while monitoring supply and return temperature, differential pressure, flow and leak status. The market behind it grows 22.65% a year, from USD 6.77 billion in 2026 to USD 18.79 billion by 2031 (Mordor Intelligence, 28 July 2026).

Short answer

Where the PID executes decides loop stability more than which sensor you buy: a DIN-rail module with local logic keeps the valve loop alive when the network is down, while the BMS owns setpoints, scheduling and trending. The five measurements that matter are temperature, differential pressure, flow, leak status and strainer differential pressure.

What a CDU control loop actually does

A coolant distribution unit separates two hydraulic circuits. The primary, or facility, loop rejects heat to chillers or dry coolers; the secondary, or technology cooling loop, runs through cold plates and manifolds in the racks. A heat exchanger links them, and the controller holds secondary supply temperature on setpoint while the IT load swings. That setpoint is bounded below by the room dew point: coolant colder than the surrounding air condenses water on manifolds and uninsulated pipework, so anti-condensation control is a setpoint limit, not a detail.

That is not one loop but a cascade. A temperature loop decides how much flow the racks need; a differential pressure loop decides how fast the pump must run without starving the far end. Safeties — leak, low flow, high temperature — are hard limits, not tuning problems. For where the unit sits in the hall, see our note on the CDU at the centre of liquid-cooled data centres; this article is about the loop inside it.

Secondary-loop temperature control: valve vs VFD

Two actuators can hold supply temperature. A modulating valve changes the heat exchanged — often by blending return water into the supply — and responds in seconds. A variable-speed drive on the secondary pump changes mass flow instead: pump power falls sharply as speed reduces, but the hydraulic response is slower and low flow is easier to destabilise.

Most CDUs need both, and the classic failure is letting both chase one error: the valve trims while the pump ramps, and the loop hunts. Give the pump the slow, bulk decision — differential pressure or flow — and the valve the fast trim.

Differential pressure control and why it precedes temperature

Differential pressure is the pump loop's process variable because flow at every rack depends on it. Get ΔP wrong and each rack's flow moves whenever a neighbour's valve strokes, so temperature control chases a shifting target.

Constant-pressure control holds one ΔP setpoint at any load; proportional-pressure control raises it as flow rises, saving pump energy at part load at the cost of lower ΔP at the far end. Location matters too: measure at the CDU headers for total flow, at the most remote manifold for worst-case rack control.

The five measurements that matter (and two that don't)

Five measurements control the loop and prove it did.

MeasurementWhat it tells youTypical instrumentPoint in the BMS
Secondary supply and return temperatureOn setpoint, and the ΔT the load producesImmersion RTDs in thermowellsTwo analogue inputs; supply is the PID PV
Secondary differential pressureWhether flow reaches the far end, and the cause of a deviationWet differential transmitter at the headers or remote manifoldAnalogue input; the pump PID PV
FlowMass flow, hence heat removed (Q = ṁ * cp * ΔT)Electromagnetic or ultrasonic, full boreAnalogue input and a totaliser
Leak statusLiquid where it should not be — safety, not controlPoint or rope sensors in drip traysBinary input, wired to a relay
Strainer differential pressureFouling: rising ΔP at constant flowWet differential transmitter across the strainerAnalogue input with a high-ΔP alarm

Temperature, ΔP, flow, leak, filter ΔP

Read together, they separate faults that one alarm cannot. A falling ΔT at constant flow points to a bypassing valve; a rising ΔT to fouled cold plates or a choked strainer; a falling ΔP with the pump at full speed to a blocked strainer or a failing pump; a leak contact that closes while everything else looks normal is still a hard stop.

Where to place each sensor

Placement decides whether the loop can be tuned. Put the supply sensor in a thermowell downstream of the heat exchanger and upstream of the first manifold, where flow is mixed — a probe in a dead leg measures the pipe, not the coolant. Fit the flow meter full-bore with the specified straight run, never just downstream of a modulating valve. Wire the leak sensor to a digital input as well as a relay so the interlock works without the network.

Two signals make poor control inputs. Room air temperature describes the hall, not the loop: a liquid-cooled loop can be stable while the room runs warm from hot-aisle recirculation. IT power draw works as feed-forward, to shift a setpoint ahead of the load, but as a PID input it is noisy and moves faster than the loop responds.

When a measurement sits on the air side, our range covers it: the Dwyer MS series spans 50 Pa to 5 kPa at ±1% accuracy and is specified for air and non-flammable compatible gases, the HUBA 699 covers −50 to 5000 Pa with linearity better than ±0.7% FS, and the EE65 hot-film anemometer measures air velocity up to 20 m/s. For coolant-side strainer ΔP, specify a wet transmitter rated for loop pressure.

PID tuning for coolant loops without guesswork

No universal gain set survives contact with a real CDU: the loop contains a valve, a pump curve, a heat exchanger and metres of pipe. Publish the procedure, not a table of magic numbers.

Proportional-only first, integral second

Start proportional-only, with integral and derivative at zero, and raise gain until a step in load gives a quick response with only slight overshoot. Then add integral action slowly until the offset at design load disappears. Add derivative last, and only if valve hunting persists with a working PI loop: with a transport delay of tens of seconds, D amplifies sensor noise faster than it damps the loop.

Hunting, deadband and valve wear

Hunting is not cosmetic: an oscillating valve wears its seat and actuator, and the pump pays for the movement. Usual causes are too much gain, integral windup at low load, two controllers chasing one error, a noisy sensor, and a deadband tighter than the noise. Set deadband near twice the sensor noise, limit output movement, and log a trend through a full load cycle — the trend, not the tuning screen, proves stability. On the P-IOM633 the relay outputs support latch or pulse operation with manual/automatic per channel, which a time-proportioned actuator needs.

Where should the loop live: CDU, DIN-rail module, or BMS?

Three homes are common, and the choice sets latency, failure mode and who may change the logic later.

Latency budget for a liquid loop

A coolant loop does not need microsecond control, but it needs a deterministic PID update and a defined behaviour when the network disappears. A BMS polling a point every 30 or 60 seconds adds jitter to a loop whose valve strokes in seconds, and if the head-end is down, a BMS-resident PID stops controlling at all. The practical split: fast loop local, in the CDU controller or a DIN-rail module; slow loop in the BMS, which owns setpoints, schedules, alarms and trends. Write the setpoint as an analogue value, read back the process value and status, and give the local device a fallback setpoint.

When local logic wins (with the P-IOM633 pattern)

Local logic wins when the loop is single-purpose and the module has both the I/O and the protocol uplink to stand alone. The P-IOM633 programmable I/O module is that pattern: 6 analogue inputs (0–10 V, 4–20 mA, NTC-10K or dry contact, 12-bit), 3 analogue outputs (0–10 V, 4–20 mA or PWM, manual/automatic per channel), 6 digital inputs with a 20 ms filter and 3 relay outputs, with built-in PID and browser configuration at 192.168.5.2 or over a direct USB-RS485 link. It supports BACnet MS/TP at 76800 bps and Modbus RTU from 9600 to 115200 bps. On a single CDU that I/O maps cleanly: supply and return temperature, loop ΔP, flow and strainer ΔP on the inputs; valve or VFD speed reference on an output; pump run, pump fail and leak on the digital inputs; alarm and pump enable on the relays.

What it does not replace is a plant controller for several CDUs with sequence logic and redundancy, or a listed dedicated CDU controller demanded by the specification. There, keep the dedicated controller and use the module for I/O expansion.

When the CDU has to reach the BMS

If the CDU's own controller exposes only Modbus RTU and the head-end is BACnet/IP, a gateway is cheaper than replacing the controller. The DSP-200B converts Modbus RTU to BACnet/IP or BACnet MS/TP across up to 512 points on two RS-485 master ports; the QSP-400B raises that to 1,024 points on four ports, and the QSP-800B to 2,048 points on the same four ports. All are AC/DC 9–24 V, under 3 W, DIN rail, configured from a built-in web page. None of these gateways advertises MQTT — MQTT belongs to MatyContrE software, not the gateway — so plan that layer separately. All are in our gateway and converter range.

Designing for the 2026 density reality

The load behind the loop is what makes it worth engineering. Mordor Intelligence put data centre liquid cooling at USD 6.77 billion in 2026 and USD 18.79 billion by 2031, a 22.65% CAGR, with direct-to-chip taking 42.85% of 2025 share; the same report notes NVIDIA's H200 at roughly 700 W per device and a practical liquid threshold near 20 kW per rack (Mordor Intelligence, 28 July 2026). Above that density, loops per hall grow, and so does the BMS point count.

Two more signals point the same way. Grundfos and CAICT estimate that better coordination of compute, cooling and water can cut total electricity use by up to 12% (China Daily, 9 September 2026) — a claim about controls and pumping, not hardware. The build-out is also greenfield: Saudi Arabia alone holds 3.81 GW, about 67% of the GCC pipeline (Research and Markets, 27 October 2026), while Johor has 825 MW live and 3.9 GW planned, with grid connection the constraint (TNGlobal, 26 October 2026). Specifications are written once, and the instrumentation list is easier to get right then than to retrofit.

The consequences are concrete: more loops per hall means more points per gateway, so size by point count. Heat-recovery evaluation above 1 MW (Mordor Intelligence, 28 July 2026) means rejected heat must be measured — flow, ΔT and something that logs both. And the client's specification will use the vocabulary of ASHRAE TC 9.9 and Standard 90.4.

Commissioning checklist

Commission in this order; each step produces a record the client can keep.

  1. Verify sensors first. Compare supply and return sensors in a bath, confirm the ΔP transmitter is across the strainer, not the pump, and check the flow meter's straight run.
  2. Stroke the actuators. Drive the valve 0 to 100% and back, confirm the VFD follows its speed reference, and record where flow starts.
  3. Run the pump loop alone. Use three loads with temperature control disabled and confirm ΔP holds at the remote manifold.
  4. Step the temperature loop. Run 25%, 50% and 100% load steps and log overshoot and settling time as the acceptance record.
  5. Test the failure modes. Unplug the network and confirm the local loop holds setpoint; simulate a leak and confirm the relay interlock stops the pump.
  6. Check every BMS point. Compare values, statuses, alarms and setpoint write-back against the schedule.
  7. Hand over the tuning record. Gains, deadbands, setpoints, sensor serial numbers and the load-step trends.

The application scenario in our library, data centre liquid cooling loop monitoring and control, follows the same layering: sensors on the loop, local PID at the CDU, gateway to the BMS.

Frequently asked questions

What does a CDU controller actually control?

Secondary-loop supply temperature via valve position or pump speed, plus the differential pressure or flow that sets how much coolant each rack receives. Leak, low flow and high temperature are hard limits.

Should the PID run in the CDU, in a PLC, or in the BMS?

Run the fast loop locally, in the CDU controller or a DIN-rail module; let the BMS own setpoint, scheduling, alarms and trending. A BMS-resident PID adds latency and stops controlling when the network drops.

Which points should be monitored on a coolant loop?

Supply and return temperature, differential pressure, flow, pump status and speed, valve position, leak detection and strainer ΔP. Pump speed and valve position show whether the controller is doing what you think.

Why differential pressure and not just temperature?

Temperature is the result; differential pressure is the cause — a fouled strainer, air in the loop, a closed valve or a failing pump. Without ΔP you see a deviation but cannot diagnose it.

What PID parameters suit a coolant loop?

There is no universal set: proportional-only, then integral until offset is gone at design load, then derivative only if hunting appears. Publish the procedure and the load-step test, not borrowed numbers.

Can a DIN-rail I/O module replace a dedicated CDU controller?

Yes when the loop is single-purpose and the module offers local PID, browser configuration and a Modbus RTU or BACnet MS/TP uplink — the P-IOM633 class. For a multi-CDU plant with sequence logic, keep a dedicated controller.

Sources

Previous Post
Next Post