Power Planning for High-Density AI Racks

This guide walks you through planning power conversion and backup for high-density AI racks, which typically draw 30-100+ kW per rack. You’ll learn how to size your power infrastructure, choose between AC and DC distribution, and design a resilient backup system that keeps GPUs running through outages.

What You’ll Need

• Rack power consumption data (nameplate and measured)
• Facility electrical single-line diagram (if available)
• UPS or generator specifications
• Power distribution unit (PDU) specifications
• Battery backup capacity requirements
• Budget and physical space constraints

Step-by-Step

1. Determine your rack’s true power draw. Start with the nameplate ratings of all equipment in the rack, but also measure actual consumption over a week using a power meter or BMC logs. AI workloads can spike, so use the peak sustained load, not the average or the theoretical maximum.
2. Calculate total rack power and redundancy needs. Multiply the per-rack peak by the number of racks to get total IT load. Decide on redundancy: N (no backup), N+1 (one extra unit), or 2N (full duplicate). For AI training, 2N is common because a single failure can waste thousands of dollars in compute time.
3. Choose the voltage and phase. High-density racks often run on 208V or 415V three-phase. 415V is more efficient because it reduces current and copper losses. If your facility only has 120/208V, you may need a step-up transformer or a dedicated feed.
4. Select the power conversion path. For AC distribution, you’ll use a PDU that takes facility power and distributes it to rack outlets. For DC distribution, you can use a rectifier to convert AC to 48V DC and then feed DC-powered servers or use a DC-DC converter. DC distribution is more efficient (fewer conversions) but requires specialized equipment.
5. Size the UPS. The UPS must cover the total IT load plus overhead for battery charging and cooling. For a 100 kW rack, a 150 kVA UPS with a power factor of 0.9 gives 135 kW of usable power. Ensure the UPS has enough battery runtime to cover generator startup (usually 5-15 minutes) or the expected outage duration.
6. Plan the backup generator. If you need extended runtime, size a generator to handle the full facility load, including cooling. The generator must be able to accept the UPS’s non-linear load (harmonic distortion) and start under load. Include an automatic transfer switch (ATS) to switch between utility and generator.
7. Design the distribution to the rack. Use a rack PDU with sufficient capacity and the right plug type. For 415V, you’ll need IEC 60309 or similar. For 208V, use NEMA L6-30 or L21-30. Consider a switched PDU that lets you monitor and reboot outlets remotely.
8. Add battery backup at the rack level (optional). For short power blips, a rack-level battery backup unit (like a battery cabinet or a BESS) can ride through without switching to generator. This is especially useful for GPUs that don’t tolerate power interruptions.
9. Test the system. Simulate a power failure by turning off the utility feed and verify that the UPS takes over, the generator starts, and the transfer switch operates. Test at full load to ensure voltage and frequency stay within tolerance.
10. Document and monitor. Create a single-line diagram, label all breakers, and set up monitoring for power usage, temperature, and UPS status. Use a DCIM tool or simple scripts to alert you to anomalies.

Common Mistakes

• Underestimating peak load. Relying on nameplate ratings alone often overestimates, but relying on average power can underestimate. GPUs can draw 30% more during training spikes.
• Ignoring cooling power. High-density racks generate enormous heat. Cooling can consume 30-50% of the total facility power. Include cooling in your backup sizing.
• Mixing voltage phases incorrectly. Plugging a 208V device into a 240V outlet can damage equipment. Always check the PDU’s voltage and plug compatibility.
• Forgetting about inrush current. When power returns, equipment draws a surge. Ensure UPS and generator can handle inrush without tripping.
• Not planning for maintenance bypass. You need a way to power the rack while the UPS is serviced. Include a maintenance bypass switch.

Tips

• Use a power management tool that tracks per-outlet power consumption to identify underutilized racks.
• Consider lithium-ion batteries for UPS—they have a smaller footprint and longer life than VRLA.
• For AI workloads, a 415V feed is worth the extra cost because it reduces energy loss by about 2-3% compared to 208V.
• Plan for future expansion: leave spare capacity in the UPS and generator, and run extra conduits for additional circuits.
• Coordinate with your utility to ensure they can provide the required power, and consider demand response programs to reduce costs.

Troubleshooting

• UPS overloads during normal operation. Check if the load exceeds the UPS rating. If so, redistribute loads or upgrade the UPS. Also verify that the UPS’s power factor is not too low.
• Generator fails to start. Check fuel levels, battery condition, and the automatic transfer switch. Test the generator weekly under load.
• Voltage drops when GPUs ramp up. This could be due to undersized wiring or a poor connection. Measure voltage at the PDU and at the server inlet; if the drop exceeds 5%, consult an electrician.
• PDU trips a breaker. Verify that the PDU’s input current is within its rating. Some PDUs have a higher inrush current; consider a soft-start feature.
• Battery runtime is less than expected. Batteries degrade over time. Test them regularly and replace them according to the manufacturer’s schedule.

Frequently Asked Questions

Q: How many kW does a typical AI rack draw?
A: It varies widely, but high-density AI racks can draw 30-100 kW. A single NVIDIA DGX H100 system can draw up to 10 kW, and a rack can hold several of them.

Q: Should I use AC or DC distribution for AI racks?
A: DC distribution is more efficient because it eliminates multiple AC-DC conversions. However, it requires specialized equipment and is less common. For most facilities, AC distribution with a high-efficiency PSU is simpler and more cost-effective.

Q: How much battery runtime do I need?
A> That depends on your generator. If you have a generator, you only need enough runtime to cover the generator’s startup time, typically 5-15 minutes. Without a generator, you need to cover the expected outage duration, which could be hours.

Q: Can I use a standard 120V outlet for a high-density rack?
A> No. High-density racks require 208V or 415V three-phase power. 120V circuits are limited to 15-20 amps, which is far too little for 30+ kW loads.

Q: What is the difference between N+1 and 2N redundancy?
A> N+1 means you have one extra UPS or generator beyond the required capacity, so if one fails, the others can handle the load. 2N means you have two complete, independent systems, so any single failure is fully backed up. 2N is more reliable but costs twice as much.

Q: Do I need a generator for AI racks?
A> Not necessarily, but if your operations are critical, a generator is recommended. Without one, you’ll rely on battery backup, which is expensive for long outages. Many AI facilities use a generator to provide extended runtime.

Related

• Apple Copy Paste Photos? How It Works
• Haze Seas Unlocking Fishman Style & Sea Beast Armor V2
• FCC Apple Beeper Mini: Status & Legality