Core Definition & Working Principle

Main Structures & Manufacturing Processes

• Skived fin / microchannel cooling plates (mainstream for supercomputers): 0.1–1 mm microchannels or fins are precision-machined or etched on copper or aluminum substrates. They feature large heat exchange areas and low thermal resistance (down to 0.02 °C/W). MLCP (Microchannel Cooling Plate Integrated Package) further integrates the cooling plate with the chip IHS, eliminating the intermediate TIM layer and reducing thermal resistance by more than 40%, suitable for 1500–2000 W GPUs/CPUs.
• Tube-embedded cooling plates: Copper tubes are embedded into milled grooves on the base plate and sealed by welding. Costs are about 30% lower than microchannel types, making them suitable for medium-to-high power general nodes, though with slightly higher local thermal resistance.
• 3D-printed cooling plates: Produced via SLM technology using copper alloys with topology-optimized flow channels. Complex channels can be customized, improving heat dissipation efficiency by 30%, but high mass-production costs limit use to customized supercomputer components.
• Blown / extruded cooling plates: Low cost and high production speed, but limited thermal performance; generally not used for core computing chips.

Key Technical Specifications & System Support

• Materials: Copper (thermal conductivity 401 W/(m·K), preferred for heat exchange), aluminum (lightweight and low-cost for auxiliary components). High-end models use copper‑tungsten alloys to balance thermal conductivity and coefficient of thermal expansion.
• Sealing & safety: Dual O‑rings + vacuum brazing, leakage rate < 10⁻⁶ mL/h. Equipped with pressure / liquid leakage sensors and automatic shut-off valves.
• Coolants: Deionized water (low cost, high specific heat capacity), water‑glycol (antifreeze), electronic fluorinated fluid (insulating, for leak‑sensitive applications).
• CDU & control: Temperature control accuracy ±0.5 °C, adjustable flow rate to avoid excessive temperature differences between chips.
• Thermal performance: Heat flux density up to 100 W/cm²+, chip surface temperature difference < 5 °C.

Typical Supercomputer Applications

• Summit / Sierra (Oak Ridge / Lawrence Livermore National Laboratory, USA): Adopt hybrid cooling with direct cooling for all CPUs and GPUs. Cooling plates handle 90% of the heat load. Cooling water temperature exceeds 40 °C, greatly reducing system energy consumption.
• New-generation domestic exascale supercomputers (e.g., follow-up models of Sunway, Tianhe): Widely use cold plate liquid cooling. Some adopt MLCP and two-phase cooling (boiling heat absorption of phase-change fluids), further improving cooling efficiency and reducing pump power consumption by 30%–60%.

Challenges & Development Trends

• Cost & manufacturing: Microchannels and MLCP require extremely high machining precision, and yield directly affects cost.
• Maintenance: Long-term operation requires high coolant purity and clean pipelines to prevent corrosion and fouling.
• Trends: Integration of cooling plates and chip packaging, two-phase cooling, hybrid immersion + cold plate solutions, and AI‑based predictive control for CDU flow and temperature.