Machining, castings and heat sinks

Heat sinks for data centre equipment: high density and liquid cooling

Where air cooling stops being practical, how cold plates change the heat sink design, and the flatness and pressure calls that decide whether a liquid loop works.

QKQumek Engineering Desk··9 min read·Updated
Ø 12 h7Ra 1.6

Key takeaways

  • At high heat flux, the limit is often the heat spreading and the interface, not the amount of fin area.
  • Air cooling still works for many rack parts, but the fan power and the acoustic budget can become the constraint.
  • A cold plate is a pressure vessel and a thermal part at the same time, so flatness, wall thickness and leak integrity all matter.
  • The interface material, the clamp load and the cold plate flatness should be specified as one system.

Why data centre cooling has changed

A server used to spread a few hundred watts across a large board. A modern accelerator can put a much higher heat flux into a small package, and the heat sink has to spread that heat before it can convect it away. The fin field is no longer the only problem.

The air side has also become harder. More heat per rack means more air, higher fan speed and more acoustic noise. In a dense hall, the air delivered to the inlet may be warmer than the design assumption, which reduces the temperature difference available for cooling.

That is why liquid cooling has moved from a specialist option to a mainstream one for high density equipment. It is not that air cannot remove the heat. It is that the air path, the fan power and the floor space become expensive.

The limits of air cooling

An air cooled heat sink has three resistances in series: the interface, the spreading resistance in the base and the convection resistance at the fins. At high heat flux, the spreading resistance can dominate. A larger fin field does not help if the heat cannot reach it.

The base thickness, the heat pipe or vapour chamber and the heat source footprint all affect spreading. A thin base with a small source leaves the outer fins cool. A thick base adds mass and cost but can lower the spreading resistance.

Fan curves and pressure drop decide the actual airflow. A dense fin stack has a high pressure drop, so the fan may operate far from its free delivery point. The real airflow can be much lower than the datasheet headline.

The acoustic limit is often the practical limit. A heat sink that needs a high speed fan to pass the thermal test may fail the acoustic requirement in the same product.

Fin density and pressure drop

Finer fins add surface area but raise the pressure drop and the chance of dust blocking the channels. In a data centre, dust is controlled but not absent. A fin gap that is too tight can clog over the service life.

The useful fin density depends on the airflow available. For a low airflow, wider fins may outperform a dense stack because the air actually passes through the channels. For a high airflow, a denser stack can use the available pressure.

The fin thickness also matters. A thin fin is efficient in material but can be fragile and hard to braze or solder. A thicker fin conducts better along its height but adds weight and blocks more of the channel.

The base to fin joint is a thermal resistance. A bonded or soldered joint is not the same as an extruded or machined joint. The joint quality should be part of the specification, not an assumption.

Liquid cooling and the cold plate

A cold plate replaces the air side with a liquid loop. Heat crosses the interface into the plate, then into the metal, then into the coolant. The fin field becomes an internal channel with a much higher convection coefficient, so the external fins disappear.

The cold plate is a pressure vessel. It has to hold the loop pressure, survive a pressure test and resist corrosion from the coolant. The wall thickness, the joint method and the internal channel design all affect both the thermal and the mechanical result.

A machined cold plate with a skived or milled channel is common for high performance parts. A brazed or friction stir welded assembly can give a larger internal area but adds a joint that must be leak tight. The choice depends on the heat load, the pressure and the volume.

Internal geometry and flow

The internal channel should put coolant where the heat is. A simple straight passage can leave a hot spot under the die. A pin field, a serpentine path or a set of parallel microchannels can raise the wetted area, but each adds pressure drop.

Flow distribution matters as much as the total flow. If one part of the plate sees more flow than another, the surface temperature will vary. A manifold or a flow straightener can help, but it also adds volume and cost.

The coolant chemistry affects the material choice. Aluminium, copper and stainless steel behave differently with water, glycol and inhibited coolants. A mixed material loop can create a galvanic couple, so the wetted materials should be reviewed together.

A filter or a strainer protects a narrow channel, but it adds a service item. The channel size should be set with the loop filtration and the expected service life in mind.

Flatness, clamp load and the interface

A cold plate has to sit flat on the package. A bowed plate leaves a thicker bond line at the edges and a thin one in the centre, or the reverse. The thermal interface material cannot fix a shape error that is larger than its intended bond line.

The flatness callout should be tied to the package footprint, not the whole plate. A plate that is flat over the die area is what matters. The clamp load should be high enough to compress the interface material without bending the package.

Screws and springs set the pressure distribution. A four screw pattern with a stiff lid can leave the centre under compressed. A spring loaded or a belleville washer arrangement can give a more even load, but it adds parts and assembly steps.

Cooling options for high density rack equipment
OptionHeat removalMain constraintService impact
Air with finned sinkConvection to room airFan power, acoustics and spreadingSimple, no liquid in the rack
Heat pipe or vapour chamberSpreading to a remote fin fieldOrientation and gravityPassive, but limited by the wick
Direct to chip cold plateLiquid to a local plateFlatness, pressure and leak integrityLoop connection and drip protection
ImmersionLiquid to the whole boardFluid compatibility and service accessDifferent hardware and procedures

The right option depends on the heat flux, the rack density, the facility and the service model. Many halls use a mix.

Materials, manifolds and serviceability

Copper has a higher thermal conductivity than aluminium, but it is heavier and more expensive. Aluminium is often used for the body with a copper interface plate or a copper liner where the heat flux is highest. The joint between the two metals needs its own review.

The manifold and the quick disconnect are part of the thermal system. A restrictive connector adds pressure drop, and a drip on disconnection is a service risk. The connector should be chosen with the loop and the maintenance procedure.

Serviceability is often forgotten in the thermal design. A cold plate that needs the whole tray removed to access a fitting is a support cost. The layout should leave room for the tool and the drip tray.

What to specify

Give the heat load, the source footprint, the coolant inlet temperature and flow, the allowed pressure drop, the loop pressure and the service life. State the flatness over the source area, the surface finish, the interface material and the clamp load.

For a cold plate, add the leak test, the burst pressure, the wetted materials and the cleaning requirement. For an air heat sink, add the airflow, the pressure drop and the acoustic limit.

The interface is part of the cold plate

A flat plate and a good pad still fail if the clamp load is uneven. Specify the flatness, the pad and the mounting force as one requirement.

Frequently asked questions

When should a data centre part move from air to liquid cooling?

It depends on the heat flux, the rack density, the facility and the acoustic limit. Liquid cooling becomes attractive when the air path needs high fan power or when the inlet air is too warm for the required case temperature.

What flatness does a cold plate need?

It depends on the interface material and the bond line. The flatness should be specified over the heat source footprint, not the whole plate, and matched to the pad or paste and the clamp load.

Is copper better than aluminium for a cold plate?

Copper conducts heat better but is heavier and more expensive. Many plates use an aluminium body with a copper interface or liner. The choice also depends on the coolant and the galvanic pair.

How is a cold plate leak tested?

Usually with a pressure decay or a helium leak test at a defined pressure, followed by a burst test on a sample. The test method, pressure and accept criteria should be agreed before production.

Sources and further reading

  • ISO 8062-3, Geometrical product specifications (GPS). Dimensional and geometrical tolerances for moulded parts
  • ISO 2768-1, General tolerances. Linear and angular dimensions
  • Qumek data centre thermal and DFM notes
QK

Qumek Engineering Desk

Engineering team

The engineering team behind Qumek drawings, DFM notes, tolerances and inspection standards. Every guide is reviewed against the drawings and processes used on live programmes.

Designing a high density thermal part?

Send the heat load, the source footprint and the cooling route. The engineering desk will review the spreading, interface and flatness calls.

Send a drawing
Busbars9 min read

Busbar for data centres and AI racks

Rack density has risen faster than room layouts. The bar now has to carry more current in the same aisle, with less tolerance for heat and voltage drop.

Read

The Qumek dispatch

One useful email a month

Drawing notes, supplier checks, tariff changes and the questions buyers keep asking. No newsletter filler, no daily drip.

One email a month. No spam, no list sharing.