Heat, Airflow, and Thermal Derating

An ASIC turns nearly all the electrical energy it draws into heat. A 3,500 W machine therefore heats a room like a 3,500 W radiator. If that heat does not leave, die temperature rises, firmware sheds load or trips, and the hashrate printed on the spec sheet stops existing.

Where the heat comes from

Every hash trial dumps energy into silicon and power stages. Heatsinks, heat pipes, and plates do not make cold: they widen the exchange surface and the path into a fluid, air or liquid. That fluid then has to leave the chassis, and after that the building.

Two identical machines do not share a thermal regime if one inhales the exhaust of the other. So derating often starts before the miner, in how the hall is organized.

Airflow is a path, not a fan

Useful flow is a directed one: cool inlet, through the fins, hot outlet with no short recirculation. A fan spinning hard into a blocked plenum, behind a packed filter, or in a leaky hot aisle moves noise rather than heat.

The signs of a bad path are recognizable without instruments:

• exhaust air loops back in front of the intake;
• hot spots appear on one side of the chassis only;
• boards drop in sequence, always on the downstream side;
• fans run at full tilt while the room feels fine from the office end.

In hydro or immersion the same reasoning applies to a fluid: flow, return, final exchanger. A dry cooler breathing its own hot plume is the equivalent of an unsealed hot aisle.

What thermal derating is

Derating means reducing frequency, voltage, or the number of active chips to stay inside a thermal and electrical envelope. Vendor firmware already does it from its sensors. Custom firmware may trigger it earlier, later, or on less reliable logic.

It is an automatic response to a constraint, not a yield setting. In a hot hall, a derating machine makes less hash and often less heat, until it settles into a new balance that is sometimes unstable, oscillating between fan speed and power.

The vendor sheet generally assumes inlet air and a lab or purpose-built hall regime, not a sealed container in full sun. Environmental conditions are published in the model manuals, at Bitmain as elsewhere (bitmain.com). Those are the values that count.

Hot warehouse against lab bench

On a bench: few neighbours, inlet air close to room temperature, little dust, a clean measurement. The machine reaches its advertised behaviour without difficulty.

In a warehouse or farm: density, recirculation, filters, altitude, season, doors open toward the generators, and the maintenance carton left in front of an intake. The same boards see hotter, dirtier, less stable fluid. Sensors climb, protections trip, boards desynchronize.

Comparing a bench hashrate to a hall hashrate without discussing air is comparing two different machines.

Diagnose before you flash

Four questions to settle in this order, before touching firmware:

1. Is the problem local (one miner, one board, one fan) or hall-wide (an entire row)? Compare inlet temperatures across neighbouring machines, not just their hashrate.
2. Is inlet air genuinely cool, or only cooler than the exhaust? Measure at the chassis intake, never in the middle of the room and never at the office end.
3. Does cleaning a filter or restoring an air path change the behaviour? It costs nothing, it is reversible, and it resolves a large share of cases.
4. Do the measured temperatures sit inside the model manual's range? If they do, the derating comes from something other than ambient.

There is no universal target in degrees. The right temperature is the one the manual and the site policy accept, measured at the chassis intake. Until that measurement point is trustworthy, any firmware adjustment is correcting a number you do not actually know.