Direct-to-chip cooling for tomorrow's AI densities
Air cooling has reached its limit. At 1,000 W and more per accelerator, the cold plate decides clock rate, PUE and rack density. We develop and manufacture cold plates, manifolds and distributors for direct-to-chip (D2C) systems. Each one is simulated and tested before it goes into series delivery.
More than the accelerator gets hot in the rack.
Direct to chip solves the problem at the die, but voltage regulators, optics, memory and the infrastructure behind them remain. We supply the cooling components for the entire path from chip to dry cooler.
-
CPU & GPU cold plates
Direct to chip with a fine channel structure right beneath the die.
-
Voltage converters & VRMs
Reliably dissipating high current densities in a small area.
-
Memory modules
Even temperature control across long, narrow installation spaces.
-
Optical transceivers
Temperature stability for stable link budgets.
-
Busbars & power rails
Absorbing waste heat in the power distribution.
-
Manifolds & distributors
Distributing volume flow evenly across all nodes.
-
UPS & power supply electronics
IGBT and rectifier stages in the rack infrastructure.
Density · Efficiency · Operational safety
Liquid cooling in the data centre always operates under three constraints at once. More density must not blow the pump budget, and no efficiency gain justifies a leak above the board. We design along all three axes together.
-
Density
Channel structures that absorb 1 kW and more per chip in the smallest area without a hotspot at the edge of the die.
-
Efficiency
Higher inlet temperatures allow free cooling. We design so that your PUE benefits from it.
-
Operational safety
Leak tightness, corrosion resistance and material compatibility across the entire life of the circuit.
What we supply for liquid cooling systems in the data centre
- Cold plates : for CPU, GPU and accelerators, designed around your die layout.
- Manifolds : for even distribution within the rack.
- Busbar coolers : for power distribution in the rack.
- Sonderbauteile : for retrofitting existing air-cooled racks.
We name references in conversation. Many projects run under NDA.
Three designs that make the difference in data centres.
Multipress
Press-fitted tubes in the cold plate: a robust, process-reliable design for large areas and long parts. Available in copper, copper-nickel, stainless steel or aluminium.
Structureflow
Cooling channels integrated directly into the load-bearing structure. That means fewer interfaces, less thermal resistance and one part instead of an assembly.
Turbulators
Up to 40 % more cooling capacity at an acceptable pressure drop. We position them to the millimetre beneath your hotspots.
Everything from one source, from development to series production.
COOLTEC handles your thermal management completely: one contact, one process, from the first idea to series delivery. You join wherever you stand. Development, simulation and design are optional.
- Optional
Development & simulation
We develop, simulate and design your cooling solution. We test it virtually before we produce anything physically.
-
Manufacturability & cost
You already know what you need? We assess manufacturability and optimise your design for series production.
-
Samples & prototypes
Samples in around 6-8 weeks as standard, or 2-4 weeks by express, depending on product and complexity.
-
Series production
From one supplier: highly automated high-tech manufacturing in Germany, low-cost series from Asia.
-
Supply & lifecycle
Framework contracts, stockholding, last-time buy: we secure your supply across the entire product lifecycle.
Datacenter Thermal Guide: direct to chip in practice
Not a product catalogue but design knowledge: which decisions on cold plates and liquid circuits have to be made early, and what they cost later if they go the wrong way.
Sebastian Krüger
Bring your thermal challenge: heat load, installation space, medium, ambient conditions. In the first call Sebastian works out the right approaches with you and puts together the team we need for it, from application engineering, production engineering and thermal engineering. You will not be put on hold.
The questions data centre engineers actually ask us.
Does COOLTEC manufacture cold plates for data centres?
Yes. Cold plates for data centre technology are technically the same task as in power electronics, with three different constraints: low inlet temperature, high volumes and very tight tolerances at the contact surface. The quantities we cover range from a single piece to 100,000 units per year; data centre projects sit at the upper end of that range, where dedicated tooling and automated production steps pay off. Where we are strongest: customer-specific geometry, power supply and busbar cooling, special formats outside standard socket dimensions. Limit, honestly stated: for direct-to-chip cold plates on standard CPU and GPU sockets you are competing with specialised high-volume manufacturers optimised for exactly one dimension. We are not the cheapest supplier there, and we say so upfront.
Read more –>How are busbars cooled in the rack?
With busbars the heat is not generated at a single component but along the entire length of the conductor, which makes it a different task from a chip cold plate. Two routes are common: a cold plate bonded to the bar across its face, or a cooling channel directly inside the conductor profile. The second route is thermally much better but requires the coolant and conductor to be electrically compatible. With coolers at potential you work with insulating hose sections of defined length or with deionised water. Important for your enquiry: state current and permissible conductor temperature, not just a heat load; we calculate from that. Limit: the electrical design of the bar stays with you, and we assess the thermal and manufacturing side.
Read more –>Why does reproducibility matter more than the best single value for data centre cold plates?
Because cold plates in a rack are connected hydraulically in parallel. If the pressure drop scatters between plates, the volume flow distributes unevenly: the plate with the lowest resistance gets the most flow, the one with the highest gets the least, and that is exactly the one that then runs too hot. A batch with consistently good values is therefore worth more than one with individual peak values. In practice that means fixed process parameters, a defined material form and inspection of the critical features to an inspection plan with a measurement report. What creates scatter is rarely the machining but the raw material. Cast and rolled material behave differently during joining, even under the same alloy designation. Limit: a scatter figure for thermal resistance has to be determined by measurement; it cannot be derived from a simulation.
Read more –>Which coolant is used in the data centre, and what does that mean for the material?
Data centres usually run water with inhibitors or a water-glycol mixture, less often deionised water. The material follows the medium, not the other way round: deionised water attacks aluminium because it draws ions out of the material and the protective oxide layer does not stay stable, which produces pitting corrosion and particles in the circuit. For DI water we therefore use copper or stainless steel at the media-wetted surface, if necessary via tubes press-fitted into an aluminium body so that the medium only touches the tube material. What else is in the circuit matters just as much: aluminium and copper mixed together form a galvanic pair without a suitable medium. Limit: two-phase and immersion cooling with dielectric fluids are not our field. A specialist is the right partner there.
Read more –>How narrow can the channel be, and what happens if there are particles in the circuit?
Narrow channels are the route to low thermal resistances, and at the same time the risk in operation: the finer the structure, the more sensitive it is to particles, deposits and biofilm. In a closed circuit with a filter this is manageable; in an existing system without a filter a finely structured plate will clog over time, and the fault does not show up as a leak but as a slowly rising component temperature. That is why we ask about the filter rating in the system before we fix a channel structure, and in case of doubt we recommend the slightly coarser geometry. Limit and recommendation: anyone without a filter in the circuit should not use a finely structured plate. A filter is cheaper than replacing a hundred cold plates.
Read more –>How quickly can series production start for a data centre project?
The schedule is not set by manufacturing but by two steps before it: release of the geometry and procurement of the raw material in the required quantity. For series in the four-digit range, raw material is no longer a stock item but an order with lead time, and a new extrusion profile adds tooling and press dates that lie outside our control. Realistically such a project runs in three stages: design and prototype, pilot series for validation in the rack, then series production with framework call-offs. Practical advice: the prototype should already be hydraulically identical to the series part, otherwise the validation is not transferable. Limit: we state binding dates in the quotation relative to order receipt. A general lead time figure would be unserious at this project size.
Read more –>How do you design the thermal resistance of a cold plate?
We start from the load case: heat load, die size and position, permissible junction temperature, inlet temperature, volume flow and installation space. This yields the Rth budget, which we split across contact resistance, material conduction and the convective share and verify in simulation.
How much pressure drop is realistic?
That depends on your pump and the number of nodes in the circuit. We optimise channel geometry and turbulator use so that the critical spot gets performance while the system pressure drop stays within budget.
Which materials are suitable for the liquid circuit?
Copper for maximum conductivity, copper-nickel against aggressive media, stainless steel for chemical resistance, aluminium for weight and cost. What matters is the compatibility of all materials in the circuit. Mixed installations are the most common cause of corrosion.
How do you ensure leak tightness?
Pressure testing of every part is our standard; for critical applications we additionally test with a helium leak test and document the leak rate per part. We set the target leak rate before design because it determines material and joining process.
In which sizes do you manufacture heat sinks and liquid cold plates?
From a 30 mm microchannel cooler to a 3 m (30,000 mm) cold plate, right up to a fully cooled large precision part. What is possible is not decided by size alone but by material, design and tolerance: give us your dimensions and we will tell you in which design we can produce them.
Can existing air-cooled racks be retrofitted?
In many cases yes, using retrofit cold plates on existing sockets.
Do you also build manifolds and distributors?
Yes. We supply not only the cold plate but the components along the secondary circuit: distributors and collecting lines up to the interface with the primary circuit.
Which quantities and lead times are possible?
Samples in around 6-8 weeks as standard, 2-4 weeks by express. For series production we manufacture in a highly automated plant in Germany; cost-sensitive standard parts come from our joint ventures in Asia.
Do you work under NDA?
Yes, that is the norm for us. We sign your non-disclosure agreement or provide ours. We never pass on customer projects.
Liquid cooling in the data centre: what matters technically
With AI accelerators, the heat load per socket has multiplied within a few generations. Air cooling hits a physical limit here, not a design limit. The following sections summarise how we at COOLTEC approach these tasks.
CPU and GPU cold plates: one design principle, two load patterns
Direct to chip (D2C) describes the heat path, not the component. Whether the cold plate sits on a CPU, a GPU or an AI accelerator changes nothing about the principle, but it does change the channel structure. The heat flux density across the die and the position of the hotspots decide the layout: the same heat load on a smaller area calls for a finer structure and more pressure drop budget.
We manufacture cold plates for both cases to your die layout and socket dimensions. Depending on the case that means an internally structured plate, microchannels or additively manufactured channels, and an embedded copper heat spreader beneath the densest sources where needed. We also supply the components along the secondary circuit: manifolds, distributors and collecting lines up to the interface with the primary circuit.
CPU cold plates have a moderate heat flux density across a larger base plate, but tight tolerances at the socket and high volumes. Reproducibility of pressure drop across the batch matters more here than the best value of a single plate.
GPU and accelerator cold plates carry 1 kW and more on a small area, often with several dies and memory stacks on one package. Here the fine structure directly beneath the die counts, together with a channel layout that covers memory and voltage regulators as well.
How we design a cooling solution
Every design starts from the same boundary conditions: heat load Pv, maximum permissible component temperature, ambient or inlet temperature and the available contact area. Only these determine which cooling strategy is feasible at all.
The four key parameters of a thermal system
Five steps to the right cooling solution
What influences the thermal resistance of an air-cooled heat sink
Materials: thermal conductivity and flow limits
The choice of material determines thermal performance and service life. If the flow velocity in the tube exceeds the maximum recommendation, the medium mechanically removes the protective passive layer, which leads to erosion and leakage.
| Material | Thermal conductivity at 20 °C | Max. flow velocity |
|---|---|---|
| Copper / copper alloys | 305–394 W/m·K | 2,0 m/s |
| Aluminium / Al alloys | 125–210 W/m·K | 1.8 m/s (structures 1–2 m/s) |
| Copper-nickel (CuNi) | - | 3,5 m/s |
| Stainless steel | low conductivity, very good corrosion resistance | 4,5 m/s |
| Graphite (parallel to the layer plane) | up to 2000 W/m·K, technically 140–160 W/m·K | - |
| Sintered ceramics (BN, SiC) | 100–200 W/m·K, electrically insulating | - |
The data we need for your design
The more complete the boundary conditions, the faster we get from a first estimate to a reliable CFD simulation. This checklist comes from the Cool How Report and is the basis of every design discussion.
Direct-to-chip cold plates
With direct to chip, the cooler sits directly on the package. Contact resistance often dominates the entire Rth budget: flatness, roughness, contact pressure and the interface material decide the outcome more than the channel geometry does.
We design the channel structure around the actual heat flux density: fine structures where the die sits, larger cross-sections in the feed and return. That keeps the pressure drop manageable while the critical area gets full performance.
Pressure drop and pump sizing
Every additional node in the secondary circuit costs pressure. Optimising the cold plate alone just moves the problem into the pump, and with it into operating costs. That is why we treat cold plate, manifold and piping as one system.
We use turbulators selectively: only in the sections beneath the hotspots, not along the entire channel length. That delivers performance where it is needed without raising the system pressure drop unnecessarily.
Materials, media and corrosion
The most common cause of failure in liquid circuits is not cooling capacity but the material pairing. Copper, aluminium and stainless steel in the same circuit without a suitable medium and inhibitor reliably lead to corrosion.
We agree material choice, medium and surface treatment together and document the release. For mixed installations we recommend separate circuits rather than a shared one.
From prototype to rack rollout
The most expensive mistake is a prototype that works thermally but cannot go into series. That is why we assess manufacturability and cost in parallel with the thermal design, not afterwards.
Because development, manufacturing, testing and delivery are all in one hand at COOLTEC, nothing is lost at these transitions, from the first load case to the last-time buy.
Get the Datacenter Thermal Guide
Three details, one click, instant download. After that you decide whether we talk about your project.
Are you planning the switch to liquid cooling?
Send us your load case: heat load per socket, inlet temperature, volume flow, installation space. We will tell you whether and how we can solve it.