Platforms
A Fan Speed Command Is Different from Rotation Feedback
A PWM duty percentage requests a fan response; correctly interpreted tach feedback supplies evidence of rotor speed. Assess the exact part, compatible controller, declared response and feedback convention together. Then assess the separate evidence for the supported complete system: neither the command nor a valid RPM reading establishes delivered server cooling.
Read the command and the feedback as different signals
Pulse-width modulation, or PWM, expresses a control request through the fraction of each signal period spent high. Carrier frequency describes how often that control period repeats. Tachometer feedback is a separate return associated with rotation.
Noctua's undated PWM white paper covers its 5 V, 12 V and 24 V four-pin fans unless otherwise stated. It specifies a 25 kHz control target and two tach cycles per revolution. These are different signal conventions. A four-pin or PWM label alone does not establish controller compatibility or explain how a reported percentage was interpreted.
This manufacturer interface is an illustration for evaluating CPU-side air cooling in AI equipment. It does not establish either example fan below as a supported part in an OEM GPU server. For the offered system, retain the exact part and controller identities with their applicable documentation.
The exact model determines what a duty command requests
The current 12 V NF-A9 PWM specifications, checked October 21, 2026, list a maximum of 2000 rpm, 500 rpm at 20% PWM and 0 rpm at 0% PWM. Dividing 500 by 2000 gives 25%. Even this published point disproves a universal rule that the duty percentage must equal the same percentage of maximum RPM.
At zero duty, the current 12 V NF-A20 PWM table instead lists 350 rpm. A zero command can therefore coexist with intentional continued rotation in a named model. The two physically different models illustrate why the exact response matters; they are not substitute candidates.
Use those declared points for their stated purpose. They do not define all intermediate speeds or justify interpolating an unlisted response. For a supplier's claimed command-to-speed relationship, ask for the applicable response information and its conditions rather than deriving it from the connector label.
Work one fictional record from command to rotation
Consider fictional record FAN-60 for one identified part. Compatibility with the supplied 25 kHz PWM carrier is expressly stipulated, and the duty command is 40%. A separate, complete feedback record supplies 600 full tach cycles over 10 seconds under steady conditions. Its declared convention is two cycles per revolution, with no missing, duplicate or partial cycles and no timing ambiguity. These are invented inputs, not an actual measurement or instructions for collecting one.
The feedback frequency is 600 cycles / 10 seconds = 60 Hz. At two cycles per revolution, that represents 30 revolutions per second. Multiplying by 60 seconds per minute gives 1800 rpm. Equivalently, 60 Hz x 60 / 2 = 1800 rpm. Counted cycles are complete waveform repetitions, not a count of both signal edges.
The 25 kHz figure means 25,000 control periods per second. It does not mean 25,000 rotor revolutions per second. The separate 60 Hz feedback and its declared convention provide the basis for the rotation calculation. The 40% command alone supplies no missing maximum-speed value and is not automatically 40% of maximum RPM, airflow or power.
The useful result is positive but bounded: the stipulated feedback supports 1800 rpm for this identified rotor during the stated interval. Whether that matches the declared response requires that part's response information. Whether it satisfies the system's cooling requirement requires system evidence. Neither conclusion follows from the arithmetic alone.
References: [NWP]
Keep an absent return distinct from a reported zero
A supplier response with no usable feedback record leaves rotation unestablished. A displayed zero also needs the supplier's supported interpretation: what fan or channel does it identify, what quantity does the field represent, and does the return establish valid feedback? Those are documentation questions, not a diagnosis of the equipment.
In FAN-60, the supplied completeness and identity assumptions make the positive calculation possible. If those assumptions are removed, the record no longer supports the same inference. An unavailable signal and established absence of rotation cannot be silently treated as equivalent. Conversely, the NF-A20's declared zero-duty behavior shows why a zero command cannot itself be read as an instruction that guarantees a stopped rotor.
Ask the equipment supplier to state what its return establishes and what remains unresolved under the applicable system documentation. A missing or unexplained value supplies no basis here for declaring a failed fan or directing corrective work.
References: [NA20]
A sound RPM result still leaves the airflow question open
Rotation, airflow and thermal performance answer different questions. ebm-papst describes the airflow operating point through the relationship between the fan characteristic and system resistance; its performance discussion also distinguishes curves for different speeds. A speed value alone supplies neither relationship for the offered chassis.
The existing fan airflow and static-pressure article develops that downstream decision, including a complete fictional curve example. Use it when an offer moves from RPM to a delivered-flow claim. FAN-60 deliberately stops at rotation: it contains no chassis airflow, temperature, thermal-headroom or fan-unavailability result.
For the selected complete server, relate the supplier's supported configuration and operating conditions to the actual system requirement. If the offer claims sufficient cooling for a named component population, the returned evidence must address that claim at the system boundary. A clear rotor record is useful component evidence within that assessment; it does not fill a missing system result.
References: [E1]
Compare two returns for the same selected system
Suppose a buyer has already selected a complete server configuration and asks two suppliers to explain a fan-control and feedback offer. These are fictional documentation returns for the same proposed system, not competing replacement-fan recommendations.
Return A says four-pin PWM, 40% fan speed. It identifies neither the meaning of that percentage nor the part/controller pairing, response information or feedback basis. The buyer cannot tell whether the number is a request, a derived speed percentage or a shorthand label. A cannot yet substantiate even the narrower rotation claim, much less delivered cooling.
Return B identifies the approved fan/controller pairing for the selected configuration, cites its compatible command specification and declared response, and supplies FAN-60 with the correct fan identity and tach convention. It also identifies the applicable supported-system requirement and the separate OEM evidence addressing the offered component population and operating conditions.
B lets the buyer follow the claim in order: compatible command, declared response, interpretable rotation evidence, then the relevant system evidence. The buyer can recognize the 1800 rpm calculation as supported under the stated assumptions and assess the cooling claim against the separately identified requirement. If B omitted the system evidence, its control/feedback explanation would remain useful while its cooling claim remained unresolved.
The practical choice is which return supports the claim being assessed. Ask A to resolve its percentage and evidence gaps before making a like-for-like comparison. Assess B's system evidence on its own applicability; a more complete component explanation does not make a supplier's cooling conclusion true by itself.
Carry the selected equipment into the physical installation scope
Once the customer's appointed OEM and thermal specialists have resolved the equipment and control decisions, a separately approved rack installation or hardware-refresh project can use the selected complete configuration as an input. The rack-and-stack guide explains the physical package and handoff; the service page provides the relevant project discussion.
Leviathan's rack-and-stack service covers agreed equipment installation against approved rack positions, mounting instructions and handling conditions. Share the selected equipment configuration, rack layout, external cable plan, site constraints and remaining prerequisites through the rack-and-stack project inquiry. Fan selection, control changes and thermal approval remain with the appointed specialists.
Primary sources
- [NWP] Noctua PWM specifications white paperChecked 2026-10-21. Undated eight-page manufacturer paper. Its stated scope is Noctua 5 V, 12 V and 24 V four-pin fans unless otherwise noted. Used for control/feedback separation and tach interpretation; exact current product tables supply the named endpoint speeds.
- [NA9] Noctua NF-A9 PWM specificationsChecked 2026-10-21. Current named 12 V PWM model. Performance entries support 2000 rpm maximum, 500 rpm at 20% PWM and 0 rpm at 0% PWM. These are component specifications, without server compatibility or delivered-cooling evidence.
- [NA20] Noctua NF-A20 PWM specificationsChecked 2026-10-21. Current named 12 V PWM model. The 350 rpm entry at 0% PWM illustrates continued rotation at zero command. Other response points are not interpolated; this physically different model is not a substitute recommendation.
- [E1] ebm-papst North America FAQ: PWM and fan performanceChecked 2026-10-21. Undated manufacturer FAQ. PWM discussion explains duty and part-specific response; fan-performance discussion explains the operating point and speed-specific curves. Its general mechanisms do not establish any particular server's cooling performance.