Cold Plate Cooling Troubleshooting: Why Your Cooling Capacity Is Still Limited (and How to Fix It)

Introduction

You've moved from air to liquid. The cold plates are mounted, the CDU is running, and the chips still throttle under full load. Or seven of the eight GPUs on a loop look fine while one runs 10°C hotter than the rest. That's usually when cold plate cooling troubleshooting starts, and the instinct is to blame the cold plate and order a bigger one.

Often, though, the cold plate isn't the only bottleneck. Heat leaves a chip through a chain: the package, the thermal interface material (TIM), the cold plate base and channels, the secondary coolant, the CDU heat exchanger, and finally the facility water. When any link in that chain falls short, the symptom looks the same: the chip won't cool down. Swap the cold plate without finding the real cause, and you've spent money on a problem you still have.

This guide walks through the problem from a cold plate manufacturer's point of view. For each of the eight most common causes, you'll see the symptom, how to confirm it, and how to fix it, followed by a field checklist ordered from the cheapest checks to the most expensive.

Why Is Your Cold Plate Cooling Capacity Still Limited?

Limited cold plate cooling capacity usually comes down to one of eight causes:

1. Poor contact between the chip and the cold plate (mounting pressure, flatness, or TIM)
2. Too little flow through an individual cold plate
3. Uneven flow across parallel cold plates
4. Coolant supply temperature that's too high
5. Trapped air or bubbles in the loop
6. Fouling, clogging, or corrosion in the channels
7. A cold plate that doesn't match the chip's real heat map
8. A CDU that has run out of secondary-side margin

Start by measuring the temperature stack to find where the temperature is being lost, then fix that link.

Start Here: Find Where the Temperature Is Being Lost

The costliest mistake in troubleshooting is swapping parts on a hunch. Before you touch any hardware, collect these readings. Most come from the server BMC, the CDU controller, or a few temporary sensors:

• Chip junction or case temperature (Tj / Tcase), along with actual power draw at the time
• Cold plate inlet and outlet coolant temperatures (T_in / T_out)
• Flow through each individual cold plate, not just total CDU flow
• Pressure drop across the cold plate (ΔP)
• CDU secondary supply temperature and primary (facility water) supply temperature

With that data in hand, run a quick energy balance: the heat carried away by the coolant equals mass flow × specific heat × (T_out − T_in). With PG25, a design flow of about 1.5 LPM per kW produces a coolant temperature rise of roughly 10°C. If the measured rise is far off what the energy balance predicts, the flow reading, the power reading, or the heat path itself is wrong. Expect a small shortfall even on a healthy system, since some heat always leaves through the air. Our liquid cold plate design calculations guide covers the math in detail.

Next, figure out where the temperature difference is concentrated. Two numbers do most of the work:

• Coolant temperature rise (T_out − T_in). For a given power and coolant, this depends only on flow.
• Chip-to-inlet thermal resistance, (Tj − T_in) ÷ power. This is the thermal resistance of the entire path from the chip into the coolant, in °C/W. Dividing by power lets you compare it against the datasheet, a known-good plate, or the same plate's commissioning baseline. It also rises somewhat as flow drops, so read it alongside the flow.

Cause 1: Poor Contact Between the Chip and the Cold Plate

When flow, supply temperature, and coolant rise all look normal but the chip-to-inlet thermal resistance is clearly above its baseline, the heat is getting stuck between the chip and the cold plate. It's a common problem in the field and an easy one to miss.

Typical culprits:

• Low or uneven mounting pressure. Screws tightened out of sequence or below spec leave one side of the plate clamped and the other side gapped. For paste and gel TIMs, clamping force, not the material's bulk thermal conductivity, largely sets interface resistance.
• Out-of-spec flatness. Both the chip package and the cold plate base have some warpage. When their tolerances stack up, local gaps can open across the contact area.
• Protective film left on. The film shipped on a new cold plate base gets clamped straight onto the chip. It's an easy slip to make during high-volume installs.
• Wrong TIM amount, pump-out, or dry-out. Too little leaves gaps; too much makes the bond line too thick. Over many thermal cycles, paste can also get pumped out of the contact zone.

How to check: Pull the cold plate and look at the TIM imprint. A good imprint shows full, even coverage. Thick on one side and thin on the other, or a bare patch in the middle, points to a pressure or flatness problem.

How to fix: Clean both surfaces, apply fresh TIM (never reuse the old layer), and reinstall to the specified torque and sequence. Switch TIM types if needed; our guide to thermal interface materials for cold plates compares the options.

Cause 2: Not Enough Flow Through the Plate

A cold plate that gets less flow than its design point runs hotter, because its thermal resistance rises as flow drops. Keep in mind that adequate total flow at the CDU doesn't guarantee adequate flow through every plate.

Common reasons:

• Higher system resistance than designed. Pressure drop adds up across quick disconnects, hoses, the manifold, and the cold plate itself. Longer runs and extra fittings push the pump's operating point down its curve toward lower flow.
• Valves not fully open or balancing valves set wrong. A valve adjusted during commissioning never got returned to its correct position.
• Kinked or crushed hoses, or a quick disconnect that isn't fully seated. A half-seated quick disconnect may still pass coolant, but at much higher resistance.
• Pump speed set too low, or the CDU running in an energy-saving mode.

How to check: Measure per-plate flow, or back-calculate it from the coolant temperature rise. Compare the measured pressure drop against the supplier's flow vs. pressure drop curve to see whether it matches normal flow.

How to fix: Reset valves, straighten hose runs, and confirm every quick disconnect is fully seated. If system resistance simply exceeds what the pump can deliver, adjust pump speed or revisit the piping and cold plate pressure-drop budget at the design level.

Cause 3: Uneven Flow Across Parallel Cold Plates

When one or two chips on a loop run hot, the rest are fine, and the hot positions stay the same over time, flow distribution is the usual suspect.

In a parallel layout, coolant takes the path of least resistance. Depending on the manifold layout, branches at one end of the header, or with longer runs, can end up with noticeably less flow. The same thing can happen inside a cold plate: in one parallel-channel layout, simulation showed flow maldistribution between channels becoming more pronounced at higher inlet flow rates. And when cold plates are plumbed in series, the downstream plates receive coolant that's already been preheated, so they naturally run warmer.

How to check: Measure flow in each branch, or compare the coolant temperature rise across each plate. Map the hot chips against the manifold's inlet and outlet positions.

How to fix: Rebalance the valves, move to a reverse-return layout, or redesign the manifold so branch resistances are closer to equal. See our liquid cooling manifold guide for how manifolds distribute flow.

Cause 4: Coolant Supply Temperature Is Too High

When every chip runs hot while coolant rise, flow, and chip-to-inlet thermal resistance all look normal, the problem is upstream of the cold plate: the coolant is arriving too warm.

Chip temperature is roughly the supply temperature plus the rise across the whole thermal path. Raise the supply temperature a few degrees and the chip goes up by about the same amount. No cold plate can make that back. Common causes include:

• Rising facility water temperature, such as in summer conditions or when a cooling tower loses capacity
• A larger approach temperature at the CDU heat exchanger (the gap between the coolant it supplies and the water or air cooling it), often from fouling or a load above design
• A CDU supply setpoint that's simply set too high

That's also why CDU cooling capacity is quoted at a specific approach temperature difference: a given unit removes less heat when that gap is squeezed, for example when facility water runs warmer but the supply setpoint stays the same. So when you troubleshoot supply temperature, look at the facility water temperature and the temperature difference across both sides of the CDU together.

How to check: Log CDU supply temperature and facility water temperature over time, and see whether chip temperatures track them.

How to fix: Adjust the CDU supply setpoint, inspect and clean the heat exchanger, or address the water temperature on the facility side. Our data center liquid cooling design guide explains how facility water classes are chosen.

Cause 5: Trapped Air and Bubbles in the Loop

Air is one of the sneakiest problems in a liquid cooling loop. It collects in cold plate channels and at high points, shrinking the effective heat transfer area. Once the pump draws in air, flow becomes unstable too.

Typical symptoms:

• Temperatures and flow swing in a sawtooth pattern instead of settling at a steady value
• The pump gets noisy, or pressure gauges flutter
• The problem shows up right after a new install, a cold plate swap, or service work

The usual causes are incomplete venting during fill, no vent valve at a high point, low expansion tank pressure that lets the pump inlet drop below atmospheric pressure (which can pull air in through fittings or cause cavitation), or air introduced while connecting quick disconnects.

How to check: Watch for periodic swings in flow and temperature. Inspect high points and vent valves, and confirm pump inlet pressure is in its normal range.

How to fix: Refill and vent in the right order: fill from the low point, vent from the high points, and circulate at low speed for a while to work the air out. Top up the expansion tank pressure if needed. Make venting a mandatory step after every cold plate swap or service visit.

Cause 6: Fouling, Clogging and Corrosion

If chip temperatures have crept up over weeks or months while cold plate pressure drop has climbed, suspect fouling or clogging first.

Microchannel cold plates have very narrow passages and are highly sensitive to particles. Current systems are typically specified so that the coolant carries no particles larger than 50 microns, since anything that reaches skived micro-fins can block flow and degrade thermal performance. Common sources of blockage include:

• Brazing residue, machining chips, and sealant left over from installation
• Biofilm and microbial growth
• Corrosion products formed after the inhibitor is depleted
• Corrosion products from galvanic corrosion between mixed metals in the loop

How to check: Watch the pressure-drop trend. If ΔP keeps rising at constant flow, a blockage is the most likely explanation; first rule out a change in glycol concentration or coolant temperature, since both change viscosity. Inspect the filters and send a coolant sample for analysis.

How to fix: Clean the filters and flush the loop. In severe cases, pull the cold plates and clean them individually. Then fix the root cause: restore fluid quality, replenish the inhibitor, and verify the filtration rating. Our direct-to-chip coolant guide covers water quality parameters and corrosion control.

Cause 7: The Cold Plate Doesn't Match the Real Heat Map

Sometimes mounting, flow, water temperature, and fluid quality are all fine and the average chip temperature looks normal, yet one hot spot keeps exceeding its limit. That usually means the cold plate wasn't selected or designed around the chip's real heat distribution.

Common situations:

• The plate was sized on uniform heat flux (total power ÷ die area), while the real chip concentrates heat in a few core regions
• The densest channel region doesn't sit directly over the hot spots
• The datasheet was checked for a single thermal resistance number, without confirming it holds at the actual operating flow
• The chip was upgraded to a higher-power part, but the cold plate was never re-evaluated

How to check: Get the chip vendor's power map and compare it with the cold plate's channel layout. Then read the cold plate's thermal resistance vs. flow curve to confirm the resistance at your actual operating point. Our explainer on cold plate thermal resistance shows how to read the datasheet.

How to fix: This is the case where redesigning or replacing the cold plate is genuinely the answer. Run a thermal simulation with your measured data first to confirm the new design solves the hot spot, then build and test a prototype. Our liquid cooling thermal simulation guide explains what simulation can and can't validate.

Cause 8: The CDU Has Run Out of Secondary-Side Margin

The system ran fine at launch. Then more servers or GPUs went in, and temperatures started rising across the board. That pattern usually means the CDU's margin is gone.

A CDU can hit its limit in three ways:

• Heat rejection: total heat load exceeds what the CDU can transfer at the current facility water temperature, so supply temperature rises
• Pump head: added cold plates and piping raise system resistance beyond what the pump can push
• Flow: total flow demand exceeds pump capacity, so every cold plate's share gets thinner

How to check: Review the CDU's load, pump speed, and supply temperature trends, and compare them before and after the added load.

How to fix: Re-run the CDU sizing, then add a CDU or move to a larger model if needed. See our overview of what a CDU is and how it works, and our in-rack vs. in-row CDU comparison for sizing at different scales.

A Field Troubleshooting Checklist

Work through these steps in order, starting with checks that don't require downtime or disassembly:

1. Review the data. Pull historical trends for chip temperature, power, CDU supply temperature, flow, and pressure drop. Did the problem appear suddenly or develop gradually?
2. Check the settings. Verify the CDU supply setpoint, pump speed, and operating mode.
3. Check valves and quick disconnects. Confirm every valve is in its correct position, every quick disconnect is fully seated, and no hose is kinked.
4. Check for air. Look for swings in flow and temperature, and inspect vent points and expansion tank pressure.
5. Check flow distribution. Compare flow or coolant rise across branches and individual cold plates.
6. Check fluid quality and filters. Inspect the filters and send a coolant sample for analysis.
7. Check the installation. Power down and isolate the node, pull the cold plate on the hot chip, inspect the TIM imprint and the base, then reinstall with fresh TIM to spec and vent the loop.
8. Check the selection. Compare the chip power map with the cold plate's performance curves to decide whether a redesign is needed.

This order keeps you from tearing down hardware or replacing cold plates before you need to. One server vendor's over-temperature procedure follows the same logic: confirm that room temperature and rack inlet water temperature and flow are within spec, then verify the cold plate is securely installed.

How to Rule These Problems Out Before Deployment

Several of these eight causes can be ruled out at the cold plate design and production stage. Here's what we build into custom projects as a cold plate manufacturer, and what we recommend every buyer ask for:

• Design channels around the real heat map. On custom projects we work from the chip's heat distribution and real operating conditions, then use simulation to confirm cooling over the hot spots instead of designing to average heat flux.
• Inspect before shipping. APALTEK cold plates go through dimensional inspection, leak testing, and functional testing before they leave the factory. Our helium leak testing guide covers leak test methods and limits, and our testing and validation page outlines the full inspection process.
• Get the performance curves. Ask your cold plate supplier for thermal resistance vs. flow and pressure drop vs. flow curves. With them, your integration team can size the CDU and piping during design instead of discovering in the field that the pump can't keep up.
• Spell out the conditions in your RFQ. Coolant type, supply temperature range, allowable pressure drop, and test requirements all belong in the request for quote. Our guide on how to prepare an RFQ for a custom cold plate lists what to include.

FAQ

How can I tell if a cold plate is clogged?

The most reliable sign is a pressure drop that keeps rising over time at constant flow, while chip temperature slowly climbs. If ΔP returns to its original value after you remove and flush the cold plate, a blockage is confirmed.

Can air in the loop cause a chip to overheat?

Yes. Air trapped in the cold plate channels shrinks the effective heat transfer area, and air drawn into the pump makes flow unstable. The telltale sign is a periodic swing in temperature and flow, especially right after a new install or service work.

Why is one GPU hotter than the others on the same loop?

Start with its position. If the hot plate always sits at the same end of the manifold, or on its longest branch, flow distribution is the likely cause. If the position seems random, the mounting or TIM on that plate is more likely. Comparing that plate's data with the others tells you which: a much higher coolant rise means it's short on flow, while a normal coolant rise with a high chip-to-inlet thermal resistance points to mounting or TIM.

Should I raise the pump speed or lower the supply temperature first?

It depends on where the temperature difference sits. If the coolant rise across the cold plate is too high, flow is short; once you've ruled out a closed valve or a clog, raising pump speed helps more. If the coolant rise is normal but every chip runs hot, lowering the supply temperature helps more. If the problem is poor contact, neither will do much.

Will a bigger cold plate fix an overheating chip?

Only if the cold plate itself is the limit. If the cause is low flow, warm supply coolant, air, fouling, or a CDU without margin, a new plate barely moves chip temperature. A higher-performance plate often has denser fins and a higher pressure drop, which can make a flow problem worse. Confirm the cause with the temperature stack before you change hardware.

Is the problem the cold plate or the TIM?

Look at where the temperature difference sits. If flow and coolant rise are normal but the chip-to-inlet thermal resistance is clearly above baseline, the heat is stuck between the chip and the cold plate, so check the TIM and mounting first. Only if mounting and TIM check out and the hot spot still exceeds its limit should you question the cold plate design itself.

Conclusion

When a cold plate seems to come up short, the cold plate itself often isn't the weak link. Something else in the thermal chain is. Measure the temperature stack, work through the checks from cheapest to most expensive, and you'll typically find the cause without replacing hardware you didn't need to.

Send your data to our engineering team. If your system shows any of these symptoms, send APALTEK's engineers your chip power, cold plate inlet and outlet temperatures, per-plate flow, and pressure drop. We'll help you pin down whether the problem is contact, flow, or the cold plate itself. If a new cold plate is the right answer, we'll design it around your measured data and supply samples for validation on your system. Explore our cold plate solutions or contact us to get started.