01
What is added
Condenser, receiver with a defined fill level, pressure control and a second, independent heat balance on the condenser side. A plant section of its own, not an accessory kit.
← Capabilities
Capability · 02 · Fluids · Electrical
Data centre cooling and chip emulator
On a single accelerator module two temperature limits apply side by side: 105 °C for the logic, 85 °C for the memory stack next to it. In a published multi-chip module the figures behind them are 295.86 against 52.08 W/cm². A uniformly heated test body averages away precisely this difference. We build the test bench that reproduces the heat source zone by zone, conditions the circuit and makes the gradient visible.
The emulator
How we reproduce the heat source
The chip is not simulated but replaced by a heated surface of known power and known area — it applies the same heat flux density over the same area. The heater plates themselves come from a specialised supplier; we build the test bench around them.
| Quantity | Range (design) | Note |
|---|---|---|
| Power per unit area | up to 200 W/cm² | Design target; in a published multi-chip module the logic chip already reaches 295.86 W/cm² |
| Electrical power | up to 5.1 kW per zone | typical operating point 3,500 W at 240 V; a separate power supply for each zone |
| Zones per emulator | 1 to 4 | a separate control channel and a separate power supply for each zone |
| Load changes | up to about 2 Hz | edge 50 ms rising, 100 ms falling under load — limited by the response time of the power supply |
| Flow temperature | +15 to +60 °C | warm-water cooling included |
| Flow rate | 0.5 to 120 l/min | per test item; test field supply above that |
| Operating pressure | 0.5 to 6 bar | typical secondary circuit in the rack |
| Media | Water-glycol · deionised water · dielectric fluids | conductivity is monitored |
| Heat balance | fluid side ± 3.7 % · electrical ± 0.2 to 0.5 % | at a 10 K temperature difference; the error chain is set out further down |
The characteristic values of the heated surfaces come from the supplier's data sheet and are to be confirmed afresh before every design. What we warrant is the plant around them — control, measuring chain, safety and the test record.
Multi-zone emulation
A heat map instead of a single point
The most important change of the coming years is not the level of the power loss but its distribution. Different limits apply within the same package — and the power is spread very unevenly over the area.
A uniformly heated test body does not represent this. It yields a mean thermal resistance over the whole area — and that says nothing about whether the memory stack stays below its 85 °C while the logic area next to it is allowed to reach 105 °C. That is why we build the emulator with separately controlled zones: for each zone its own control channel, its own power supply, its own power measurement.
A copper spreader in the cooling concept smooths this map over the area but does not remove it: the difference between the logic area and the memory area remains, because different limits apply to the two and the heat flow is still introduced unevenly. What the spreader changes is the resolution required, not the question of whether more than one measuring point is needed. In our view four zones are the sensible upper limit.
- What limits the number of zones is not the electrical side. Power supplies can be stacked. The limit is set by the heat removal of the fluid circuit — four zones at full power are something different from four zones sharing the power of one test item.
- The zone geometry follows the test item, not the test bench. How many zones lie where and how large they are follows from the interface to the test item. An emulator built to the wrong division is scrap, not stock.
- Memory power is growing with it. HBM4E up to 80 W per device, HBM5 around 100 W per stack from 2028. With eight stacks that is 800 W alongside some 2 kW of logic power — on one component whose part that has to be kept cooler occupies the larger share of the area.
Where the figures come from: the temperature limits from the OCP paper “30 °C Coolant — A Durable Roadmap for the Future”, the power densities from a study of jet impingement cooling in the International Journal of Heat and Mass Transfer. The statements on future component generations rest on trade press and market reports, not on manufacturers' data — they indicate the direction, not a committed specification.
Measurement uncertainty
How accurate the thermal resistance really is
A thermal resistance is the quotient of a temperature difference and a heat flow. Both are measured, both carry an error — and the error of the heat flow is regularly underestimated.
Calculated over the fluid circuit, the heat flow is the product of density, flow rate, heat capacity and temperature difference. At a difference of 10 K the flow rate contributes ± 1.5 %, the density derived from the temperature ± 0.2 % — and the temperature difference alone ± 3.4 %, measured with two independent class A sensors. Combined in quadrature, ± 3.7 % remains, and a good four fifths of that comes from the temperature measurement, not from the flow, which is what one suspects first.
√( 1,5² + 3,4² + 0,2² ) % = ± 3,7 %
The error grows in inverse proportion to the temperature difference: at 5 K it is ± 7.0 %, at 20 K still ± 2.3 %. If you run a small temperature difference because the test item demands it, you pay for it in uncertainty. Two routes lead below that:
- A matched sensor pair to 1/10 DIN lowers the temperature share from 3.4 to 0.7 % and the total uncertainty to about ± 1.7 % at 10 K. That is the effective lever — not a more accurate flow meter.
- Determining the heat source electrically instead of on the fluid side: ± 0.2 to 0.5 % via a power measurement independent of the power supply. That is the more accurate route and the reason why we always provide a separate measurement at the emulator.
A commitment such as “U(Rth) ≤ 2 %” cannot be verified without stating the temperature measurement and the temperature difference that was run. With two separate sensors and a 10 K difference it cannot be met. We therefore state the error chain and not only the result — and say which measuring equipment carries the value.
Electrical side
Why the power supply has to be larger than you think
Conventional programmable power supplies have a rectangular performance map: the current limit at 100 V is the same as at 300 V. The adjustable power limit of these units only caps further downwards — at low voltage it does not permit a higher current. Autoranging units with a true power performance map deliver their rated power over a wide voltage range; the calculation here applies to the conventional class.
The consequence is a trap when selecting within this class. At the operating point what counts is voltage times rated current: a unit with 300 V and 17 A carries 5.1 kW on the nameplate but provides 4,080 W at 240 V. Anything up to 4,000 W per zone is thus attainable; 5,000 W would need 20.8 A there and only work above 295 V. One size down, 11.5 A and 3.45 kW on the plate, stays at 2,760 W at 240 V — and that already falls short of 3,000 W.
Between the two sizes there is nothing in a single rack unit. Anyone who selects by rated power therefore easily takes one size too small — and notices it only at commissioning, when the zone fails to reach its setpoint.
The read-back of the power supply is a display of the manipulated variable, not a measurement. For the heat balance we measure the power fed in a second time via a separate shunt with an isolated transducer. Above 120 V DC, extra-low voltage protection no longer applies — DC-side isolation, protection against contact and integration into the safety circuit belong to the emulator and not to the accessories.
Layout
Two racks, one circuit
A test bench of this kind consists of two separate fluid circuits and two separate racks. The separation is not a matter of aesthetics but the precondition for servicing the media side and the electrical side independently and extending them later.
The conditioning circuit carries mains water: through a filter and a circulating pump, an electric heater adds heat, a control valve lets warm water out into the drain and carries heat away with it. The process circuit carries the test medium — water-glycol, deionised water or a dielectric fluid — from the tank via the manifold, the test item and the collector and back again. The two are coupled solely through the heat exchanger. Mains water and test medium touch nowhere, and that is precisely why the medium stays clean and in a known condition.
What is controlled is the temperature at the transition, not the one at the test item. The manipulated variable is split for this: in one half of the control range the cooling valve opens, in the other the heating element heats, and between them lies a dead band so that the two do not work against each other. A feedforward from the commanded emulator power anticipates the load step instead of only seeing it as a control deviation.
The process pump runs on demand via a frequency converter. It does not deliver the maximum flow continuously but follows what the connected branches draw, plus a control reserve for the overflow valve that holds the manifold pressure. That does not only save energy — it also keeps the pump's heat input into the circuit small, and that input otherwise goes straight into the heat balance.
- Water and electrical equipment stand apart. Tank, pump, heat exchanger and valves in one rack; incoming supply, control, measuring cards and the emulator power supplies in the other. A leak in the media rack then does not reach the electronics by a short path.
- One power supply per zone, one rack unit per power supply. Extending to several zones is therefore a question of free rack units and an adequately sized incoming supply — both can be provided for from the outset without any of it having to be paid for at the same time.
- Kept free means space, not a port. The manifold and the collector get no blind outlets held in reserve. A capped port is a dead corner: air cannot be got out of it, and when the medium is changed residual medium stands in it. What is kept free is the mounting level, not the connection.
- The operator panel hangs on an arm, not in the front. A 21-inch panel does not fit into a 19-inch rack front — the opening is 482.6 mm wide, the housing distinctly wider. It therefore sits on a height- and swivel-adjustable arm on the rack, which in any case gives the better working posture at the test bench.
Option · Two-phase
You can add the two-phase module at any time
Evaporation removes heat at a nearly constant surface temperature. To test that you need a plant section of its own — and it can be added at any later point, because the basic build keeps space, the connection point and the necessary channels in the control system free from the outset.
02
What changes in the measurement
In single-phase operation the balance of mass flow, heat capacity and temperature difference is enough. As soon as part of the fluid evaporates, the heat sits in the enthalpy of vaporisation and the temperature difference tends towards zero — what is then measured is the vapour quality via the condenser.
03
What the basic build already provides
Rack units in the media rack, the connection point on the return and the inputs and outputs in the control system. What is kept free is the mounting level, not a capped port — that would be a dead corner in the circuit.
04
When the design becomes firm
As soon as the fluid, the pressure level and the power range are settled. After that the price is a figure instead of a range — and the scope follows from what you want to test, not from an assumption of ours.
That we list it as an option and not as a fixed part has a reason outside the technology. At the end of 2022, 3M announced that it would stop manufacturing fluorinated substances by the end of 2025 — the products on which a large part of two-phase immersion cooling was built. What is offered as a replacement today falls largely into the same class of substances.
In parallel the European restriction of the whole substance group is under way: the opinion of the Committee for Risk Assessment is available, that of the Committee for Socio-economic Analysis as a draft; the consultation on it closed in May 2026 and the final version is expected at the end of the year. Only after that does the legislative procedure begin; for semiconductor and medical applications, time-limited exemptions of up to twelve years are under discussion.
For you that means: the test bench does not have to anticipate this decision. It stays ready to be connected to, and you buy the two-phase section once you know which fluid you will be working with.
Sources: the manufacturer's withdrawal according to trade reporting, the state of the procedure according to the publications of the European Chemicals Agency and accompanying legal analyses, as of August 2026. The procedure is ongoing — the dates named are expected ones, not decided ones.
Measuring chain
What makes a measuring chain usable
A cold plate is only as good as the measurement by which it is judged. Four points at which it tips:
Differential pressure before flow
A flow characteristic without a clean Δp measurement is worthless. Resolution, zero point, mounting position and the influence of the pipework belong in the test record, not in a footnote.
Temperature at the right place
Flow, return and surface, with a documented mounting position. Two kelvin of measurement error at the wrong place turn a good cold plate into a bad one.
Leak tightness before commissioning
Water in a rack is more expensive than any test that would have prevented it. Pressure decay in series production, helium where the detection limit counts.
Traceability
Every measurement with a time stamp, the test equipment number and the calibration status. Anyone who has to explain a result a year later is glad of it.
Test items
Four test items, four tasks
From the single component to the finished distribution unit — on the same test bench.
01
Cold plate and cooling manifold
Emulation of the heat source with a defined power per unit area, measurement of thermal resistance, pressure drop and temperature distribution. The influence of the contact pressure is measured along with it.
02
Manifold and distributor
Even distribution across many outlets, pressure drop over the flow rate, cavitation limit. A manifold that is right on average and not at the edges only shows up in the field.
03
Quick coupling (UQD)
Mating cycles up to the manufacturer's figure — commonly up to 5,000 — with measurement of residual leakage and flow over the cycle, not only before and after.
04
Coolant Distribution Unit
Final test of the finished unit: capacity over the flow temperature, control quality, leak tightness, behaviour during a hot swap of the pump.
Design
This is the window we work in
| Feature | Value (design) | Note |
|---|---|---|
| Test items | Cold plate · manifold · UQD · CDU | Component to finished unit |
| Power per unit area | up to 200 W/cm² | Chip emulator |
| Load step | Tenths of a second | Control quality test |
| Flow temperature | +15 to +60 °C | incl. warm-water cooling |
| Flow rate | 0.5 to 120 l/min | per test item |
| Operating pressure | 0.5 to 6 bar | Secondary circuit |
| Media | Water-glycol · deionised water · dielectric fluids | Conductivity monitored |
| Leak tightness | Pressure decay · helium | before commissioning and in series production |
| Traceability | Test record per test item | Time stamp, test equipment no., calibration status |
Design values, not a project commitment. Reference from practice: an emulation of a chip cooling system at 2,500 W has been built; a complete direct-to-chip test bench with load-step capability has been designed. What is possible in your plant is stated in the functional specification.
From the project
A look inside
Photographs of the plant will follow — instrumentation, cooling circuit and the emulator.
A cooling solution to test?
Tell us the test item, the power per unit area, the medium and the load profile — we will design the direct-to-chip test bench for it.
