Liquid Cold Plate Design Calculations: Heat Load, Flow Rate (L/min per kW) and ΔT

On a lot of data center liquid cooling projects, the numbers don’t line up. The CDU and rack manifold get sized to the rack’s total kilowatts. The cold plate supplier quotes thermal resistance for a single chip, measured under its own test conditions. Nobody notices the mismatch until power-on, when one GPU runs noticeably hotter than its neighbors, or the pump is maxed out and the loop still can’t hit its design flow.

Most of these problems show up on paper long before anyone draws a flow channel, if you run a handful of equations first. This guide breaks liquid cold plate design calculations into five steps, from heat load to flow rate, coolant ΔT, thermal resistance budget, and pressure drop. Each step comes with the formula, the unit conversions, and a worked example, so you can check your own project or sanity-check a supplier’s numbers.

The Five Steps of Liquid Cold Plate Design Calculations

Liquid cold plate design calculations use the energy balance Q = ṁ·cp·ΔT to link heat load, coolant flow rate, and temperature rise. Work through five steps:

1. Define the heat load the liquid must remove.
2. Choose an allowable coolant ΔT.
3. Calculate flow rate and L/min per kW.
4. Check the junction-to-inlet thermal resistance budget.
5. Check pressure drop, velocity, and pumping power.

Key Formulas at a Glance

The Core Equation Behind Every Cold Plate Calculation

This guide covers single-phase cold plates only, where the coolant stays liquid throughout the loop. Two-phase cold plates remove heat through the latent heat of vaporization, so the math is fundamentally different and the equation below doesn’t apply. If you’re evaluating a two-phase approach, take a look at APALTEK’s pump-driven two-phase cooling system.

Single-phase cooling comes down to one energy balance:

Q = ṁ · cp · ΔT

• Q: heat carried away by the coolant (W)
• ṁ: mass flow rate (kg/s)
• cp: specific heat of the coolant (J/kg·K)
• ΔT: coolant outlet temperature minus inlet temperature (K or °C)

In practice, flow is usually specified as a volumetric rate, V̇, in L/min, so you multiply by density, ρ, to get mass flow. Rearranged in L/min and kW, the equation becomes:

V̇ (L/min) = 60 × Q (kW) ÷ [ρ (kg/L) × cp (kJ/kg·K) × ΔT (°C)]

Throughout this guide, Q means heat and V̇ means volumetric flow, so the two never get mixed up. If you work in US units, 1 gpm = 3.785 L/min, a 10°C rise equals an 18°F rise, and for water the equation collapses to the familiar Q (BTU/hr) ≈ 500 × gpm × ΔT (°F).

Always use the ρ and cp of the coolant actually in your loop. Direct-to-chip (D2C) loops most often run deionized water or a 25% propylene glycol-water mix (PG25):

Fluid properties shift slightly with temperature. At an inlet temperature around 40°C, the difference barely moves a flow-rate estimate, but detailed simulation should use properties at the actual operating temperature.

Step 1: Define the Design Heat Load

The Q in the equation isn’t the chip’s nameplate power. It’s the heat that actually ends up in the coolant. Three things matter when you pin it down.

Size for sustained power. AI training keeps GPUs at full load for long stretches, so design the cold plate for sustained operation. For the differences between TDP, cTDP, and other power metrics, see our server CPU cold plate selection guide.

Leave out the heat that still goes to air. Not everything on the board dumps its heat into the liquid. In OCP’s definition of cold plate thermal resistance, the denominator is the ASIC power cooled by liquid, and in many systems the VRs, optical modules, and switch chips still rely on fans. The gap can be large: in a 2014 National Renewable Energy Laboratory evaluation of a rack-level liquid cooling system, the liquid captured up to 64% of server heat, with the rest still going to air. Q should only include the components your cold plates actually cover.

Add it up across the loop, then add margin. One cold plate handles one chip, but the loop has to carry the sum of every cold plate on it, plus whatever headroom your upgrade roadmap calls for. Once you have a prototype, you can run it at peak load with thermocouples at the inlet and outlet and a flow meter in line, then back-calculate the real heat load with the same equation. Researchers characterizing liquid-cooled server racks use exactly this method, and in one study the heat captured by the coolant, estimated from measured flow and temperature rise, landed within 5% of the measured total. That’s more reliable than any spec sheet.

Step 2: Choose an Allowable Coolant ΔT

Coolant ΔT is the temperature difference between the coolant leaving and entering the cold plate or loop. Once the heat load is fixed, flow rate and ΔT are two sides of the same coin. Pick one, and the equation hands you the other.

• A small ΔT means high flow and cooler liquid near the chip, but it costs more pumping power and gives the CDU heat exchanger a smaller temperature difference to work with.
• A large ΔT means low flow, less pumping power, and warmer return water, which helps heat exchanger efficiency and PUE. The catch is a hotter cold plate outlet, which eats into the margin for downstream chips and series loops.

OCP’s OAI liquid cooling guidelines, looking at it from the CDU and facility heat exchanger side, recommend keeping the coolant temperature rise between 7.5°C and 12°C, with 10°C as a typical design target. The guidelines also note the trade-off: a higher temperature rise helps PUE at the expense of cooling performance, and a lower one does the opposite.

That trade-off gets tighter in warm-water loops. When the inlet is already warm, adding a 10°C rise on top leaves noticeably less temperature margin for the chip.

Step 3: Calculate Cold Plate Flow Rate in L/min per kW

L/min per kW is the coolant flow rate supplied for each kilowatt of heat load, and it’s the quickest way to compare flow requirements across chips and racks of different power. Plug the ΔT you picked in Step 2 into the core equation, and you get the flow needed per kilowatt of heat. Because ρ × cp is essentially constant for a given coolant, it simplifies to a quick rule of thumb:

• Water: L/min per kW ≈ 14.4 ÷ ΔT
• PG25: L/min per kW ≈ 15 ÷ ΔT

The table also shows where the industry’s familiar 1.5 L/min per kW comes from. It’s the typical design target for PG25 at a 10°C rise, and OCP’s mapping of 1.25–2.0 LPM/kW to a 7.5–12°C temperature rise is based on PG25’s properties. In other words, 1.5 L/min per kW is a design target worked backward from the loop and heat exchanger side. It isn’t a hard requirement of the cold plate itself. If your loop uses a different ΔT, coolant, or inlet temperature, run the numbers again instead of borrowing the figure.

Worked example: One GPU dissipating 1,000 W of sustained power, cooled with PG25 at 1.5 L/min.

ΔT = 60 × 1 ÷ (4.01 × 1.5) ≈ 10.0°C

With a 40°C inlet, the outlet sits at about 50°C.

For context, OCP’s guidelines put typical flow through a single OAM cold plate at 1–2 L/min, and chassis-level flow for an eight-accelerator system at 4–16 L/min, depending on the loop layout and module power.

Prefer not to do the math by hand? Use the calculator below to size flow, ΔT, and port velocity for your own loop.

[EMBED: Cold Plate Flow Rate Calculator]

Step 4: Turn Temperatures into a Thermal Resistance Budget

With the flow rate in hand, the next question is whether the chip stays within its temperature limit. The most direct way to answer it is to convert that limit into a thermal resistance budget: the maximum total thermal resistance allowed between the chip junction and the coolant inlet.

R_budget = (T_j,max − T_in) ÷ P

• T_j,max: the chip vendor’s maximum junction temperature
• T_in: coolant inlet temperature under worst-case conditions
• P: chip power removed by the liquid

APALTEK’s engineers set cold plate thermal resistance targets on a junction-temperature basis. Here’s how it works with illustrative numbers (not the spec of any specific chip):

Assume T_j,max = 95°C, T_in = 40°C, and P = 1,000 W: R_budget = (95 − 40) ÷ 1,000 = 0.055 °C/W

That 0.055 °C/W has to cover the entire heat path:

• Inside the package (junction to case): set by the chip and its package, and out of the cold plate supplier’s hands. Assume 0.025 °C/W.
• Case to coolant inlet: covers the thermal interface material (TIM), the cold plate itself, and the coolant’s own temperature rise. That leaves 0.055 − 0.025 = 0.030 °C/W.

The coolant’s own temperature rise uses up part of the budget. This is the link between flow and thermal resistance that most often gets missed. When thermal resistance is referenced to the inlet, the coolant has already warmed up by the time it reaches the area under the chip, and that rise counts against the resistance. It scales inversely with flow:

• At 1.5 L/min and a 10°C total rise, the coolant is about 5°C above the inlet halfway through the plate and 10°C above it at the outlet. A hot spot at mid-plate loses about 0.005 °C/W (9% of the budget) to the coolant’s rise; one near the outlet loses up to 0.010 °C/W (18%).
• Cut the flow in half, and the rise doubles. So does the share of the budget it consumes, which leaves less for the TIM and cold plate.

That’s why thermal resistance can’t be discussed apart from flow rate.

Before comparing supplier data, make sure everyone defines thermal resistance the same way. There are at least three common conventions:

The same cold plate produces a different number under each convention. You can only compare thermal resistance figures directly when they share the same convention, flow rate, coolant, and inlet temperature.

For a sense of scale, system-level tests in OCP’s guidelines measured case-to-inlet thermal resistance of roughly 0.014–0.019 °C/W for cold plates on 1 kW OAM thermal test vehicles running deionized water at 1.5–3.5 L/min per plate, with up to a 15% performance penalty when switching to PG25. By that yardstick, the 0.030 °C/W case-to-inlet budget in our example is achievable. Keep in mind that test vehicles are idealized: real packages add hot spots, warpage, and TIM variation, so the margin on an actual chip will be smaller.

Finally, always check the budget against the worst case: the lowest flow and the highest inlet temperature occurring at the same time.

Step 5: Check Pressure Drop, Velocity and Pumping Power

Flow rate isn’t just a question of how much heat you need to move. It’s also a question of what the loop can actually deliver.

Pressure drop climbs faster than flow. In laminar flow, pressure drop rises roughly in proportion to flow rate. Once the flow turns turbulent, it grows closer to the 1.75 to 2nd power of flow. Double the flow, and pressure drop can triple or quadruple, while the temperature improvement is usually far smaller. In one rack-level study, raising rack flow from 4.5 to 8 L/min, an increase of nearly 80%, cut cold plate thermal resistance by 24–26% at most.

The cold plate is only part of the pressure drop. In one set of eight-accelerator (OAM) system tests from OCP’s guidelines, the loop pressure drop reached 17 psi at 8 L/min, with roughly half coming from the cold plates and the rest from hoses, manifolds, and quick disconnects. That’s why the cold plate’s pressure drop budget has to be worked out alongside the rack manifold and the CDU pump head, not on its own.

Pumping power = pressure drop × volumetric flow. Using the numbers above, 17 psi is about 117 kPa and 8 L/min is about 0.000133 m³/s, so hydraulic power comes to roughly 117,000 × 0.000133 ≈ 16 W. Divide by pump efficiency to get electrical power, which will be higher. To compare cold plate options side by side, divide the heat each one removes by the pumping power it needs. This ratio is often called the cold plate’s coefficient of performance (COP). When thermal resistance is similar, the design with the higher COP uses less energy.

Set a velocity limit, too. High velocity wears away copper over time. The Copper Development Association’s design guidance caps water velocity at 4–5 ft/s for hot water systems below 140°F, or roughly 1.2–1.5 m/s, which covers the 40–50°C range typical of warm-water cold plate loops. Keeping ports and tubing at or below about 1.5 m/s is a reasonable ceiling. A quick check: through a port with a 6 mm inner diameter, 1.5 L/min works out to about 0.88 m/s. Push it to 3 L/min, and velocity hits about 1.77 m/s, which is over the limit. At that point you either go to larger ports and tubing or rethink how flow is split.

Worked Example: Cold Plate Flow Rate Calculation for an 8-GPU Loop

Here’s a complete example that ties all five steps together. The numbers are illustrative only and don’t represent any specific platform.

Conditions: Eight GPUs, each with 1,000 W removed by liquid; PG25 coolant; 40°C inlet; designed for 1.5 L/min per kW.

• Total heat load: 8 kW
• Total flow: 8 × 1.5 = 12 L/min (about 3.2 gpm)
• Total loop temperature rise: 10°C, with return water at about 50°C

The two most common layouts for eight-accelerator servers, as described in OCP’s guidelines, are four parallel paths with two cold plates in series each (4×2), and all eight in parallel (8×1). At the same total flow of 12 L/min:

In the 4×2 layout, the downstream GPUs see coolant about 5°C warmer at the inlet. But each of those cold plates also gets twice the flow, and the stronger convection partly offsets the preheat. The price is a much higher pressure drop per path and inlet velocities that can creep past 1.5 m/s. The 8×1 layout avoids preheat entirely and has a lower pressure drop, but it only works if flow divides evenly across all eight paths.

An analysis in OCP’s guidelines backs this up. With PG25 at a system flow of 10 L/min, the 4×2 loop had a pressure drop of about 10 psi versus about 7 psi for 8×1. In the 8×1 loop, all chips stayed within 1°C of each other, compared with about 7°C for 4×2.

Which layout makes sense depends on how much pump head the CDU has to spare, how much room there is for tubing in the chassis, and how tightly chip temperatures need to match. For the engineering considerations behind that choice, see our server GPU cold plate selection guide.

Common Calculation Mistakes

• Using water properties for a glycol mix. PG25’s ρ × cp is about 4% lower than water’s, and the gap widens at higher glycol concentrations.
• Plugging L/min in as kg/s. Multiply volumetric flow by density first, and keep the units consistent.
• Looking at ΔT and ignoring thermal resistance. To hit the same ΔT, PG25 needs only about 4% more flow than water. But it’s more viscous and transfers heat less effectively, so the penalty in thermal resistance and pressure drop is far bigger than the difference in flow.
• Comparing thermal resistance measured under different conditions. If the flow rate, coolant, inlet temperature, or resistance convention differs, the numbers aren’t comparable.
• Splitting flow evenly by rack kilowatts. Each cold plate has its own power and flow resistance, and flow in parallel paths drifts toward the path of least resistance.
• Ignoring how viscosity changes with temperature. Glycol solutions get much more viscous at low temperatures, so actual flow on a cold winter day can fall short of the design point.
• Only checking nominal conditions. The condition you actually have to hold is the lowest flow combined with the highest inlet temperature.

From Hand Calculations to a Validated Cold Plate

Hand calculations get the flow rate, ΔT, thermal resistance, and pressure drop constraints lined up before you spend money on simulation and prototypes, and they rule out options that were never going to work. Getting a cold plate to a final design takes a few more steps:

• Thermal resistance targets driven by junction temperature. APALTEK’s engineers work backward from the chip’s junction temperature to set the cold plate’s thermal resistance target, budgeting the package, TIM, and coolant temperature rise together instead of chasing the lowest possible number in isolation.
• Simulation, then measurement. Simulation settles the flow channel design. Then prototypes go on the test bench, where we measure flow rate against thermal resistance and pressure drop, so you can pick the operating point that fits your loop. Specific test items are defined around each customer’s validation requirements.
• System-level matching. APALTEK’s server and GPU cold plates use brazed copper construction. APALTEK also develops and manufactures its own manifolds and CDUs in-house, so a single team can work out cold plate pressure drop, manifold flow distribution, and CDU pump head together. That cuts down the interface mismatches that come with juggling several suppliers.

For cold plate materials, flow channel structures, and manufacturing processes, see our cold plate design guide, or browse APALTEK’s custom cold plates.

FAQ

How do you calculate the flow rate for a liquid cold plate? Divide the heat load by the coolant’s heat capacity and the allowed temperature rise: V̇ (L/min) = 60 × Q (kW) ÷ [ρ (kg/L) × cp (kJ/kg·K) × ΔT (°C)]. For example, a 1 kW chip cooled with PG25 at a 10°C rise needs about 1.5 L/min.

How many L/min per kW does a liquid cold plate need? There’s no single number. It depends on the coolant temperature rise you allow. For PG25, flow per kilowatt is roughly 15 divided by ΔT. A common design target is 1.5 L/min per kW, which corresponds to a rise of about 10°C. OCP’s guidelines recommend a 7.5–12°C rise, which works out to 1.25–2.0 L/min per kW with PG25.

What coolant ΔT is typical for direct-to-chip cooling? The usual range is 7.5–12°C, with 10°C as a typical design value. A larger rise means warmer return water, which helps heat exchanger efficiency and PUE but squeezes the chip’s temperature margin. A smaller rise does the reverse.

Does higher flow always lower chip temperature? Not proportionally. More flow reduces the coolant’s own temperature rise and improves convection, but the gains taper off. In one rack-level study, nearly 80% more flow cut cold plate thermal resistance by about a quarter at most. Meanwhile, pressure drop and pumping power climb faster, and high velocities raise the risk of erosion. The better approach is to pick an operating point on the cold plate’s flow vs. thermal resistance and pressure drop curves, rather than just turning up the flow.

How does water vs. PG25 change the calculation? To hit the same ΔT, PG25 needs only about 4% more flow than water, so the flow difference is small. The real differences are in thermal resistance and pressure drop. OCP’s guidelines note up to a 15% drop in cold plate performance with PG25, and its higher viscosity also means more pressure drop. If you’re running PG25, leave more room in both the thermal resistance budget and the pump head.

Send Your Numbers to Our Engineering Team

Liquid cold plate design calculations aren’t complicated, but every number is tied to the rest of the loop. Heat load sets the flow. Flow sets the ΔT. The ΔT eats into the thermal resistance budget. And pressure drop limits how much flow you can push in the first place.

If you’re planning a liquid cooling loop for AI servers or other high-power equipment, send APALTEK’s engineering team the following:

• Sustained power and maximum junction temperature for each chip
• Coolant type and maximum inlet temperature
• Available flow and pressure drop budget (or CDU pump head)
• Loop layout (parallel, series, or mixed) and port sizes
• Chassis space constraints

We’ll check whether your flow rate, ΔT, and thermal resistance budget can all be met at once, flag any bottlenecks in the loop, and recommend a cold plate channel design and system-level match.