The Thermal-Power Convergence

Glossary

Conservation identity (power in equals heat out)

The rule that a chip dissipates essentially all the electrical power it consumes as heat, so the thermal load a cooling plant must reject is approximately equal to the electrical load the power plant delivers.

Coolant distribution unit

The device between the facility water loop and the technical loop that flows over the chips; it holds a heat exchanger and the pumps that circulate coolant, and its pump power is a critical, backed-up load in a dense design.

Coupled system (power-and-thermal)

The view, forced by high rack density, that power delivery and heat removal are not two separate problems but one interdependent system in which each choice constrains the other and neither can be designed alone.

Density ceiling

The maximum power per rack a given coolant can carry. Air caps out roughly between twenty and forty kilowatts per rack; direct liquid and immersion lift the ceiling well past a hundred kilowatts.

Overhead ratio

The share of a facility's total energy spent on everything other than compute. It lives in the interaction of power and cooling and can only be improved by optimizing the two systems together.

Parasitic power ledger

An honest accounting of every watt the cooling and power support systems consume, including pumps, fans, compressors, and conversion losses, all of which the power plant must feed and which by conservation return as heat.

Stranded capacity

Provisioned electrical or thermal capacity that cannot be used because the other system is the bottleneck; it represents capital spent on a capability the coupled system forbids you to sell, and co-design is what prevents it.

Thermal mass and the short clock

The heat a system can absorb before overheating. Dense liquid-cooled racks store little heat, so a loss of cooling or pump power can push them into danger within tens of seconds rather than minutes.