The mechanism
A coldplate places a network of channels close to the heat source. Heat conducts through its metal base and enters coolant moving across the internal surfaces. Increasing contact area with small channels can improve heat transfer, but those channels also resist flow. The plate, pump, plumbing, coolant, and allowed temperatures therefore have to be designed together. Coolant travels from the inlet through channels near hot regions and leaves through the outlet, carrying the absorbed heat. Coolant temperature describes thermal energy transport; the flowing liquid is separate from the electrical power path. The warmed coolant still carries energy that must be transferred through the rack and facility cooling chain; the plate is only the point where that journey begins.
- Heat conducts from the package through the plate’s base into the internal channel surfaces.
- Coolant flowing over those surfaces absorbs thermal energy and warms as it crosses the plate.
- The outlet carries warmed coolant to the larger cooling loop, where energy is transferred onward and the fluid is conditioned for reuse.
Science and assumptions
Flow and temperature rise account for captured heat
Transported heat equals mass flow times specific heat times outlet-to-inlet temperature rise.
Symbols and units
- Q̇
- Heat transported by the coolant
W
- ṁ
- Coolant mass flow rate
kg/s
- c_p
- Coolant specific heat capacity
J/(kg·K)
- T_out
- Coolant outlet temperature
°C
- T_in
- Coolant inlet temperature
°C
ṁ is coolant mass flow in kilograms per second and cp is specific heat capacity in J/(kg·K). Multiplying by the outlet-to-inlet temperature rise gives heat carried by the stream in watts. This measures transported energy, not the die temperature: the chip-to-fluid resistance still determines how hot the silicon must become to transfer that heat.
Assumptions Steady flow of a single-phase coolant with nearly constant heat capacity, negligible heat leakage, and no significant pressure-work or kinetic-energy change. Boiling and mixed-flow loops require a more general enthalpy balance.
Engineering example
Removing 1 kW with water-like at 0.05 kg/s implies a temperature rise of roughly 4.8 K before plant losses.
Distinctions and tradeoffs
Blue and red coolant in the atlas encode temperature, not electrical charge. The liquid transports heat; the chip receives power through a separate electrical path.
Smaller channels can improve heat transfer but increase pressure drop and pumping needs. Material compatibility, flow distribution, and reliable seals also shape the design.
Dense accelerators need continuous heat removal to sustain their operating limits. Liquid cooling makes higher rack compute density practical when the surrounding infrastructure supports it.
Specifications and interfaces
- Heat capture
- Conduction followed by convection
Metal receives chip heat; fluid transports it away.
- Useful balance
- Q = mass flow × heat capacity × ΔT
Steady-state energy balance for the coolant stream.
- Channel design
- Microchannel arrays
CoolIT publishes split-flow examples; no exact GB200 channel geometry is asserted.
Interfaces
- Package heat Input
- Thermal energy through contact interface
- Coolant inlet Input
- Conditioned fluid from server loop
- Coolant outlet Output
- Warmer fluid carrying captured heat
Check your understanding
Why can very small channels create a cooling tradeoff?
- They remove the need for a pump.
- They turn coolant into stored model weights.
- They can improve heat transfer while increasing resistance to fluid flow.
Answer and reasoning
They can improve heat transfer while increasing resistance to fluid flow.
More surface area can help heat exchange, but greater pressure drop raises pumping requirements and complicates flow distribution.
Evidence and scope
Fluid properties, flow distribution and two-phase effects are omitted.
Vary total power and a lumped junction-to-coolant thermal resistance; local flow, spatial spreading and transients remain outside the model.