ASIC Cooling: Air, Hydro, and Immersion

Cooling an ASIC means removing the heat the chips make from electricity. A miner drawing 3,500 W rejects roughly 3,500 W of heat: nearly all the electrical energy ends up as heat, with nothing useful lost elsewhere. Three families dominate operations: air, hydro (a water / glycol loop at the machine), and immersion (dielectric fluid). Each moves the constraint somewhere else, and the choice comes down to what your site can absorb.

Air

Air crosses the chassis, washes the heatsinks, and leaves warmer. Fans set the flow. Reject heat goes straight into the room or a duct.

The design rule never changes: keep incoming cold air separate from outgoing hot air. In practice that means a distinct intake face and exhaust face, plus a duct or extract that stops hot air from looping back to the intake. An air miner in a closed room with no extract heats itself, because its coolant is the room's atmosphere: inlet temperature climbs, firmware derates, hashrate drops.

What that means day to day: filters to clean (dust insulates the heatsinks and pushes chip temperatures up), fan bearing wear, sensitivity to inlet humidity, and a noise level in the range of an industrial vacuum at one metre. The mechanics stay simple and the parts are common: a dead fan costs a few tens of euros to replace, and diagnosis is visual (blades, dust, flow direction).

The limit is physical. Air's heat capacity is low, around 1 kJ per kilogram per kelvin against roughly 4 for water, and the gap is far wider per unit of volume. So you move a lot of air for the same heat, which sets the floor on noise and duct cross-section.

Hydro (liquid at the machine)

So-called hydro units replace part of the air path with a cold plate or liquid block on the chips or sinks. A fluid (often treated water or a water-glycol mix, depending on the vendor) circulates, carries heat away, and dumps it into an external exchanger: dry cooler, tower, or plate exchanger.

Noise does not disappear, it moves to pumps and outdoor coils, which you can place outside or on the roof. That is the real operational gain: heat travels in a pipe instead of a duct, and the machine depends far less on hot-aisle temperature.

The sizing figure to watch is the dry cooler's approach, the gap between outdoor air temperature and fluid temperature leaving the exchanger. That gap cannot be shrunk at will: at 35 °C outside, a correctly sized dry cooler returns fluid noticeably warmer than 35 °C, and it is that temperature which has to stay inside the model's spec. A site designed around the annual average rather than the August peak derates every summer.

The risks specific to this principle: leak (short circuit and corrosion), blockage, dead pump, air trapped in the loop, fluid incompatible with the seals, and freezing if the site is not glycol-protected. A badly hosed hydro unit falls off its thermal contract and derates, even when the spec sheet promises more than an air model. The flows, pressures, and temperatures to respect are the ones in the model manual, not an average read on a forum; see hydro product lines on bitmain.com, for example.

Immersion

Immersion puts the hashboards (and sometimes more) in a dielectric fluid: electricity must not travel through the liquid, heat should. It then leaves through a secondary loop, an exchanger or a condenser depending on whether the fluid is single-phase or two-phase.

In single-phase, the fluid (mineral or synthetic oil) stays liquid, heats up, and circulates to an exchanger. In two-phase, it boils against the components and recondenses on a cold coil at the top of the tank; thermal performance is better, the fluid costs considerably more, and the tank has to be vapour-tight.

Contact with the components is intimate, which removes two air-cooling problems at once: hot spots from a badly aimed jet, and dust building up on the dies. In exchange you prepare each machine (remove fans, remove labels that peel, check glues and ribbon cables), handle filtration and periodic fluid analysis, control humidity above the tank, and plan an exit procedure with drip and dry time before any intervention. An air-designed miner dunked without preparation is not documented immersion, and voids the vendor warranty in most cases.

The point that matters for the operator: the fluid is an operations asset, with a per-litre purchase cost, a volume tied up in the tank, an analysis routine, and an end of life to dispose of.

What does not change

In all three cases, heat leaves the chip and then leaves the site. If the final sink (roof air, dry cooler, contracted water source) is saturated, the machine derates or stops, whatever the chassis is. Firmware still reads its sensors and trips if thermal policy says so. Hydro and immersion move the thermal problem outside the building; they do not remove it.

Another invariant: derating is gradual long before it is abrupt. A machine running 3 °C above its target does not stop, it gives back a few percent of hashrate for months. That is the most common loss and the least visible, because it raises no alert.

Choosing for your site

The right principle is the one the building, the water or air contract, and the homologated model can carry. Five questions to settle before ordering:

• Where does the heat go on the way out, and what temperature does that exchanger return at peak summer rather than annual average?
• Can the site take the noise at source (chassis fans), or does it have to move away (pumps, outdoor coils)?
• How much electrical capacity is available, and does the chosen cooling add auxiliaries (pumps, dry cooler fans) inside that envelope?
• Who responds to a leak, a blockage, or a contaminated tank, and within what contractual delay?
• Is the model you want homologated by the vendor for this cooling mode, or repurposed?

A poorly installed principle returns less than a well-installed different one. If you cannot answer any of those five questions, air stays the mode most tolerant of improvisation, and the only one you can walk away from without reselling the building.