1 of 4

MPCDF Raven�Direct Liquid Cooling

Karsten Kutzer

2022 Lenovo internal.

2 of 4

How and why adopt liquid cooling?

  • How to adopt liquid cooling?
    • Different technologies available – e.g. Direct-Water-Cooling (DWC) and Immersion Cooling
    • Different implementations of DWC are available, e.g.
      • DWC assemblies by specialized providers added to systems or vendor designed and integrated DWC systems
      • Loop materials and fusing (e.g. fused plastic/EPDM vs. welded copper)
      • Components on cooling loops (e.g. CPU only vs. all components up to the PSU)
    • Direct-water-cooling with welded copper loops is a mature technology for a decade now in large HPC systems

  • Why to adopt liquid cooling?
    • Energy savings on the cooling
      • at system level (eliminating fans and air cooled power supplies)
      • at data center level (allowing “free cooling” concepts with dry or hybrid recoolers)
    • Increasing cooling efficiency, supporting higher TDPs / lower Tcase of latest CPUs and GPUs
    • Decreasing the carbon footprint of the data center
      • reducing data center consumption of electricity
      • opening a commercial viable option to re-use heat energy

3 of 4

Story: Max Planck Computing and Data Facility (MPCDF)

  • Max Planck Computing and Data Facility (MPCDF)
    • Is a cross-institutional competence centre of the Max Planck Society in Germany
    • supports computational and data sciences by running HPC systems in their own data center
    • updates HPC system to next generation technology at regular procurement intervals

  • Main goals that drove the decision for introducing direct-water-cooling at MPCDF
    • Reducing of the overall operational energy cost of the system
    • Minimizing/Eliminating the heat energy going to air in the data center

  • “Raven” – a direct-water-cooled HPC system at MPCDF (installed 2021)
    • All compute nodes are direct-water-cooled
      • 1592 dual-socket compute nodes with Intel Xeon processors
      • 192 compute nodes with additional GPU acceleration (4 Nvidia A100 GPUs)
    • Two Coolant Distribution Units (CDUs) with two separate Tichelmann Loops
    • Rear Door Heat Exchangers (RDHX) on the racks remove remaining heat
    • 100% water cooled system

3

4 of 4

Roadblock and/or key learning from “Raven”

  • Operation of the Coolant Distribution Unit (CDU)
    • CDUs are part of the system managed by the admins, but the facility management specialists have the skills in working with water
    • Key learning: Early involvement of system administrators and facility management specialists is important to clarify organizational questions

  • Missing secondary cooling loop implementation skills on supplier side
    • Special considerations on the materials being used (or to be avoided – e.g not using hemp for sealing connections between pipes and valves)
    • Key learning: Tight oversight of supplier work even in small details is important to achieve the quality that is important to operate the system flawlessly

  • Limited communication during Covid pandemic lead to issues
    • Onsite meetings and inspections were not possible during the implementation phase, resulting in misunderstandings and issues
      • inlet/outlet water connections from secondary loop to CDU were confused initially, causing bad cooling behaviour
      • Primary side pumps were not pressure regulated initially, causing high pressure with the CDU primary side 2-way valve regulating
    • Key learning: onsite communication with all involved parties (suppliers, vendor, facility specialists) is very important for details

  • Overall, the project was a great success
    • Smaller challenges and issues were resolved quickly, after they have been identified
    • Direct-water-cooling works as expected in operation, main goals for going to direct-water-cooling were achieved
    • Next generation of the system will be direct-water-cooled as well

4