High-Density AI Rack Cooling & Power Planning

Planning cooling and power for high-density AI racks ensures your hardware runs reliably without overheating or tripping breakers. This guide walks you through calculating heat loads, sizing cooling systems, and balancing electrical capacity for racks exceeding 20 kW per cabinet.

What You’ll Need

• Facility floor plan with rack locations and ceiling height
• Server/GPU specifications: power draw (kW), airflow (CFM), and operating temperature range
• Electrical single-line diagram (SLD) and available utility capacity
• Cooling system specs: CRAC/CRAH capacity, chilled water supply/return temps, or direct-to-chip capability
• Thermal imaging camera or temperature sensors for validation
• Power meter (e.g., PDU with monitoring)

Step-by-Step

1. Calculate total IT load per rack. Sum the nameplate power of all devices, but use realistic operating power (typically 60-80% of nameplate). For AI GPUs, check the thermal design power (TDP) and typical workload power. Example: 8 GPUs at 700W each plus 2 CPUs at 350W = 6.3 kW per server; with 4 servers per rack = 25.2 kW.
2. Determine cooling capacity needed. Use the formula: Cooling load (kW) = IT load (kW) + overhead (lighting, people, wall heat gain). Add 10-15% safety margin. For 25.2 kW IT load, plan for ~28-30 kW cooling per rack.
3. Choose cooling architecture. For racks >10 kW, traditional room cooling is insufficient. Options: (a) In-row cooling (rear-door heat exchangers or in-row units) for 10-30 kW/rack; (b) Direct-to-chip liquid cooling for >30 kW/rack; (c) Immersion cooling for >50 kW/rack. Evaluate your facility’s chilled water availability and budget.
4. Calculate airflow requirements. Each kW of IT load requires approximately 160 CFM (cubic feet per minute) of airflow. For 25.2 kW, you need ~4,032 CFM per rack. Ensure your cooling units deliver that at the rack inlet.
5. Design hot/cold aisle containment. Place racks in alternating hot/cold aisles, seal cold aisles with doors and ceiling panels to prevent mixing. This improves cooling efficiency by 20-30% and allows higher inlet temperatures.
6. Set supply air temperature. ASHRAE recommends 18-27°C (64-80°F) for inlet air. For AI workloads, aim for 22°C (72°F) to balance efficiency and safety. Increase supply temp by 1°C can save 2-3% cooling energy.
7. Calculate electrical power per rack. Convert IT load to kVA: kVA = kW / power factor (typically 0.9-0.95). For 25.2 kW, at 0.9 PF = 28 kVA. Add 10% for PDU losses and cooling fans inside racks.
8. Verify branch circuit capacity. Each rack should have dedicated circuits. For 208V single-phase, a 60A circuit provides 12.5 kW; for 415V three-phase, a 60A circuit provides ~43 kW. Choose voltage and amperage to match load with 20% headroom.
9. Plan for redundancy. Decide on N+1 (one extra cooling unit and UPS) or 2N (dual everything) based on criticality. For AI training, N+1 is common; for inference, N may suffice.
10. Install power distribution. Use rack PDUs with monitoring. For high density, use 3-phase PDUs with phase balancing. Ensure UPS capacity covers the total load plus startup surge (usually 1.5x steady-state).
11. Validate cooling with CFD modeling. Use computational fluid dynamics software to simulate airflow and temperature distribution before installation. Adjust vented tiles or in-row unit placement to eliminate hotspots.
12. Commission and test. After installation, run a load test at 100% for 24 hours. Monitor inlet temperatures, power draw, and humidity. Adjust cooling setpoints and airflow as needed.

Common Mistakes

• Using nameplate power instead of actual. Overestimating leads to overbuilding; underestimating causes overheating. Use real measured data.
• Ignoring airflow direction. Racks must have front-to-back airflow; mixing front-to-front causes recirculation.
• Insufficient cooling redundancy. A single CRAC unit failure can shut down a whole row.
• Overloading circuits. Don’t exceed 80% of breaker rating for continuous loads.
• Not accounting for cooling unit power draw. Cooling equipment itself consumes electricity, which adds to the facility load.
• Skipping phase balancing. Unbalanced phases cause neutral currents and inefficiency.

Tips

• Use cold aisle containment to allow higher supply temps (up to 24°C) without risk.
• Consider liquid cooling for GPUs to reduce airflow needs and energy consumption.
• Install blanking panels in empty rack spaces to prevent hot air recirculation.
• Monitor inlet temps at multiple points; use sensors with alarms.
• Plan for future expansion: oversize cooling and power infrastructure by 20-30%.
• Use variable speed fans and pumps to match load and save energy.

Troubleshooting

• Hotspots in rack: Check for blocked airflow, missing blanking panels, or failed fans. Increase cooling unit fan speed or lower supply temp.
• Circuit breaker trips: Measure actual current; if near 80% of rating, redistribute loads to other phases or add a new circuit.
• Cooling unit short-cycling: If unit turns on/off frequently, reduce number of active units or adjust deadband.
• Inlet temp too high: Verify cold aisle containment is sealed; check for leaks and adjust airflow.
• Power factor low: Use power factor correction capacitors or active PFC power supplies.
• Humidity too high/low: Maintain 40-60% RH; use humidifier/dehumidifier if needed.

Frequently Asked Questions

What is the typical power density for AI racks?

AI training racks can range from 20-100 kW per rack, with GPU servers consuming 5-10 kW each. Inference racks are usually 10-20 kW.

Can I use air cooling for 40 kW racks?

Yes, with in-row cooling and containment, air cooling can handle up to ~30-40 kW per rack. Beyond that, liquid cooling is recommended.

How do I calculate BTU/h from kW?

Multiply kW by 3412. So 25 kW = 85,300 BTU/h. Use this for comparing cooling unit capacity.

What is the difference between CRAC and CRAH?

CRAC (Computer Room Air Conditioner) uses refrigerant and is good for smaller loads. CRAH (Computer Room Air Handler) uses chilled water and is more efficient for large loads.

How much floor space do I need per high-density rack?

Plan for at least 3 square meters (32 sq ft) per rack including service clearances and cooling units. For liquid cooling, add space for CDUs.

Is it better to use 208V or 415V power?

415V three-phase reduces current and copper losses, and is common in Europe. In North America, 208V is standard, but 415V is becoming popular for high density.

Related

• AI Reasoning Engine: How It Works & What It Can Do
• Steam Kit for Kids: What to Look For
• Batman Arkham City vs Arkham Knight Combat Story Open World