Data Centre Rack Power and Cooling Capacity Planning

Adding one more server is trivial right up to the moment it is not. The rack that was comfortably running at 4 kW cannot take a GPU chassis that draws 8 kW on its own, and the UPS that was sized for a 40% load factor suddently runs at 85% — where efficiency drops and a single feed failure takes out the whole row. This article covers the practical capacity arithmetic: rack power tiers, the A/B feed model, BTU conversion, PUE, and the measurement discipline that turns a facility decision into a spreadsheet rather than an argument.

The Power Density Tiers You Are Actually Designing For

30 kW per cabinet is the widely used planning threshold: below it, airflow engineering solves the problem; above it, you are choosing a cooling technology, not tuning one.

What a Rack PDU Can Actually Deliver

# Single-phase, 208 V, 30 A PDU
#   usable = 208 x 30 x 0.8 (derate to 80%) = ~4.99 kW
# Three-phase, 208 V, 60 A PDU (3 x 208 x 60 x 0.8) = ~29.9 kW
# Three-phase, 415 V, 100 A PDU (~1.73 x 415 x 100 x 0.8) = ~57.5 kW

The 80% derating is not optional — it is what keeps breakers and connectors inside continuous-duty ratings and gives you headroom for inrush. Anyone quoting a PDU's nameplate figure as usable capacity is quoting a number that will trip a breaker at the worst possible time. Also confirm the branch circuit protection upstream: a 30 kW rack needs both the PDU and the upstream panel to support it.

The A/B Feed Model

# Per rack: place half the load on feed A and half on feed B.
# Each feed must be able to carry the FULL rack load, because a feed failure
# transfers everything to the survivor.
# Example: 40 racks x 5 kW = 200 kW total IT load
# Each UPS line must therefore support 200 kW
# With a 90% maximum occupancy: 200 / 0.9 = 222 kW per line

This is where capacity planning stops being arithmetic and becomes a failover exercise. Model the failure explicitly: if the A feed drops, the B feed now carries 100% of the row, and the closest upstream device becomes the constraint. Doing this on paper costs an hour; discovering it during a transfer test costs an outage. Record the redundancy topology (N, N+1, 2N) next to every number you publish, because "we have 200 kW of UPS" means three different things depending on redundancy.

Cooling: Every Watt Becomes Heat

# Convert IT power to cooling load
# 1 kW = 3412 BTU/hr
# 5 kW rack  = 17,060 BTU/hr
# 20 kW rack = 68,240 BTU/hr
# 30 kW rack = 102,360 BTU/hr

# Airflow check (approximate): 1 kW needs ~160 cfm at a 20 F delta-T
# 30 kW rack => ~4800 cfm, which no 1U-mounted fan arrangement will deliver

Three efficiency metrics belong in the design document:

• PUE = total facility energy / IT equipment energy. An average facility sits around 1.6; well-run ones reach 1.1–1.2. Publish it with the measurement boundary stated, because PUE computed from UPS output is not comparable to PUE computed from the utility meter.
• Cooling efficiency in kW/ton: 0.8 kW/ton is good practice, 0.6 kW/ton is a strong design.
• Airflow efficiency in W/cfm: 1.25 W/cfm standard, 0.5 W/cfm better.

Raise supply air temperature rather than lowering it. ASHRAE's allowable range extends to 90 °F at the intake, and running as close to 80 °F as the equipment permits reduces compressor energy substantially for the same reliability level.

Hot Aisle / Cold Aisle Is Not Optional Above 5 kW

At low densities, mixed air is a minor inefficiency. At high densities it is the failure mode: rack exhaust recirculating into the intake raises component temperatures until CPUs throttle. Practical measures, in order of cost effectiveness:

1. Seal cable openings and blank unused U positions — the cheapest airflow fix available.
2. Rearrange to hot aisle / cold aisle with correct inlet-to-inlet orientation.
3. Add aisle containment (chimney doors, then full end-of-row doors).
4. Move to rear-door heat exchangers or in-row cooling.
5. Direct-to-chip liquid cooling for 30 kW+ racks.

Capacity Data to Collect Before Every Deployment

• Rack-level: measured draw at the PDU (not nameplate), free U, weight against floor loading, feed A and B headroom.
• Row-level: upstream panel capacity, breaker ratings, containment type.
• Room-level: UPS load factor, battery runtime at the new load, cooling tonnage available and utilised, PUE trend.
• Growth: what the same rack looks like after the next generation of hardware — the trend on a 5 kW rack is measured historically in lower numbers, and in K, double-digit kW going forward.

Monitor, do not assume: SNMP-enabled PDUs and UPS management cards give you per-outlet current so the next capacity review is data-driven. See our posts on APC UPS network management cards, out-of-band console servers and structured cabling and labelling practice.

原文链接：https://www.energy.gov/sites/default/files/2024-07/best-practice-guide-data-center-design_0.pdf