Abstract

What is this page about? This page looks at the two opposite ends of the data center: (1) the AI computer chips that create the heat, and (2) the heavy building equipment outside that dumps that heat into the outdoor air.

The Big Takeaway: A single modern AI server cabinet (120–132 kW) uses 17 times more electricity and creates 17 times more heat than a traditional server cabinet (8–12 kW), with future cabinets heading toward 600 kW. Because air fans physically melt down past 15 kW, liquid cooling is now mandatory - creating massive demand for both inside-the-room liquid hardware and outside-the-room chillers, valves, and prefabricated power skids.

1. AI Silicon & Rack Thermal Density Progression

  • Historical Legacy Baseline: Average rack density in Singapore nearly quadrupled from 2.4 kW/rack (2011) to 8.4 kW/rack (2020), with typical legacy enterprise racks at 8 to 12 kW.

  • AI Retrofit Racks: 40 to 60 kW/rack (Direct-to-Chip cooling cuts energy consumption by ~23% in AirTrunk deployments).

  • Current Production AI Racks (Gigabyte / NVIDIA / AMD Benchmarks):

    • AMD Radeon AI PRO R9700S (32 GPUs, 4x 4U DLC): Up to 30 kW.

    • DLC APU Servers (8x 3U): Up to 50 kW.

    • High-Density CPU DLC (Intel Xeon 6700/6500 or AMD EPYC 9005/9004, 80-128 CPUs): 76 kW to 105 kW.

    • NVIDIA HGX H200 (64 GPUs, 8x 4U DLC): Up to 80 kW to 109 kW.

    • AMD Instinct MI355X (64 GPUs, 8x 4U DLC): Up to 110 kW to 120 kW.

    • NVIDIA GB200 NVL72 Flagship: Nominal 120 kW to 132 kW per rack (~17x legacy density).

    • NVIDIA HGX B300 (64 GPUs): Up to 120 kW to 140 kW (6x 1U 33 kW power shelves, 54V DC 1400A busbar, 9x 1U NVLink switches).

    • NVIDIA Vera CPU + Liquid-Cooled Busbar Rack: Up to 230 kW per rack (36x Vera CPUs, 4x 3U 110 kW power shelves, 1x liquid-cooled 54V DC 5000A+ busbar).

  • Future Silicon Roadmap (NVIDIA Rubin Ultra): Projected to hit 600 kW per rack, requiring complete redesign of existing B300-class cooling infrastructure.

  • Single-Phase Immersion Alternative: Submerges servers in dielectric fluid inside sealed tanks (e.g., Envicool Coolinside integrated cabinets rated for 10-80 kW IT load, 280-1,000 L liquid volume, supply water ≤ 35 degC, flow 72-965 L/min).

2. Facility Balance-of-Plant (Primary Chilled-Water & Power Loop)

  • Heat Rejection Plant: External air/water-cooled chillers (e.g., LG Chip-to-Chiller, Schneider Uniflair XCA MW-class chillers, Modine Airedale), cooling towers, and reclaimed-water storage.

  • Primary Loop Isolation Valves: KSB BOAX-S/SF butterfly valves allow maintenance shut-off without interrupting cooling circuit operations.

  • Power Management & Redundancy Topologies:

    • 80% Circuit Rule (Singapore Best Practice): A 32A single-phase circuit carries max 25.6A continuous load; capacity planning uses actual measured watts (or 0.85 power factor) and keeps base utilization >40%; monitored via 48-outlet Raritan PX3 PDUs over SNMP/Modbus TCP/REST API.

    • Redundancy Tiers: N+1 (edge/small facilities), A+B Feed (Tier III colocation), and 2N Full Duplication (Tier IV hyperscale/financial; held 61.63% share of Malaysia's AI DC market in 2024, while Tier III grows at 31.77% CAGR to 2030).

Built with LogoFlowershow