Very Early Warning Smoke Detection in a Data Center: What the Codes Require Where

A data hall is the hardest room in a commercial building to detect a fire in, and it is hard for a reason that has nothing to do with the fire. The room recirculates the same air rather than changing it, and it recirculates it fast. A white space running at 40 to 100 air changes per hour is moving more air through the rack intakes than a hospital operating theatre or a laboratory fume hood bench, and it moves that air in a closed loop. Any smoke that forms is diluted into the full recirculating volume within seconds and carried wherever the return path leads, which is usually not where the smoke was born.

Against that, the ordinary spot-type photoelectric smoke detector is listed to respond somewhere in the range of 2 to 4 percent obscuration per foot, and its spacing tables assume a room with still air and a smooth ceiling [10][15]. An incipient fault in a cable termination or a power supply capacitor produces particulate at concentrations one to three orders of magnitude below that for the first several minutes. So the spot head waits, and by the time it announces the event the room has been breathing combustion products long enough for them to have travelled through the cooling air path and into equipment that was never near the fire.

Very early warning fire detection — VEWFD in the code language, and air sampling or aspirating smoke detection (ASD, the category the VESDA product family belongs to) in the equipment catalogue — exists to close that gap. The codes do not require a brand. They require a set of numbers, and those numbers are a specification: coverage per sampling port, sensitivity at that port, transport time from that port to the detector, and the locations where ports have to be. Get those four right and the system does what it was bought for. Get any of them wrong and you have purchased an expensive way to detect a fire late.

This is an installation guide, not a product comparison. It works through the standard stack, the three detection classes and the numbers that separate them, the five places in a data center where ports are required, the pipe network arithmetic that decides whether the system meets its transport limit, and the commissioning and maintenance steps that determine whether it still works in year three. It closes on the two things that changed most recently: containment and liquid cooling in the room, and batteries in the corner. The companion piece on clean agent suppression selection covers the agent side; this one covers the sensing layer that has to fire before the agent is allowed to.

The standard stack, and which document owns which decision

There is no single code for very early warning detection in a US data center. There is a stack of five documents, plus a product listing standard, plus whatever the authority having jurisdiction has adopted and whichever edition it has adopted. Most detection design failures trace back to reading only the top of the stack.

Table 1 — The detection stack. Note that NFPA 76, not NFPA 75 or 72, is where the three detection classes are actually defined, and that the insurer data sheet is often the document that forces VEWFD in a room that the code would let you protect with spot detectors.

A facility that combines IT equipment with carrier or telecommunications gear falls under both NFPA 75 and NFPA 76 and has to satisfy the stricter of the two for each condition. A modular or containerised unit brings NFPA 75 Chapter 13 into scope, which has its own detection provisions for factory-built enclosures — a topic covered in more depth alongside modular and containerised build-outs. And where the facility is pursuing or holding a Tier certification, the detection redundancy expectations are assessed as part of that programme as well, in the same way the Tier standard comparison treats the mechanical and electrical layers.

Three classes of detection, and the four numbers that separate them

NFPA 76 is the document that sorts detection into classes, and the class you are designing to determines every number downstream of it. There are exactly four quantities that matter, and all four move together.

Table 2 — The three NFPA 76 detection classes as they apply to air sampling systems [2][8][9]. Where two levels of sampling are provided — one high, one low — each level is permitted 400 ft2 per point, with the combined coverage of the two levels still not exceeding 200 ft2 [2][11].

Two consequences follow immediately, and they are the reason a half-designed aspirating system underperforms.

First, the coverage figures convert into a spacing grid, and the grid is finer than most designers expect. At 200 ft2 per point the implied spacing between adjacent ports is about 14 ft (4.3 m), with the end port no more than 7 ft from a wall; at 400 ft2 it is 20 ft and 10 ft; at 900 ft2 it is 30 ft and 15 ft, which is exactly the conventional spot-detector grid that NFPA 72 uses as its baseline [9]. The rule of thumb to carry: an aspirating system that is claiming VEWFD performance while the ports are laid out on the 30 ft spot-detector grid is not a VEWFD system, whatever the detector's brochure says.

Second, the coverage and the transport time are in tension with each other, because you cannot add ports for coverage without also adding pipe, and pipe is what transport time is made of. A 1,000 m2 hall at 200 ft2 per port needs 54 sampling points. That is one detector's worth of ports — a four-pipe detector with 25 holes per pipe is rated for 100 — but not necessarily one detector's worth of pipe, because the same detector carries total-run and single-run length limits that bind well before the hole count does. Working through that arithmetic is the difference between a design that passes commissioning and one that gets value-engineered during construction. The arithmetic is in the pipe network section below.

Where the ports go: the five required locations

NFPA 75 does not leave placement to judgement. It names the locations, and the list is longer than most of the rooms it describes actually have detection in.

NFPA 75 requires automatic detection in an IT equipment area at four locations: at ceiling level throughout the ITE area; below the raised floor of the ITE area containing cables; in the exhaust or return air stream where aisle containment systems are used; and in the return air stream where the above-ceiling space is used as a return air plenum [1][12]. To that list NFPA 75 adds a fifth condition for the equipment itself: where detection is used to monitor fire in IT equipment with close-coupled cooling units, detectors or sampling ports must be provided at the return inlets [13].

Table 3 — The NFPA 75 detection locations, with the hazard each one exists to catch. Table 3 continues onto the equipment itself where in-cabinet sampling is used.

Inside equipment, NFPA 75's guidance is to put the sampling point where smoke is most likely to migrate rather than where it is convenient to drill. In an unventilated, effectively sealed cabinet that means the top 10 percent of the cabinet interior; in a ventilated cabinet it means at the point where the ventilation exits the cabinet; in a naturally vented cabinet it means the upper ventilation vent [13]. Where the equipment is protected by in-rack or in-cabinet sampling, FM Global goes further in the opposite direction from the room-level design: for the fastest possible response, sampling points are to be installed inside the electrical and electronic cabinets, not merely above them [7].

For rooms where the detection exists partly to control the building, NFPA 75 accepts one of three approaches to the plenum and damper problem: detection throughout the above-ceiling or below-floor air circulation space; detection located at each smoke damper; or detection at the other openings through which smoke could enter the IT space, such as a door from a non-ITE area [14]. Sampling at each damper is the cheapest version of this, because one aspirating network can watch a dozen dampers that would otherwise need a dozen spot heads.

The return-air port: the highest-value location in the room

If one location in a data hall is worth more than the others, it is the return air path. Every cubic metre of air in the room passes through it, so the concentration of any contaminant reaching it is the room-average concentration rather than a local sample. That makes it the most reliable place to put a port and, in a contained hall, the only place where the concentration has not yet been diluted by the full recirculating volume.

The rules for it are specific. NFPA 76 requires sensors or ports mounted at the return air location, installed where return air is brought back to the unit for stand-alone packaged HVAC units, and sized so that each sensor or port covers no more than 0.4 m2 (4 ft2) of the return air opening [2][18]. FM Global states the same 4 ft2 limit and adds a placement rule that is easy to get wrong: do not locate the ports within 3 ft (0.9 m) of a supply duct register [6].

The 4 ft2 limit is far tighter than the 200 ft2 room coverage figure, and that difference is the point — a return air opening is a small, high-velocity area, so a port covering 200 ft2 of it would be sampling a jet of air that has already passed the port before it had time to be sampled. Where the opening is small, the permitted coverage drops to 0.2 m2 (2 ft2) per port, which is a reduction, not a relaxation [18].

A worked example makes the density real. A vertical return air opening 1.5 m by 0.8 m is 1.2 m2 of face area. At the tightened 0.2 m2 per port that is 1.2 / 0.2 = 6 sampling points on that one opening [18]. Six ports, on one vent, on one pipe run — and that is before the room's ceiling, underfloor and contained-aisle layers are counted.

Two more return-air details belong in the specification. Where detection is intended to catch a failure inside the cooling unit rather than in the room, the sampling point goes downstream of the filters on the supply side of the CRAC or CRAH, so that a smouldering filter or fan-belt failure is recognised before the unit distributes the products of combustion through the equipment it is cooling [14]. And where the cooling system has an airside economiser, the return path is also the entry path for smoke from outside the building — a nearby fire or a diesel exhaust plume can arrive through that opening [14]. Detection at the return air handles both, which is why it earns its cost twice over.

High airflow is a spacing problem before it is a sensitivity problem

NFPA 72 has an explicit provision for rooms that move their air quickly, and it is one of the most commonly missed requirements in data center detection design. Where the airflow in a defined space exceeds 8 minutes per air change — 7.5 air changes per hour — spot detector spacing must be reduced in accordance with the air movement table before any other spacing adjustment is applied [3][15].

Table 4 — NFPA 72 Table 17.7.6.3.3.2, smoke detector spacing based on air movement, applied to a reference 1,000 m2 (10,764 ft2) hall. The table's own note is that it is not to be used for underfloor or above-ceiling spaces [3][15].

Three things about that table are worth stating plainly.

The scaling is close to inversely proportional to airflow. Going from 7.5 to 60 air changes per hour is an 8-fold increase in air volume moved; the permitted coverage falls by a factor of 7.2 [15]. Spacing reduction is not a penalty the code applies to be difficult — it is the code's acknowledgement that dilution is the mechanism by which smoke detectors miss fires, and that the fix is more sampling points.

Applying it to a real hall shows why aspirating detection wins on arithmetic alone. A 1,000 m2 white space at 60 air changes per hour requires 87 spot detectors, each of which is an addressable device to wire, address, test annually and eventually replace. An aspirating system covering the same area at the VEWFD class needs 54 ports on a handful of pipe runs feeding one or two detectors [2]. The hardware price difference is negligible; the difference is 87 devices to maintain against one. Set that against the cost of the event the detection exists to prevent: published downtime reference figures put a data center at around USD 300,000 per hour of operation, against USD 40,000 for a clean room and USD 1,000,000 for a large electronic data processing installation [22]. A layer that moves the response from the five-minute mark to the one-minute mark is not really a capital expense decision at that scale.

And then there is the gap. The table's top entry is 60 air changes per hour, and AI-era halls routinely run past it — high-density GPU rows push air volumes that the 2013-era table does not address at all [15]. NFPA 72's own note resolves the problem for aspirating designers: the air movement table does not apply when air sampling or projected beam detectors are used, because those systems are installed in accordance with the manufacturer's instructions instead [15]. That is why an aspirating design is not merely a nicer way to do what spot detectors do at high airflow — it is the only detection method in the code that is not scaled by a table which stops before the rooms we are now building.

NFPA 75's Annex A adds a second, independent set of limits that a designer should treat as a design default rather than a footnote, because it pairs each functional objective with both a sensitivity and a coverage area.

Table 5 — NFPA 75 Annex A guidance pairing each detection function with a sensitivity threshold and a coverage limit [1]. Above 30 air changes per hour, both the threshold and the permitted area are halved. This is the table to quote when a client asks why the alert threshold is 0.2 rather than 1.0 percent obscuration per foot.

Table 5 also explains a design subtlety that gets lost when a single detector is asked to do two jobs. The objective matters. Where detection exists primarily to summon people to a small event, the system should be arranged with high sensitivity and close spacing. Where detection exists to release an agent, the guidance reverses: lower sensitivity, spacing tighter than the listed spacing, and a form of logical confirmation so that a single indication does not release the room [1]. A detector tuned aggressively for the alert function will cry wolf into a releasing circuit; a detector tuned only for release reliability will miss the incipient event it was installed for. Well-designed halls run the two functions from the same pipe network but different thresholds, with the release logic gated behind cross-zoning.

The pipe network is the design

Everything above describes where the sample has to be taken. This section is about whether the sample arrives in time, which is a hydraulic question about pipe, and it is the part of the design that gets value-engineered after award.

The transport time is the interval between smoke entering a sampling point and the detector announcing it, and it is measured at the most remote or least sensitive port — not at the detector, and not at the nearest hole [9][19]. It is a calculated quantity, not a field discovery: transport time and system performance are required to be worked out in the design software and then confirmed during commissioning, which is why the model and the as-built network have to be reconciled rather than assumed to match [19]. The code limit depends on the class being claimed: 60 seconds for very early warning, 90 for early warning, 120 for standard, and 60 seconds flat for a data center VEWFD system under the insurer's criteria as well [2][6][9].

Transport time is not a property of the detector. It is the pipe volume divided by the airflow the aspirator pulls, and both of those are design choices. Running the arithmetic on the hardware limits that aspirating detectors are actually built to shows how quickly the 60-second limit becomes the binding constraint.

Read that table as a warning. A single 100 metre run of the largest pipe the system accepts uses 87 percent of the very early warning transport budget before a single fitting, elbow or partially loaded filter is accounted for [16]. The practical implication is that in a data hall the design that meets the class is not one detector with four long runs — it is more detectors, each with shorter runs, positioned so that no run approaches the length limit. Detector count in a well-designed VEWFD installation is driven by transport time, not by coverage.

The rest of the pipe network rules follow the same logic, and each of them exists because a field failure taught someone a lesson.

• Model it, then build it to the model. The pipe network is designed in the manufacturer's listed modelling tool, which solves the hole sizes needed to balance flow across the network and reports the transport time at the least favourable point. The as-built pipe has to match the model. A network that was modelled and then re-routed on site is an unproven network, and the commissioning test is what proves it [16].
• Size the holes progressively. Sampling holes are drilled smaller near the detector and larger toward the far end, because the near holes sit in a region of higher static pressure and would otherwise take more than their share of the airflow, starving the far end of the run. Uniform holes are the most common field error in an otherwise correct design [16].
• Keep the bore and the runs inside the limits. Sampling pipe is typically 25 mm outside diameter with a 15 to 21 mm bore, run in ABS or PVC, with a maximum total length around 200 m, a maximum single run around 100 m, and no more than about 25 holes on any one run [16].
• Use swept bends, not square corners. Right-angle elbows create pressure loss that the model did not account for unless they were modelled. Two 45-degree fittings or a swept bend are the standard substitutions.
• Cap the ends, and drill the cap. Every run terminates in an end cap, and the cap carries a small calibrated hole that is part of the balancing — it is not a plug and it is not to be sealed with tape [16].
• Return the exhaust to the room being sampled. The exhaust has to be piped back into the same space the network samples, because the whole transport calculation assumes the pressure inside the pipe and the pressure in the room are the same. Exhausting to a different space imposes a differential the aspirator was not sized for [8][20].
• Label the pipe. Sampling pipe must be conspicuously identified as a smoke detector sampling tube at every change of direction and branch, on each side of any wall or floor penetration, and at intervals no greater than 20 ft (6.1 m) so the warning is visible from anywhere in the space. The instruction is literally "Do Not Disturb" [15].
• Keep capillaries short. Where capillary tubes drop through ceiling tiles to sample the space below a false ceiling, they must be short and must land on the grid intersections that define the sampling layout — a capillary long enough to be convenient to install is long enough to add unmodelled transport time [18].
• Space vertical runs by temperature, not distance. Where a sampling run goes vertically, the interval between holes follows the smaller of a 2 degree C temperature increment or 3 m, and no hole faces upstream against the airflow [16].

Alarm levels, cross-zoning, and what the detector is allowed to do

An aspirating detector is not a yes-or-no device, and treating it as one is a common specification error. A typical unit reports four alarm levels and two fault levels: Alert, Action, Fire 1 and Fire 2, plus Maintenance and Major Fault [16]. Those levels exist to be wired to different actions, and the value of an aspirating system lies largely in what happens in the long interval between the first of them and the last.

Table 6 — Alarm and fault levels on an aspirating detector, and the actions each is normally wired to. The exact number and naming of levels varies by product; the principle that each level drives a different output does not [16].

The release sequence for a data hall is where detection design meets suppression design, and the insurer's criteria are the clearest statement of it. FM Global requires that a clean agent system be actuated either by very early warning detection or by an approved cross-zoned smoke detection system. Where VEWFD is the actuator, an alert condition must be annunciated at a constantly attended location; at the pre-alarm setting the control system should automatically shut off air conditioning and close fire dampers; and at the alarm setting it should discharge the agent after a time delay not exceeding 30 seconds, with the alarm setting comparable to the alarm setting of conventional smoke detectors [6].

Underpinning that, NFPA 2001 requires a pre-discharge delay with an abort station so that a person who has correctly identified a false alarm can stop the discharge [4], and NFPA 72's cross-zoning provision requires two independent detection inputs within a defined window before the release initiates [4][12]. The reason all of this is worth designing carefully rather than accepting a default is the cost of being wrong in both directions. A release into a live hall involves agent replacement, a purge, and a restart procedure. A missed event involves the room. The alert and action levels on the aspirating detector are how the design gets the best of both: the smoke is caught two levels before anything discharges, which is the entire point of paying for the sensitivity.

The purge itself is arithmetic, and it is worth having on hand when the sequence of operations is written, because the re-entry time is often the constraint the business cares about rather than the fire. The decay of agent concentration in a ventilated room follows an exponential: the ventilation rate required to fall from the design concentration to a safe re-entry concentration in a target time is Q = -(V / t) * ln(C_safe / C_0), where V is room volume, t is the target purge time, and the concentration ratio is safe level over design level [21]. Purge targets of 10 to 30 minutes are typical, and for halogenated agents the ventilation rate is often driven by the need to clear decomposition byproducts rather than the agent itself [21].

Commissioning: what has to be proven before handover

A very early warning system that has not been proven on site is an assumption. The 2024 edition of NFPA 75 backs that statement with a new Annex F, drawn from optional test procedures that had been in NFPA 76, which sets out performance test procedures for early fire detection systems: a heated wire test and a lactose-potassium chlorate test, both of which produce a controlled, reproducible smoke challenge of the kind a real incipient fire produces [1].

Those performance tests sit on top of the ordinary commissioning sequence, and the sequence has a specific order for a reason.

1. Model first, then verify the model on site. Confirm the modelled transport time at the least favourable point is inside the class limit before anything is drilled, then verify the as-built network against the model [16].
2. Measure transport time at the most remote port. Introduce smoke at the farthest and least favourable sampling point with alarm delays set to zero seconds so that the delay does not contaminate the measurement, and confirm the measured transport time against both the model and the code limit. Reset the delays afterwards [16][17]. A transport time confirmed at the detector, or at the first hole, proves nothing about the network [9].
3. Run the smoke test where the room's air actually goes. Verification is about placement with respect to airflow, with particular attention to ventilation and stratification. Because smoke dilutes in a large room, the test is also the check that the spacing did not get stretched during construction [6].
4. Run the fire performance test. For a computer room the traditional challenge is a one-metre length of PVC-insulated wire energised for one minute, with the detector expected to respond within two minutes of the end of the burn; the modern equivalents in NFPA 75 Annex F are the heated wire and lactose-potassium chlorate tests [1][17].
5. Trip the airflow supervision, both ways. Block a sampling point, then open the network, and confirm the detector raises a fault in both cases. The airflow fault is the system supervising itself — it reports a blocked, broken or dust-loaded network, and finding it non-functional at commissioning means it will report nothing for years [9].
6. Drive every alarm level through the full cause-and-effect. Step the detector through Alert, Action, Fire 1 and Fire 2 and verify each output against the matrix, including damper closure, HVAC shutdown, notification and the release circuit [9][16]. A detector wired to a releasing panel and never exercised end-to-end is an untested life-safety chain.
7. Set the operating thresholds against the room's actual background. Thresholds should sit above the normal particulate level of the space and below a real event, and the values must be documented. Leaving factory or construction sensitivities in a live room produces either nuisance alarms or missing warning — usually the former during fit-out and the latter afterwards [9].
8. Fit a fresh filter at turnover. Construction dust is the single most common cause of drift in a new installation. The filter goes in after the dust settles, not before [9].

The NFPA 72 acceptance test that governs all of this sits in the initiating devices and inspection, testing and maintenance chapters, and it is worth noting that acceptance testing a high-sensitivity system is a different exercise from acceptance testing a spot detector. The tolerances are tighter, the pass criterion is a time rather than an alarm, and the record has to capture the model, the as-built pipe layout, the measured transport time, the thresholds, and the airflow baseline [9].

Maintenance: the ways a working system quietly stops working

Aspirating detection has a reputation for low maintenance, and it earns that reputation by replacing 87 devices with one. The failure modes simply move from the room to the pipe network, where they are less visible.

Table 7 — The maintenance failure modes that matter in a data hall, and the specific way each one degrades performance without producing a hard fault.

The testing intervals are set by code and by the insurer. NFPA 72 requires the sensitivity of smoke detectors to be tested after the first year in service and every two years thereafter, with the interval extendable to every five years if sensitivity has remained within the listed range on two successive tests [15]. FM Global requires the VEWFD system to be tested once every two years to verify proper operation, by a defined test method [6]. Filters are changed on condition and at turnover. Airflow baselines should be recorded at commissioning and trended, because a slow drift in the airflow reading is the earliest available warning that the network is loading up.

Containment and liquid cooling: what changed in the room

Two changes in the last several years have moved the detection design more than any code revision: aisle containment became standard, and liquid cooling arrived at rack scale. Both push in the same direction — away from the ceiling, and toward the return path and the equipment itself.

Containment creates a room within a room. A contained hot aisle with doors, end panels and an overhead baffle is a semi-enclosed volume whose air is drawn into a defined return path, and NFPA 75 responds to that directly by requiring detection in the exhaust or return air stream where aisle containment is used [1][12]. Sampling above a contained aisle at ceiling level is sampling air that has already been through the cooling unit. The same containment also interacts with the suppression system: NFPA 75 is explicit that hot or cold aisle containment shall not obstruct the free flow of a gaseous clean agent to the IT equipment or to the cooling system serving the contained aisle [13]. Detection and suppression therefore have to be designed against the same containment geometry, and a containment retrofit after a detection and suppression design is a change that invalidates both.

Liquid cooling changes the argument more subtly, and in a way that surprises people who assume that liquid means less air. In most deployments the liquid loop removes heat from the processor and nothing else. Memory, power supplies, network cards, drives and everything else on the board still reject their heat to the room's air, so the hall still moves large air volumes — and because a liquid-cooled rack concentrates far more value per square metre, the cost of a late detection is higher rather than lower. In a sealed or enclosed rack the smoke can be trapped inside the enclosure, where a ceiling-level port will not see it and a smouldering condition can persist well before any flame appears. The response in the codes is already in place: NFPA 75's in-cabinet guidance puts the port where the ventilation exits the cabinet, at the return inlets of close-coupled cooling, and inside the top of a sealed cabinet [13]; and for the fastest response the insurer's guidance is to sample inside the electrical and electronic cabinets themselves [7]. In a hall that has moved to liquid cooling, the layered strategy — in-rack or in-cabinet sampling for the equipment, return-air sampling for the aisle, and ceiling sampling for the general area — is the design that matches the physics.

Batteries, off-gas, and why the 2024 editions changed

The most consequential revision in this area is not about air sampling at all. The 2024 editions of both NFPA 75 and NFPA 76 removed their lithium-ion battery requirements and handed them to NFPA 855, and in exchange added off-gas detection requirements aimed at catching electrolyte vapour before thermal runaway — with an explicit note that conventional flammable gas sensors are not an acceptable substitute [1][5].

That split is worth understanding, because it means a battery room in a data center is no longer protected by the same detection philosophy as the hall next to it. A lithium-ion cell in thermal runaway supplies its own oxidiser. It does not need the room's air, it cannot be extinguished by reducing oxygen, and a clean agent can knock down the visible flame without arresting the reaction inside the cell, so the smoke detector and the agent together address the wrong stage of the event. The instrument that matches this hazard is off-gas detection at a low threshold, ahead of the moment the cell vents, because that is the only point at which the sequence can still be interrupted.

The battery room's own detection design therefore follows NFPA 855, which requires off-gas detection under the 2024 revisions and, in its 2026 edition, requires a hazard mitigation analysis for essentially all lithium-ion installations, moving away from deflagration venting as a primary protection strategy toward preventing the accumulation of a flammable mixture in the first place [5]. Smoke and very early warning smoke detection still have a role in a battery room — they are how the room is monitored when nothing is happening — but the layer that matters during an event is the gas layer, and mixing the two up in a specification is how a battery room ends up with excellent smoke coverage and no early warning of the hazard it actually has.

Where a UPS or battery room sits inside a larger detection scheme, the integration point is the monitoring layer. Detection outputs, alert and action levels, airflow supervision and fault history all belong in the same place as the rest of the facility's telemetry, so that an alert with a location and a trend line reaches a person who can act on it — the same argument that applies to what a compute-center DCIM has to cover. A very early warning system whose alerts terminate in an unmonitored panel in a corridor is a very expensive smoke detector.

A specification checklist

What follows is the short version of everything above, written as the items to put into a detection specification rather than as the reasoning behind them.

• Declare the class. State very early warning fire detection explicitly, and state the four numbers: 200 ft2 per sampling point, 0.2 percent obscuration per foot at alert, 1.0 percent at alarm, 60 seconds maximum transport time measured at the least favourable point. If the facility is insured to FM Global criteria, the 60-second limit applies regardless of which NFPA class is named.
• Name the locations. Ceiling level throughout the IT area, below the raised floor where cables are present, in the exhaust or return air stream inside each contained aisle, in the return air stream where the ceiling void is a return plenum, and at the return inlets of close-coupled cooling units. Add in-cabinet or in-rack sampling where equipment is sealed or where rack value justifies it.
• Require the airflow strategy to be documented before the layout is drawn. The port layout follows the air, and the air path is a mechanical design decision. Detection layout signed off before the airflow plan exists is the single most common cause of a system that detects nothing.
• Require the pipe network model, and require the as-built to match it. Name the modelling tool, require the transport time report at the least favourable point, and make the as-built survey a contractual deliverable rather than a site activity.
• Set the port density at the return openings separately. One port per 0.4 m2 of return opening, tightened to 0.2 m2 where the opening is small, and no port within 3 ft of a supply register.
• Specify the cause-and-effect matrix, level by level. Alert to the attended station and to BMS, with HVAC shutdown and damper closure; Action to operations escalation; Fire 1 and Fire 2 as the cross-zoned release pair with the pre-discharge delay and abort station defined; Maintenance and Major Fault wired to supervision, not to a silenced input.
• Specify the commissioning tests by name. Transport time at the most remote port with delays zeroed, the Annex F performance test, the airflow fault tripped in both directions, and a walk-through of every alarm level through the full matrix.
• Specify the turnover artefacts. Fresh filter fitted after construction dust settles, thresholds set against the measured room background, airflow baseline recorded, and a commissioning record that carries the model, the as-built layout, the measured transport time, the thresholds and the baseline.
• Specify the maintenance interval and the trending. Two-yearly operational test, sensitivity testing on the NFPA 72 cycle, filter change on condition, and airflow trend review as a standing item.

If you are specifying detection for a hall that is still in design and the containment, cooling and battery layouts are not yet fixed, that is the right time to settle the sequence — the detection and the suppression both depend on the room's geometry, and both become retrofits if the geometry moves first. Send us the project scope with the hall dimensions, the cooling strategy and the battery arrangement, and we will come back with the detection and suppression layers that match them.

Frequently asked questions

Is an aspirating smoke detection system required by code in a US data center?

Not by name, and not by any single standard. NFPA 75 requires automatic detection providing early warning in an IT equipment area and names the locations it has to cover, but it does not mandate air sampling as the technology [1]. In practice the requirement is created by two forces rather than by a clause. The first is physics: at the air change rates a modern data hall runs, the NFPA 72 spacing table reduces spot detector coverage so far that an aspirating system becomes the cheaper way to achieve the same coverage [15]. The second is insurance: FM Global's data center criteria require very early warning fire detection in telecommunication equipment areas and return air systems, and set installation criteria that a spot detector cannot meet [6]. A specific brand is never mandated; the class and the numbers are.

How many sampling points do I need, and how many detectors?

Points first, detectors second, and never the other way around. For a very early warning design at 200 ft2 per point, a 1,000 m2 hall needs 54 points [2]. The detector count is then set by transport time rather than by hole capacity: a four-pipe detector with 25 holes per pipe is rated for 100 holes, but the same unit carries a total pipe length limit of about 200 m and a single-run limit of about 100 m [16]. Because a single 100 m run at the largest bore consumes 52 of the 60 seconds the class allows, a well-designed hall usually ends up with more detectors and shorter runs than the hole count alone would suggest. Design the pipe, then count the detectors.

Where exactly do the sampling points go in a raised-floor hall?

In four places for a contained hall, plus the equipment where applicable: at ceiling level throughout the IT area, inside the raised floor voids that contain cables, in the exhaust or return air stream of every contained aisle, and at the return air path of the cooling units. Where the ceiling void is used as a return air plenum, the return air stream at ceiling level is also required [1][12]. Underfloor points sit high inside the void rather than on the slab, and inside equipment they follow the smoke migration path — top 10 percent of a sealed cabinet, at the ventilation exit of a ventilated one, at the upper vent of a naturally vented one [13].

Should I design to 60, 90 or 120 seconds transport time?

Sixty. The 90-second figure belongs to the early warning class and 120 seconds to the standard class [2]. Any data center that is designing for detection at the incipient stage is claiming very early warning, and the very early warning limit is 60 seconds. The insurer's criteria for data center very early warning detection also state 60 seconds explicitly [6]. In practice the design target should be well inside the limit — a modelled 40 to 45 seconds leaves room for the fittings, the filter loading and the field tolerances that the model does not capture.

Can a high-sensitivity spot detector satisfy a very early warning requirement?

Yes, under defined conditions. NFPA 76's very early warning class specifies 0.2 percent obscuration per foot at the alert and 1.0 percent at the alarm, and a multi-criteria spot detector can be listed to those thresholds [10]. The obstacle is coverage rather than sensitivity: a spot device claiming very early warning sensitivity still has to cover its assigned area, and at 200 ft2 per device the device count is four and a half times what the standard spacing table would give you — before the high-airflow reduction table is applied at all [2][15]. There are applications where this is the right answer, particularly in small rooms and inside individual enclosures. In a full hall, the device count is what makes aspirating detection the practical choice rather than the sensitivity.

How often does an aspirating system need testing and maintenance?

Operational testing is required every two years under the insurer's data center criteria, by a defined test method [6]. Sensitivity testing follows the NFPA 72 cycle for smoke detectors — after the first year in service, then every two years, extendable to five years if sensitivity has held within the listed range across two successive tests [15]. Filter changes are on condition and at turnover after any construction activity. Airflow baselines and threshold values should be recorded at commissioning and reviewed as a trend, because drift in the airflow reading is the earliest indication that the network is loading up or has been damaged.

Will an aspirating detector false-alarm on dust?

It can, and the controls are designed in rather than switched off. Aspirating detectors typically include a filtration stage that removes larger particles before the sample reaches the sensing chamber, plus a self-adjusting baseline that tracks the normal particulate background of the space, plus multi-level thresholds so that a rise in particulate produces an alert rather than a discharge [12]. The residual risks are real: construction dust, aerosol activities, and zinc whisker release from raised-floor work will all be seen first by the most sensitive detector in the room. The correct response is to keep the thresholds set against a measured background, change the filter after the work is finished, and re-baseline deliberately — not to raise the thresholds until the system stops reporting, which is how a very early warning system becomes a standard one without anyone deciding to make it so.

References

1. National Fire Protection Association, "NFPA 75: Standard for the Fire Protection of Information Technology Equipment," 2024 edition — Chapter 9 protection and detection equipment including the required detection locations, in-cabinet placement guidance, close-coupled cooling return inlets, Annex A.9.2.2 sensitivity and spacing guidance by functional objective, new Annex F performance test procedures for early fire detection, and removal of lithium-ion battery requirements to NFPA 855. nfpa.org
2. National Fire Protection Association, "NFPA 76: Standard for the Fire Protection of Telecommunications Facilities" — definitions of very early warning, early warning and standard fire detection; coverage per sampling point, alert and alarm sensitivity, and maximum transport time for each class; return air port requirements and the 4 ft2 per port limit; 2024 edition changes. nfpa.org
3. National Fire Protection Association, "NFPA 72: National Fire Alarm and Signaling Code" — Chapter 17 initiating device spacing and installation, including Table 17.7.6.3.3.2 smoke detector spacing based on air movement and its exclusion of underfloor and above-ceiling spaces and of air sampling systems, sampling pipe identification requirements, supply diffuser clearances, and Chapter 14 acceptance testing. nfpa.org
4. National Fire Protection Association, "NFPA 2001: Standard on Clean Agent Fire Extinguishing Systems" — automatic release provisions, pre-discharge time delay and abort station requirements, and the requirement that a single detection indication not release the agent. nfpa.org
5. National Fire Protection Association, "NFPA 855: Standard for the Installation of Stationary Energy Storage Systems" — lithium-ion applicability thresholds, separation and dedicated room requirements, the 2026 hazard mitigation analysis requirement, the shift from deflagration venting to active prevention, and off-gas detection. nfpa.org
6. FM Global, Property Loss Prevention Data Sheet 5-32, "Data Centers and Related Facilities," section 2.3.9 and section 2.3.10 — very early warning fire detection installed to manufacturer's criteria in telecommunication equipment areas and return air systems; 200 ft2 per sensor or port with one level and 400 ft2 with two staggered levels where cable trays obstruct smoke flow; return air monitoring at 4 ft2 (0.4 m2) of opening per port; no port within 3 ft of a supply duct register; transport time not to exceed 60 seconds; smoke testing at installation with attention to ventilation and stratification; two-yearly operational testing; alarm transmission to a constantly attended station; and the release sequence in which pre-alarm shuts off air conditioning and closes fire dampers and alarm discharges after a delay not exceeding 30 seconds. fm.com (PDF)
7. FM Global, Property Loss Prevention Data Sheet 5-48, "Automatic Fire Detection," April 2025 interim revision — very early warning fire detection for high-value and critical areas including cleanrooms, telecommunication centres and critical computer rooms, and the recommendation to install detectors or sampling points inside electrical and electronic cabinets for fastest response. studylib.es
8. System Sensor (Honeywell), "Aspirating Smoke Detection Applications Guide" — NFPA class requirement table with coverage, alert sensitivity, alarm sensitivity and transport time; requirements of standard, early warning and very early warning systems; two-level high and low coverage rules; the aspirator, sensing chamber, filtration and exhaust components; and the requirement that the detector raise a trouble signal when airflow falls outside the manufacturer's specified range. intelligentlifesafety.com (PDF)
9. System Sensor (Honeywell), FAAST aspirating smoke detector pipe installation user guide — sampling port spacing by classification (200 ft2 / 14 ft / 7 ft, 400 ft2 / 20 ft / 10 ft, 900 ft2 / 30 ft / 15 ft), transport time measured from the furthest port including the test port, sensitivity at pre-alarm and alarm for each class, modelling requirements, and exhaust piping returning to the sampled area. manualslib.com
10. Siemens, "Model OOHC941 Multi-Criteria Fire/CO Detector installation instructions" — UL 268 seventh edition listing, very early warning classification with a programmable alert threshold at 0.2 percent per foot obscuration and an alarm threshold at 1.0 percent per foot, the 30 ft and 900 ft2 spacing basis, minimum distances from walls and ceilings, and the prohibition on installation in the airflow of a supply diffuser. cache.industry.siemens.com (PDF)
11. Honeywell Analytics, "Codes Address Aspirating Smoke Detection" — the respective roles of NFPA 72, NFPA 75 and NFPA 76 for aspirating detection, the recognition of early and very early warning detection, and the statement that each sensor or port in a space is limited to 200 ft2 with an exception permitting 400 ft2 per port or sensor where two levels of detection are provided. honeywellanalytics.com (PDF)
12. Fike, "Data Center Fire Hazards" — verbatim NFPA 75 detection requirements at ceiling level throughout the IT equipment area, below the raised floor containing cables, in the exhaust or return air stream where aisle containment is used, and in the return air stream where the ceiling void is a return air plenum; the Annex A guidance that detection for summoning people should be high sensitivity and close spaced while detection for agent release should be lower sensitivity with logical confirmation; and the sensitivity contrast between spot and aspirating technologies. fike.com
13. National Fire Protection Association, report of the technical committee on NFPA 75 — new guidance on sampling port and detector placement where smoke is most likely to migrate (top 10 percent of a sealed cabinet, at the ventilation exit of a ventilated cabinet, at the upper vent of a naturally vented cabinet), the close-coupled cooling return inlet requirement, and the provision that hot or cold aisle containment shall not obstruct the free flow of a gaseous clean agent to the IT equipment or the cooling system serving the contained aisle. docinfofiles.nfpa.org (PDF)
14. Security Info Watch (Bernard Sokol, System Sensor), "Data Center Fire Protection" — the three permitted approaches to protecting above-ceiling and below-floor air circulation spaces (detection throughout, detection at each smoke damper, or detection at other smoke entry points), aspiration sampling downstream of the filters on the supply side of CRAC and CRAH units, and the risk that an airside economiser cycle introduces smoke from outside the building. securityinfowatch.com
15. ACHR News, "Smoke Detection Requirements in Data Centers" — NFPA 72 Table 17.7.6.3.3.2 with 7.5 air changes per hour or fewer giving 30 ft and 900 ft2, and 60 air changes per hour reducing coverage to 125 ft2 per detector; the note that the table does not apply when air sampling or projected beam detectors are used; the observation that modern data centers often exceed 60 air changes per hour and that the table does not yet address it; sampling pipe labelling as "Smoke Detector Sampling Tube - Do Not Disturb" at changes of direction and penetrations and at intervals not greater than 20 ft; the very early warning alert and alarm thresholds; and the NFPA 72 sensitivity testing cycle with the extension to five years. achrnews.com
16. Xtralis VESDA aspirating detector product data (VLP series and VLS series) — sampling network limits of 200 m total pipe length and 100 m per run with 25 mm outside diameter and 15 to 21 mm internal diameter, maximum coverage in the order of 2,000 m2 subject to local code, ASPIRE2 pipe network modelling software, four alarm levels (Alert, Action, Fire 1, Fire 2) and two fault levels (Maintenance and Major Fault), alarm sensitivity range from 0.005 to 20 percent obscuration per metre, detector to first sampling hole distance of 1 to 15 m, pipe fixing spacing, end cap provision, the prohibition on right-angle elbows in favour of swept bends or paired 45-degree fittings, and vertical sampling hole spacing by temperature increment or distance. Xtralis product literature (Chinese datasheets)
17. FFE, "Sensis Azure 2000S user manual," commissioning section — the widely used computer room commissioning challenge in which a one-metre length of 10/0.1 mm PVC-insulated wire is electrically overloaded for one minute with the detector expected to respond within two minutes of the end of the burn, and the instruction to introduce smoke at every pipe end to verify maximum transport time with alarm delays set to zero seconds during the test. manualslib.com
18. Notifier and System Sensor, air sampling application note — capillary sampling through ceiling tiles landing on the grid intersections that define the sampling layout, sampling pipe installed on the upper surface inside floor voids with grid spacing per local code, return air vent sampling at a maximum 0.4 m2 (4 ft2) of coverage per sampling hole reduced to 0.2 m2 (2 ft2) where the vent is small, and the worked example in which a vertical return air vent of 1.5 by 0.8 m requires six sampling holes. notifier.lu (PDF)
19. System Sensor, "Aspirating Smoke Detection" technical overview — definitions of response time and maximum transport time, the requirement that transport time be measured at the most remote or least sensitive sampling hole, the requirement that transport time and system performance be calculated by design software and confirmed during commissioning, and the requirement for a trouble signal when airflow is outside the manufacturer's range. thefirepanel.com (PDF)
20. Canadian Fire Safety Association (Mike Clouthier, Xtralis by Honeywell), "Aspirating Smoke Detection" — the three components of an aspirating system (sampling pipe network, detector with aspirator and sensing chamber, exhaust pipe), the transport time and coverage requirements of the standard, early warning and very early warning classes, and the requirement that exhaust air be piped into the same room being sampled to eliminate pressure differences that would otherwise affect transport time. canadianfiresafety.com
21. 119 Fire Control, "Clean Agent Fire Suppression Ventilation Requirements" — the exponential decay relationship C(t) = C0 * e^(-(Q/V)*t) for agent concentration under continuous ventilation, the working form Q = -(V/t) * ln(C_safe/C_0) for sizing purge airflow, typical purge targets of 10 to 30 minutes, the role of decomposition byproducts in setting the ventilation rate for halogenated agents, and the practice of applying a safety factor to the calculated flow. 119firecontrol.com
22. Johnson Controls, "Introduction to Clean Agent Fire Suppression Systems" — downtime cost reference table giving USD 300,000 per hour for a data center, USD 40,000 per hour for a clean room and USD 1,000,000 per hour for a large electronic data processing installation, used here as the business case for detection at the incipient stage. gallus.si (PDF)
23. Yonhap News Agency, "Recovery exceeds 40 pct following state data center fire" (13 October 2025) and "Recovery reaches 27.2 pct after state data center fire" (9 October 2025) — the 26 September 2025 fire at the National Information Resources Service facility in Daejeon, which began when a lithium-ion battery exploded in a server room, was completely extinguished the following day, and disrupted 709 government online administrative services. en.yna.co.kr
24. UPI / Asia Today, "Audit faults broad failures in South Korea data center fire" (7 September 2026) — Board of Audit and Inspection findings identifying 18 illegal or improper practices, unregistered electrical contractors and unauthorised subcontracting in the battery relocation, battery racks left energised during dismantling with cable terminals not properly insulated, approximately 22 hours to extinguish the fire, 384 battery racks destroyed, and assessed damage revised from about KRW 107 billion in direct losses to about KRW 351.1 billion including social costs. upi.com
25. IBM Documentation, "Conductive contamination" (zSeries 3932 specifications) — zinc whiskers as the most common conductive contamination in raised-floor data centers, their origin on the underside of zinc-electroplated access floor tiles, the requirement that data centers be free of zinc whiskers, particulate cleanliness to ISO 14644-1 Class 8, and gaseous corrosivity held to ANSI/ISA-71.04 severity level G1. ibm.com
26. MCS and NASA Goddard Space Flight Center, "Data Center Contamination" — zinc whisker growth mechanism and dimensions, growth rates up to approximately 1 mm per year following an unpredictable incubation period, raised-floor tile undersides and zinc-plated pedestals, rails and rack hardware as the most commonly reported sources, and breakaway during floor-bumping and construction activity with entrainment into the cooling airflow. mcserv.com