THE HEAT PATH

From transistor to sky

🌡️

Die Heat is made in the transistors and wires

Lid and TIM A thermal interface material and metal lid spread it

Heat sink or cold plate Moves the heat into air or liquid

Facility Chillers, cooling towers or dry coolers dump it outside

Every layer in the path adds thermal resistance, and the chip must stay below about 85–105 °C.

AIR

Fans and fins

🌬️

Air cooling uses large finned heat sinks and fans. It is simple and cheap, but air holds little heat. Above roughly 30–40 kW per rack, air can't carry heat away fast enough without huge, noisy airflow.

Air cooling ran out of room around the H100 generation.

LIQUID

Direct-to-chip cold plates

A cold plate is a metal block with fine internal channels, clamped on the chip. Coolant, usually water mixed with glycol, flows through it and carries heat to a coolant distribution unit. Water holds roughly 3,500 times more heat per litre than air.

Blackwell racks such as the GB200 NVL72 are designed for liquid cooling.

SORT THEM

Which cooling fits?

🌬️

Air

Cheap, simple; up to about 30–40 kW per rack

💧

Cold plate

Liquid on the hottest chips; over 100 kW per rack

🛁

Immersion

Whole servers in a non-conductive fluid

The hotter the rack, the closer the liquid has to get to the chip.

WATER AND WASTE HEAT

Where the heat goes

♨️

Evaporative cooling towers use a lot of water, so many new sites use closed-loop or dry cooling instead. Some data centers pipe their waste heat into district heating for homes. Microfluidic cooling, with channels etched into the chip itself, is being researched for future chips.

Cooling is now designed together with the chip, not added afterwards.

UNIT 15 STUDY COMPLETE

❄️

Ready for the Fab Challenge?

You've followed heat from a transistor out of the building, and compared air, direct-to-chip liquid and immersion cooling.