Executive Summary
As next-generation AI accelerators push thermal design power beyond 700W to 1000W per socket, traditional raised-floor air cooling encounters hard physical limits at approximately 15kW to 20kW per rack. Deploying clusters of 8-way GPU servers demands a holistic rethinking of thermodynamic management. This technical paper details the engineering design and real-world operational benchmarks of South IDC's hybrid Direct-to-Chip (DLC) water loops combined with active Rear-Door Heat Exchangers (RDHx) at our Tseung Kwan O (TKO) facility.
1. Heat Flux Density and Air Cooling Saturation
High-density compute racks experience localized hot-spots that cannot be dissipated solely through increased airflow without incurring prohibitive CRAH fan energy consumption. In an unassisted air-cooled environment, thermal throttling occurs under sustained matrix multiplication loads. By deploying cold-plate liquid cooling directly atop primary compute silicon, over 75% to 80% of heat is immediately captured into the primary water loop before entering room air.
2. Secondary Heat Rejection via Active Rear Doors
Power delivery components (VRMs), high-bandwidth memory (HBM3e), and optical transceivers continue to dissipate secondary thermal energy into the chassis. To capture this remaining 20% to 25% of heat, active chilled-water rear doors are installed. Server exhaust air passes through the high-efficiency coil, returning to the data hall at room-neutral temperatures (22C to 24C), maintaining a facility PUE below 1.18 even under 100% continuous compute load.
3. Condensation Mitigation and Dew Point Safeguards
In subtropical coastal climates such as Hong Kong, dew point fluctuations present catastrophic condensation risks if supply water temperatures fall below room dew points. South IDC Coolant Distribution Units (CDUs) modulate supply water temperature dynamically between 18C and 24C via proportional-integral-derivative (PID) control, coupled with continuous optical leak detection ribbon laid along every rack tier.