Load bank testing and facility commissioning
How a data hall’s power, cooling and controls are proven before the GPUs arrive: what a load bank is, the five commissioning levels, why the test load must look like the real IT, and what a steady test cannot show about AI training load.
Testing and validation proves the IT hardware: servers, racks and the cluster. This section proves the building that hardware will sit in, before any of it arrives. If the hall cannot deliver power and remove heat as designed, nothing tested later on the racks means much.
Two kinds of testing, two owners
| Facility commissioning (this section) | IT testing and validation | |
|---|---|---|
| What is proven | Generators, UPS (uninterruptible power supplies), switchgear, power distribution, chillers, air handlers, coolant distribution units (CDUs), and the building management and power monitoring controls | Servers, racks and the cluster, from parts (L1) to a full server (L10), a full rack (L11) and the cluster (L12) |
| When | Before the IT arrives; the gate is "the hall is ready for IT" | After the hardware is built or delivered |
| Who does it | The facility operator, an independent commissioning agent, and the electrical and mechanical contractors | The server maker or systems integrator, then the operator of the cluster |
| The load | Artificial: load banks | Real: the hardware running test workloads |
| How the stages are numbered | Commissioning Levels 1 to 5 | Integration levels L1 to L12 |
The two meet at the point where real racks replace load banks. Responsibility for anything that crosses that line, such as the coolant loop inside the rack, has to be agreed in writing.
Where load banks sit: the five commissioning levels
Data centre power and cooling plant is commonly commissioned in five levels: factory acceptance testing (Level 1), site acceptance testing (Level 2), pre-functional testing and start-up (Level 3), individual system or functional performance testing (Level 4) and integrated systems testing (Level 5). Load banks are used mainly at Levels 4 and 5, and to prove generators and UPS at the factory.
Sources and status
A commonly used approach. Implementations vary — treat as a pattern, not a specification.
- Loadbank testing and commissioning for new data centres (white paper) — Aggreko, 2020-04-01
Status: Review required
Factory acceptance testingEquipment maker, witnessed by the owner
Plant is tested before it ships. Generators and UPS are run on resistive and reactive load banks to prove full load at their rated power factor.
Site acceptance testingMaker or installing contractor
Delivered equipment is inspected for damage and checked against the specification before it is installed.
Pre-functional testing and start-upInstalling contractor and maker’s technician
Each device is checked for correct installation, wiring and torque, insulation-tested, energised and started up.
Individual system testingCommissioning agent
Each system (a UPS, a generator, a busbar run, a chiller, a CDU) is tested on its own through its full range, often under load bank load and with thermal imaging to find hot spots.
Integrated systems test (IST)Commissioning agent, with every trade and the operator
The whole facility runs as one at full load: utility failure, generator start and transfer, cooling plant failures and recovery, with every alarm and sequence watched.
Generators and UPS systems are tested with resistive and reactive load banks, at leading and lagging power factors, because a purely resistive load does not prove their behaviour at rated power factor or the synchronisation of paralleled sets.
Sources and status
A commonly used approach. Implementations vary — treat as a pattern, not a specification.
- Loadbank testing and commissioning for new data centres (white paper) — Aggreko, 2020-04-01
Status: Review required
What a load bank test proves
- Power capacity and redundancy. The UPS, generators and switchgear carry the design load, and keep carrying it when one unit is taken away.
- Heat removal. The cooling plant removes the heat the load produces, through air, liquid or both, and holds supply temperatures where the design says.
- Controls, sequencing and alarms. The building management system (BMS) and electrical power monitoring system (EPMS) see the right values, raise the right alarms, and start, stage and transfer plant in the right order.
- Behaviour across the range, and recovery. The facility performs at partial load as well as full, and settles back to stable after a step change or a failure.
Controls come first
Of the four, control response is the one the others depend on. If the controls do not stage plant correctly, or an alarm fires late or not at all, a good result on load distribution or recovery tells you very little: it may only reflect how the test was run that day. Alarms and sequences are proven first, then trusted as the instruments for everything else.
The integrated systems test proves power, cooling and controls together at full load, including how they respond to varying load, staged plant failures and utility problems, rather than each system on its own.
Sources and status
A commonly used approach. Implementations vary — treat as a pattern, not a specification.
- Loadbank testing and commissioning for new data centres (white paper) — Aggreko, 2020-04-01
Status: Review required
Make the test load look like the IT load
A load bank only proves the facility for loads that look like itself. The closer it is to the equipment that will occupy the space, the more the test is worth.
- The same connection. If the racks will plug into a particular outlet, such as a 63 A three-phase-plus-neutral-and-earth connector, the test load should connect the same way, so the busway taps, plugs and breakers the IT will use carry the load too.
- Heat where the racks will be. One large load bank outside the hall proves the plant. Rack-mountable loads in the rack positions also prove the airflow, containment and hot spots of the room itself.
- The same cooling path. Liquid-cooled racks need a liquid-cooled load bank on the same loop, at the same flow, temperature rise and pressure drop as the cold plates it stands in for.
- The same mix. Most AI halls are hybrid: most of the heat goes to liquid, the rest to air from power supplies, memory and network gear. Both paths are loaded together, in roughly the design proportion.
- A realistic profile. Partial-load steps and ramps, not only a long run at 100%. Most halls spend most of their life at partial load, and controls that are stable at full load can hunt below it.
Heat for the integrated systems test comes either from freestanding load banks sized to the hall or from rack-mountable loads placed where the servers will sit; the second tests airflow and containment at the point of use.
Sources and status
A commonly used approach. Implementations vary — treat as a pattern, not a specification.
- Loadbank testing and commissioning for new data centres (white paper) — Aggreko, 2020-04-01
Status: Review required
| Air-cooled load bank | Liquid-cooled load bank | |
|---|---|---|
| Heat goes to | The room air (or outdoors, for a generator test) | A coolant loop: chilled water, or the CDU’s secondary loop |
| Proves | Air handlers and air-side cooling capacity, airflow, containment and supply and return temperatures | Chiller and facility loop, CDU pumps and heat exchangers, rack manifolds, flow balance and control response |
| Measured | Airflow, supply and return air temperature, temperature rise, capacity | Flow, temperature rise (ΔT), pressure drop (ΔP), approach temperature, capacity |
| Watch for | Heat concentrated in the wrong place; a room that passes overall but has hot rows | Loop cleanliness: the test equipment must not contaminate a loop that will feed cold plates |
A hybrid hall needs both at once, coordinated, because they share upstream power, controls and heat rejection.
Liquid-cooled load banks put a controlled heat load into a closed coolant loop, so the facility chilled-water loop, each coolant distribution unit and each rack manifold can be proven for flow, temperature rise, pressure drop, pump redundancy and control response before any liquid-cooled server is connected.
Sources and status
A commonly used approach. Implementations vary — treat as a pattern, not a specification.
- Simplifying Data Center Readiness with Liquid-Cooled Load Banks: Commissioning Best Practices for CDU and Data Center Loop Validation — Avtron Power Solutions, 2026-04-01
Status: Review required
What changes for an AI hall
Large AI training jobs alternate between compute-heavy and communication-heavy phases, so their power draw swings sharply and in step across the whole job; the swings grow with job size, and the proposed mitigations sit in software, GPU firmware and rack-level energy storage.
Sources and status
Supported by reliable public documentation.
- Power Stabilization for AI Training Datacenters (arXiv:2508.14318) — arXiv; listed as a Microsoft Research publication, 2025-08-20
Status: Review required
A conventional hall’s load changes slowly and is spread across many unrelated workloads. A large training job makes thousands of GPUs move between computing and communicating at the same moment, so the whole hall’s demand can rise and fall together, repeatedly. A load bank held steady at 100% does not exercise that. It proves capacity, not response to a load that swings.
- Ask what profile the test will run. Programmable load banks can step and ramp. Whether they can follow the speed of a GPU swing depends on the equipment; ask, rather than assume.
- Some of the protection is not in the building. Power smoothing can sit in GPU firmware, in the training software or in energy storage in the rack. None of that exists during facility commissioning, so it can only be proven once real racks are running, at the rack and cluster levels of IT testing.
- Agree who owns the gap. The facility proves it can handle the swing it was designed for; the IT side proves its racks stay inside that envelope. If neither side has written down the envelope, neither test proves the combination.
What it does not prove
- That the servers, racks or network work. That is IT testing, after the hardware arrives.
- That the hall copes with fast, synchronised load swings, unless the test was designed to produce them.
- Long-term behaviour: coolant chemistry, filter loading, component wear. Commissioning is a snapshot at handover.
- That the operations team can run the plant. A passed test run by the commissioning agent is not a trained operator; the people who will run the site should be in the room.
Where it goes wrong
| What happens | Why | What it costs |
|---|---|---|
| The integrated test is shortened or dropped | It is the last stage and the schedule has run out | The first real utility failure or plant trip becomes the test, with customer workloads running |
| Each trade signs off its own system | Contracts are split by trade, and no one owns the integrated result | A hall that passes every test and overheats, or trips, under combined load |
| The test load does not look like the IT | One central load bank is cheaper and quicker to set up | Hot spots, overloaded taps or unbalanced liquid branches are found by the first racks |
| Alarms are tested last, or not at all | They are treated as a BMS configuration task, not a test | Incidents are not noticed until something fails; results from earlier tests cannot be trusted |
| Liquid loops tested with dirty equipment | Test gear shared between sites without a cleanliness check | Contaminated coolant reaching cold plates; flushing and re-testing before the racks can go in |
Illustrative patterns, not a survey. Pass and fail criteria for each test are set in the project’s commissioning specification.
Key Takeaways
- Load bank testing proves the building; IT testing proves the hardware. Different owners, different times, different level numbers.
- Load banks matter most at commissioning Level 4 (each system) and Level 5 (the integrated systems test).
- Prove controls, sequences and alarms first; every other measurement depends on them.
- Electrical load, heat removal, airflow and alarms are verified together, not as separate sign-offs.
- The test load should connect, heat and cool like the IT that will replace it, liquid and air together.
- A steady test at full load does not prove how an AI hall copes with synchronised training load swings.