Direct-to-Chip Liquid Cooling and CDUs
Above roughly 40 kW a rack cannot be cooled by air in any practical hall, which is why every dense GPU deployment now runs liquid to the chip. A cold plate sits on each GPU, a coolant distribution unit isolates the clean rack loop from facility water, and the facility side runs warm, typically 30 to 40 degrees supply, because warm water is cheaper to make. The design numbers are flow rate and temperature rise, and both fall out of one equation that every operator should be able to do from memory.
TL;DR: Every watt that goes into a rack comes out as heat, so a 120 kW rack rejects about 409,000 BTU per hour and there is no airflow that removes that from a normal cabinet. Direct-to-chip liquid cooling puts a cold plate on each GPU and CPU, circulates a treated coolant through the rack, and hands the heat to a coolant distribution unit which transfers it to facility water through a plate heat exchanger. The CDU is sized 10 to 15 percent above rack load, and the flow rate follows from the heat equation: at a 10 degree rise, 120 kW needs about 2.9 kilograms per second, which is roughly 172 litres per minute. Facility water runs warm on purpose, commonly 30 to 40 degrees supply, because warm water can often be made with dry coolers rather than chillers. The operational requirements that get skipped are water quality, filtration to about 50 microns, conductivity monitoring and corrosion inhibitor, and they are what determines whether the loop still works in year three.
The heat equation, which is the whole design
heat carried by a liquid loop
Q = m_dot x c_p x delta_T
Q heat removed, in watts
m_dot mass flow rate, kg/s
c_p specific heat of water, about 4,180 J per kg per K
delta_T temperature rise from supply to return, in K
solve for flow at a 120 kW rack and a 10 K rise
m_dot = 120,000 / (4,180 x 10) = 2.87 kg/s
water is about 1 kg per litre, so 2.87 L/s = 172 L/min
the same rack at other temperature rises
delta_T = 5 K: 345 L/min more pumping power, gentler on components
delta_T = 10 K: 172 L/min the common design point
delta_T = 20 K: 86 L/min less flow, hotter return, tighter margins at the chip
sanity: halving the flow doubles the temperature rise for the same heat, so flow and delta_T
trade directly, and the chip's maximum inlet temperature is what bounds the choice
cross-check against the published figure
120 kW x 3,412 BTU/hr per kW = 409,440 BTU/hr
which matches the roughly 409,000 BTU/hr that deployment guides quote for a rack of this
size, so the model and the vendor number agree
The two loops, and why they are separate
The separation exists so that the fluid touching a cold plate is one the operator controls. Facility water carries whatever the building's loop carries: particulates, dissolved minerals, and biology. A cold plate has channels measured in fractions of a millimetre and fouls quickly. The CDU's plate heat exchanger moves the heat across without mixing the two, and everything on the rack side is treated, filtered and monitored.
What the requirements list actually contains
| Requirement | Typical figure | Why it exists |
|---|---|---|
| CDU capacity | 10 to 15% above rack load | Margin for peak draw and for fouling as the loop ages |
| Facility supply temperature | 30 to 40 C | Warm water can often be made with dry coolers, avoiding chiller power |
| Filtration | about 50 microns | Cold-plate channels clog below this |
| Conductivity monitoring | continuous | Rising conductivity means the coolant is picking up ions and becoming corrosive |
| Corrosion inhibitor | per the coolant specification | Mixed metals in the loop corrode without it |
| Leak detection | rack and row level | The failure that takes out hardware rather than just performance |
| Quick disconnects | dripless, rated for the cycle count | Every node service event opens and closes them |
The warm-water point is the one that surprises people coming from air-cooled halls. Cooling to 20 degrees needs a chiller and chillers are a large fraction of facility power. Cooling to 35 degrees can often be done with dry coolers using outside air, which is why liquid cooling can lower total facility power even though it adds pumps.
The first minute of a cooling failure
what happens when flow stops at 1,400 W per GPU
a cold plate has very little thermal mass, so junction temperature rises within seconds
the GPU throttles first, which shows as a sudden collective slowdown across a rack
then it shuts down to protect itself
the whole sequence is tens of seconds, not minutes
what the automation has to do in that window
detect: flow and pressure sensors on the CDU, plus leak detection, plus per-GPU temperature
act: power-cap immediately, which buys time by cutting the heat at the source
then: checkpoint if the job can, then orderly shutdown
sanity: at 120 kW a rack has no thermal ride-through worth planning around, so the response
has to be automatic; a human paged at 3 a.m. arrives after the hardware has already
protected itself
Thermal, Power and Cooling Events covers the incident side. The design point here is that liquid cooling removes the buffer air cooling gave you, so the monitoring and the automatic response are part of the cooling system rather than an addition to it.
What interviewers are listening for
The flow calculation, done from the heat equation rather than recalled. Anyone can say "liquid cooling is required above 40 kW"; the useful answer computes 172 litres per minute at a 10 degree rise for a 120 kW rack and explains the trade against a larger delta. The second signal is warm water: knowing that facility supply is deliberately 30 to 40 degrees, and why, separates people who have read about it from people who have specified it. The third is water quality, because filtration, conductivity and inhibitor are the items that determine whether the loop is healthy in year three and they are almost never mentioned.
Key takeaways
- Q = m_dot x c_p x delta_T is the whole design: 120 kW at a 10 K rise needs 2.87 kg/s, about 172 litres per minute.
- A 120 kW rack rejects roughly 409,000 BTU per hour, which matches the vendor figure and confirms the model.
- Two loops separated by a CDU plate heat exchanger, with the CDU sized 10 to 15 percent above rack load.
- Facility water runs warm at 30 to 40 degrees on purpose, because dry coolers are cheaper than chillers.
- Filtration to about 50 microns, conductivity monitoring and corrosion inhibitor are what keep the loop working over years.