ARTICLE · 1311442
AI-Era Data Center Cooling: Floor, Plenum and Fan Wall as One System
AI-Era Data Center Cooling: Floor, Plenum and Fan Wall as One SystemAn AI rack rarely fails because the floor is too weak. It fails because the air path was bought as three separate things: a raised floor, a CRAH unit and a tile layout. At 5–8 kW per rack that separation is survivable. At 60–100 kW it is not, because the floor void, the perforated tiles and the fan wall are no longer separate products. They are one pressure system, and the smallest leak or the wrong tile pattern moves the operating point. This guide sets out LP’s calculation chain for AI-era air cooling: how much air a rack needs, how the plenum and tiles are sized, how an EC fan wall is selected, how PUE and annual energy are checked, and what has to be measured at handover. The numbers are transparent, so a mechanical engineer can challenge them rather than accept a catalogue claim. Key takeaways Start with heat load and ΔT, not fan count. Airflow is a calculation, not a product category. The raised floor void is ductwork. Plenum depth, perimeter sealing, tile open area and cable layout belong in the same drawing as the fan wall. Open area is not airflow. A 35% tile and a 55% tile pass different volumes at the same pressure, and the higher-open-area panel usually carries less point load. A fan wall is selected on a P-Q curve with N+1 redundancy, not on a single nominal airflow figure. PUE is a measured result. The design model is there to compare options and to set the commissioning targets. Why 60–100 kW racks change the floor brief The air path has to be designed for the rack density the room will actually run, not the density on the original shell-and-core drawing. The bands below are the practical turning points we see in project briefs. Rack density band What changes in the air path Up to 8 kW A conventional CRAC and a modest plenum can work; airflow is rarely the limiting item 8–15 kW Hot-aisle return and a controlled tile layout become worthwhile; leakage starts to matter 15–30 kW CRAH with a fan wall, cold-aisle containment and a pressure budget become the baseline 60–100 kW High airflow, N+1 fan walls, low-leakage plenum and per-tile balancing are required to hold PUE targets Above 100 kW Air cooling usually becomes a hybrid problem; direct-to-chip liquid may carry the base load and air handles the remainder The important number is not the peak rack. It is the combination of peak rack, part-load profile and diversity. A room designed only for the peak will run its fans far from the efficient point for most of the year. A room designed only for the average will overheat on the first high-density commissioning test. The design has to state both. The four cooling layers, and where the floor sits A data center cooling system is usually described in four layers. Each layer has a different fan or pump duty, and each one has to hand over a defined condition to the next. Layer What it does What the floor team needs from it L1 — Rack Server fans and rear-door fans move air through the IT equipment Rack inlet temperature and allowed front-to-rear ΔT L2 — CRAH / CRAC Removes room heat through a coil and a fan wall Supply airflow, external static pressure, supply temperature L3 — Outdoor condenser Rejects heat from a chilled-water or refrigerant loop Water/refrigerant temperatures and part-load profile L4 — Cooling tower / dry cooler Rejects heat to ambient air Approach temperature and seasonal performance LP works at the L1–L2 interface: the raised floor, the plenum, the airflow tiles, the cold-aisle containment and the grounding that ties the floor system together. That is also where most of the avoidable losses sit, because a fan wall cannot fix a plenum that leaks or a tile layout that short-circuits supply air into the hot aisle. Start with airflow, not fan count The first calculation is the sensible-heat airflow for the rack or the room: Q = P ÷ (ρ × cp × ΔT) Where Q is airflow in m³/s, P is heat load in kW, ρ is air density (about 1.2 kg/m³ at sea level and 20 °C), cp is specific heat (about 1.005 kJ/kg·K) and ΔT is the cold-to-hot aisle temperature rise in K. For a 10 kW rack with a 12 K rise: Q = 10 ÷ (1.2 × 1.005 × 12) = 0.69 m³/s ≈ 2,490 m³/h ≈ 1,460 CFM The same calculation across the density bands makes the scale obvious: Rack load ΔT = 10 K ΔT = 12 K ΔT = 15 K 10 kW 2,990 m³/h 2,490 m³/h 1,990 m³/h 30 kW 8,960 m³/h 7,460 m³/h 5,970 m³/h 60 kW 17,910 m³/h 14,930 m³/h 11,940 m³/h 100 kW 29,850 m³/h 24,880 m³/h 19,900 m³/h Two engineering notes follow. First, a larger ΔT reduces airflow, but it also raises the rack inlet temperature and narrows the margin to the IT equipment limit. Second, the room total is not the sum of nameplate rack loads. Diversity, future growth and the actual IT schedule have to be agreed before the fan wall is frozen. Once the design airflow is known, add a margin. Ten to twenty percent is common for a first pass, but the margin should be a stated decision: future rack density, filter loading, duct leakage or commissioning uncertainty. An unexplained 30% margin is usually paid for in fan power for the life of the building. The pressure budget: the plenum is ductwork Airflow alone does not select a fan. The fan has to produce that airflow against the total pressure of the path: the pressure needed to push air through the perforated tiles into the room; the loss through the underfloor plenum, including cable trays and abrupt changes of direction; the coil and filter pressure drop inside the CRAH; the damper, grille and duct losses on the return or outside-air side; a fouling and filter-loading allowance. That total is the fan’s external static pressure at design airflow. If it is guessed, the fan will either run too fast and make noise, or run too slow and miss the airflow. Both outcomes are expensive to fix after the floor is closed. Tile open area is part of this budget, but it is not a substitute for it. Open area is the share of the panel face that air can pass through. A higher percentage passes more air at the same pressure, but it also removes load-bearing steel from the panel. Open area Typical use What to check 25% Low-density rooms, perimeter zones, low noise Enough airflow at the available plenum pressure 35% General data center cold aisle Balanced airflow and panel load rating 55% High-density cold aisle or local hot spot Panel point load, plenum pressure, risk of over-supply The practical rule is to use the tile supplier’s airflow-versus-pressure curve at the design plenum pressure, not to scale airflow linearly from the open-area percentage. Two tiles with the same open area can have different slot geometry and different pressure loss. Fan wall selection: duty point, N+1 and fan laws An EC plug fan removes the scroll and lets a number of fans sit in a common wall. That gives three things a single large fan cannot: redundancy, a wide speed range and a shorter air path. It also makes the array layout part of the performance, because fans interact when they run in parallel. The selection sequence is: Fix the design airflow per CRAH and the external static pressure. Plot the duty point on each candidate fan’s P-Q curve and keep it inside the efficient band, not on the edge. Choose the number of fans so that one fan can fail and the rest can still deliver the design airflow at a defined maximum speed. Check sound power at the actual duty point, not at the fan’s nominal best point. Check the motor and drive margin at maximum ambient, not at 20 °C. Agree the control sequence: lead-lag, automatic speed rise, alarm and BMS points. The fan laws are the sanity check for every speed change: Q ∝ n | ΔP ∝ n² | shaft power ∝ n³ If speed is reduced to 80%, airflow falls to 80%, pressure falls to 64% and shaft power falls to about 51%. That is why part-load control matters more than a small difference in nominal efficiency. It is also why a fan wall that is oversized and then run at low speed can still be a good solution, provided the controls and the minimum speed are set correctly. PUE and the operating-cost calculation PUE is total facility power divided by IT power. It is a ratio, so it can be improved by reducing cooling power, by reducing losses elsewhere, or by changing the IT load. For a cooling decision, the useful comparison is the overhead power at the same IT load. PUE = facility kW ÷ IT kW Annual energy = fan shaft power × operating hours × load factor ÷ fan efficiency The illustrative example below uses a 1 MW IT load, 8,760 operating hours and US$0.12/kWh. It compares an overhead of 420 kW (PUE 1.42) with an overhead of 280 kW (PUE 1.28). The difference is 140 kW, or about 1,226,400 kWh and US$147,000 a year before any tariff structure or demand charge. Three cautions matter. First, PUE is usually measured at part load, so the annual number needs a load profile, not a single design point. Second, a fan wall is only one part of the overhead; coil, chiller and heat-rejection efficiency also move the result. Third, a design model is not a commissioning result. The model selects the equipment; the measured data proves it. Commissioning: four numbers that prove the design A cooling system is not commissioned by walking the room and saying it feels cold. Four measurements should be recorded and compared with the design: Measurement What good looks like Instrument Airflow at each tile Within the balancing tolerance of the design value, with no dead spots Hood or hot-wire anemometer Plenum-to-room pressure At the design value, positive across the floor, no tile sucking air back Differential pressure gauge Rack inlet temperature Inside the agreed ASHRAE class range, with no rack above the limit Calibrated temperature logger Fan speed and power At or below design power at design airflow; no fan pinned at maximum speed BMS trend and power meter Add a leakage check before the floor is closed: perimeter joints, cable penetrations, pipe openings and wall/ceiling junctions. A plenum leak is not a small energy issue; it is a loss of control over which tiles get air. Where LP fits LP’s scope is the floor-to-fan interface: the raised floor and pedestal grid, the sealed plenum, the airflow tiles, the cold-aisle containment, the grounding, and the selection calculation that ties them to the fan wall. Fans, coils and controls are specified to the project duty point and supplied through qualified manufacturing partners. LP owns the interface and the commissioning numbers, because that is where most cooling problems actually start. That scope is deliberately narrow enough to be accountable. A floor supplier that only sells panels cannot own the air path. A fan supplier that only sells fans cannot fix a leaking plenum. The interface is the product. FAQ Can any raised floor be used as a supply plenum? Mechanically, the floor has to be reasonably airtight, clean and sealed at the perimeter and penetrations. Practically, the deciding factors are the understructure, panel seating, edge detail and the availability of airflow panels on the same grid as the solid panels. A floor that rocks or leaks is a poor duct. How much airflow does a 60 kW rack need? At a 12 K cold-to-hot aisle rise, about 14,900 m³/h (8,800 CFM). At 15 K it is about 11,900 m³/h. The number is a starting point; the design has to use the actual ΔT, diversity and future rack schedule. Which tile open area should we use? Use 35% as a general starting point, 25% for low-density or noise-sensitive zones, and 55% only where the pressure budget and panel load rating support it. Select from the tile’s airflow-versus-pressure curve at the design plenum pressure. Do EC fans always save energy? No. They save energy when they are selected near their efficient band, controlled at part load and not oversized to a duty point that never occurs. A bad selection with an EC motor is still a bad selection. What redundancy should a fan wall have? N+1 is the common starting point for a CRAH fan wall. The project Tier target and the operator’s maintenance philosophy decide whether more is needed. The important test is whether the remaining fans can hold the design airflow at a defined maximum speed with one fan out. What information does LP need for a first-pass cooling layout? The rack schedule (kW per rack and future growth), the target ΔT, the room plan and section, the available plenum depth, the CRAH capacity and external static pressure, the voltage and control protocol, and the noise limit. With those, LP can return an airflow, plenum and tile layout alongside the floor quotation. Next step Send the rack schedule and a room section to LP. We will return a first-pass airflow, plenum and tile layout, and a floor specification that matches the fan wall rather than fighting it. Talk to LP about a data center cooling layout References: ASHRAE TC9.9 datacom series; ISO 5801 and AMCA 210 fan performance; EN 50600; GB 50174; SJ/T 10796. Project values must be confirmed against the design brief and the supplier’s published curves. Keywords: raised access floor, access floor panels, why 60–100 kw racks change the floor brief, the four cooling layers, and where the floor sits, start with airflow, not fan count, the pressure budget: the plenum is ductworkFull guide: https://lpfloorgroup.com/ai-data-center-cooling-floor-plenum-fan-wall/
