Discussions of Cold Plate Liquid Cooling Technology and Its Applications in Data Center Thermal Management

Overview

With 5G, big data and cloud computing, data-centre size and capacity are increasing rapidly and electronic devices are increasingly integrated, raising heat generation power and heat flux density. Typical rack loads quoted are about 7 kW per rack for a traditional data centre against 50 kW per rack for new high-performance servers, expected to exceed 100 kW within 5 years. Liquid specific heat capacity is quoted as 1,000-3,500 times that of air and thermal conductivity as 15-25 times. The paper explains how a cold plate, CDU, pump and chiller fit together, and separates single-phase from two-phase cold plates.

Reported gains: open water-cooled backplanes reach 8-12 kW of cooling per cabinet, while closed water-cooled backplanes (air recirculated inside a closed cabinet) reach 12-35 kW. Cabinets with enclosed water-cooled backplanes can save nearly 35% of annual energy compared with air-cooled backplanes. A hybrid scheme with cold plates on the main loads (for example CPU) and air for the remaining heat (drives, cards) can keep average annual PUE below 1.2. Raising the cold-plate share to 90% (CPU and memory on cold plates) can cut energy use by up to 50% compared with traditional air-cooled data centres.

The heat-recovery benefit of warmer coolant (district heating, absorption cooling and power-plant integration) is covered; recovering heat from a 3.5 MW data centre for district heating is cited as reducing CO2 emissions by 4,102 tons per year. Open challenges: retrofit cost alongside air-cooled servers, leak and corrosion prevention, and the trade-off between higher coolant temperature for reuse and chip cooling margin.

Limits: this is a short opinion paper; all figures are cited from earlier literature with no new measurements, and the cited gains were reported under different conditions and cannot be compared directly.

Key points

1. Typical rack loads quoted: about 7 kW per rack for a traditional data centre versus 50 kW per rack for new high-performance servers, expected to exceed 100 kW within 5 years.
2. Liquid specific heat capacity is quoted as 1,000-3,500 times that of air and liquid thermal conductivity as 15-25 times that of air.
3. Open water-cooled backplanes reach 8-12 kW of cooling per cabinet, while closed water-cooled backplanes (air recirculated inside a closed cabinet) reach 12-35 kW.
4. Cabinets with enclosed water-cooled backplanes can save nearly 35% of annual energy compared with air-cooled backplanes.
5. A hybrid scheme with cold plates on the main loads (e.g. CPU) and air for the remaining heat can keep average annual PUE below 1.2.
6. Raising the cold-plate share to 90% (CPU and memory on cold plates) can cut energy use by up to 50% compared with traditional air-cooled data centres.
7. Recovering heat from a 3.5 MW data centre for district heating is cited as reducing CO2 emissions by 4,102 tons per year.

Using it in design

• For loads of 50 kW per rack and above, evaluate a hybrid cold-plate scheme (main loads on cold plates, remaining heat in air) for a PUE target below 1.2.
• Make the trade-off between higher coolant temperature for heat recovery and chip cooling margin a design criterion.
• Add retrofit cost alongside air-cooled servers and leak and corrosion prevention (CDU, disconnects, material compatibility) to the design specification.

Who should read it

Cooling designers evaluating cold-plate liquid cooling for high-density racks.

Document contents

1. Introduction
2. Cold plate liquid cooling system components: cold plate, CDU, pump and chiller
3. Single-phase and two-phase cold plates
4. Water-cooled rear panels and hybrid cold plate/air schemes
5. Heat recovery from warmer coolant: district heating, absorption cooling, power-plant integration
6. Open challenges and conclusion

Read the document

Publication details

Authors

Yufeng Zhang; Chengcheng Fan; Guanru Li

Publisher

Frontiers in Energy Research (Frontiers Media SA)

Year

2022

Volume

10

Pages

954718

Document language

English

Source type

Scientific

Access

Open licence — downloadable

DOI

10.3389/fenrg.2022.954718

Licence

CC BY 4.0 — share with attribution

Licence evidence

doi.org

Citation