GPU servers consuming 10 kW per rack require massive airflow — typically 800–1200 CFM delivered by 60 mm or 80 mm counter-rotating fans running at 20,000+ RPM. The smart fan control board PCB orchestrates these fans, responding to dozens of temperature sensors across the system. A fan failure in an AI cluster means thermal throttling within seconds — making reliability the paramount design requirement.

System Architecture

A typical smart fan controller PCB manages 6–8 fan modules, each containing dual counter-rotating fans. The architecture comprises:

• Microcontroller: 32-bit ARM Cortex-M4 or M7 at 100–200 MHz with hardware PWM timers. Must support 12–16 independent PWM channels with 25 kHz carrier frequency

• Fan interface: 4-wire per fan — 12V power, GND, PWM (25 kHz, 5V logic), and TACH (open-drain, 2 pulses/revolution)

• Temperature sensing: 8–16 NTC thermistors or digital sensors (TMP117, ±0.1°C accuracy) placed at GPU inlets, CPU sockets, and exhaust zones

• Host interface: I2C/PMBus or SMBus to the BMC, reporting fan speeds, temperatures, and fault status

PWM Fan Drive Circuit Design

Each fan draws 1.5–3A at 12V (18–36W per fan module). The drive circuit must handle this robustly:

• MOSFET selection: N-channel, 30V Vds, Rds(on) <5 mΩ at Vgs=10V. Gate charge below 20 nC to support 25 kHz switching with acceptable losses

• Gate drive: Dedicated gate driver with 1–2A peak drive current for fast switching transitions. Slew rate limiting via gate resistor (10–22Ω) to reduce EMI

• Flyback diode: Schottky diode (1A, 30V) across each fan output to clamp inductive kickback. Fans are essentially brushless DC motors with significant inductance

• Current sensing: 10 mΩ shunt resistor with INA181 or similar current-sense amplifier — provides per-fan current monitoring for stall detection

Tachometer Feedback and Fault Detection

The tachometer signal is an open-drain pulse train at 2 pulses per revolution. At 20,000 RPM, that's 667 Hz. The controller must:

• Measure RPM: Capture timer counting pulses over a 0.5–1 second window with 1% accuracy

• Stall detection: RPM below 30% of target for >3 seconds triggers a fault

• Degradation monitoring: Trending RPM at a given PWM duty cycle to detect bearing wear before catastrophic failure

• Redundancy logic: If one fan in a counter-rotating pair fails, the surviving fan ramps to 100% duty and the BMC is alerted immediately

Hot-Swap and Redundancy

Fan modules in enterprise servers are hot-swappable. The fan controller PCB must support:

• Hot-swap controller per fan slot: Inrush current limiting (fans have significant input capacitance — 100–470 μF per module). Inrush limited to 5A over 2 ms

• Presence detect: Short pin on the fan connector — grounded when the fan module is seated. Must be debounced (50 ms) to prevent false insertion events

• N+1 redundancy: With 8 fans in an N+1 configuration, 7 fans provide full cooling. The 8th is a hot spare. The controller must seamlessly redistribute airflow when any fan fails

• Reverse polarity protection: P-channel MOSFET in series with each fan power rail, with body diode orientation preventing damage from miswired fan modules

PCB Layout Considerations

• Layer count: 6–8 layers. Layer 2 as solid GND plane for EMI control

• Power distribution: 12V input at 20–30A total. 4 oz copper on power layers; kelvin sense traces to the VRM feedback node

• Thermal: MOSFETs rated for 2–3W each at full load; copper pours with thermal vias to internal planes

• EMI: 25 kHz PWM edges with <50 ns rise times can radiate significantly. Ferrite beads on fan outputs, RC snubbers (10Ω + 1 nF) at each MOSFET drain

• 3.3V regulator: On-board buck converter (12V to 3.3V at 1A) for MCU and digital sensors — must be low-noise to avoid corrupting temperature readings

Firmware Considerations

The fan controller implements a PID control loop with temperature as input and PWM duty cycle as output:

• PID constants: Tuned for the thermal time constant of the system (5–30 seconds). Over-aggressive tuning causes fan speed hunting

• Zone-based control: Multiple temperature sensors mapped to fan groups. The hottest sensor in a zone determines the fan speed for that zone

• Acoustic management: Fan speed ramp rate limited to 100 RPM/s to prevent audible annoyance from sudden speed changes

• Fail-safe: If communication with the BMC is lost, the controller defaults to 100% fan speed after a 30-second timeout