One review has the big air tower two degrees ahead of the 360 mm loop. Another, published the same month, has the loop six degrees ahead of the tower. Same two coolers, same processor.
The room is the usual culprit, and it is almost never in the chart.
A temperature is not a measurement unless the ambient is
Everything a cooler does is relative. It moves heat from the processor into the air around it, and the processor ends up at some number of degrees above that air. The cooler's actual property is that difference; the number on your monitoring panel is the difference plus the room.
So if one outlet works in a 21-degree lab and another in a 26-degree office, the same cooler on the same chip reports 75 in the first and 80 in the second. Five degrees of pure geography, which is larger than the gap between most competing coolers.
This is why the outlets that take the subject seriously report a delta over ambient rather than an absolute, and control the room while they do it. It is also why a review that prints "78 °C" with no ambient stated is not a measurement you can compare to anything except the other coolers in that same chart on that same day.
Open bench or inside a case
The second variable is where the cooler was when it was tested, and both choices are defensible.
On an open bench the cooler gets room air, results are repeatable, and what you measure is the cooler. Inside a case it gets whatever the case's intake has already been warmed to — commonly five to ten degrees above the room under load — and what you measure is the cooler and the case together.
Both are useful and they are not interchangeable. An air tower and a front-mounted radiator behave differently once they are in a box, because the radiator is being fed by intake air while the tower is being fed by air the graphics card has already used, or the other way round depending on the layout. A ranking established on an open bench can genuinely invert inside a case, with neither test being wrong.
The processor fights back
Here is the effect that surprises people, and it has grown more important with every generation.
Modern processors do not draw a fixed amount of power and let the cooler sort it out. They watch their own temperature and current and raise their clocks until they hit a limit. Cool one down and it responds by drawing more power and running faster — until it arrives at roughly the same temperature it was at before.
Which produces a genuinely counter-intuitive result: a better cooler can show almost no temperature advantage at all. The advantage has been converted into clock speed instead. The temperature chart, in that case, is not measuring cooling performance; it is measuring where the processor decided to stop.
This is why cooler reviews increasingly report the package power alongside the temperature, or pin the power limit to a fixed figure before testing. A comparison at an unconstrained power limit is comparing two coolers that were asked to remove different amounts of heat, which is not a comparison.
It also means a review's ordering can flip between processor generations for reasons that have nothing to do with the coolers.
Fans at maximum, or at equal noise
The third divergence is what the fans were doing.
Run every cooler at 100 per cent and you learn which one is willing to be loudest. A 360 with three fans screaming will beat a tower with one — and nobody runs a machine that way.
Normalise for noise instead — set every cooler to the same measured sound level and then compare temperatures — and the ordering frequently changes, because a large tower moving air quietly is a different proposition from a radiator that needs airflow to work at all.
Two outlets, one testing at maximum and one testing at equal noise, will publish contradictory rankings from correct measurements. The noise-normalised chart is the one that describes a machine you would live with.
The mount is a variable too
There is a source of scatter that sits inside a single outlet's own numbers, and the careful ones say so.
Remove a cooler and refit it and the result moves. Different pressure, a different paste spread, a fractionally different seating — several degrees of difference between two mounts of the same cooler is common. A review that mounted each cooler once has a few degrees of unquantified noise in every bar.
The outlets that handle this properly mount each cooler more than once and report the spread, and their charts have error bars for a reason. If two coolers are separated by less than the mounting scatter, they are not separated.
A radiator changes the rest of the case
There is a cost to liquid cooling that a cooler chart cannot show, because it lands on a component the chart is not measuring.
A radiator has to go somewhere, and wherever it goes it changes the air in the case. Mounted in the front as intake, it warms every bit of air that reaches the graphics card behind it — so the processor's heat is being handed to the card. Mounted in the top as exhaust, the processor is fed with air the card has already warmed, and the case loses an exhaust path it might have wanted.
Neither arrangement is wrong, and the trade is a real one: front intake gives the processor the coolest air in the building and taxes the card; top exhaust protects the card and feeds the radiator second-hand air.
What matters for reading reviews is that a cooler test which reports only the processor's temperature has measured one side of that trade. A front-mounted radiator can look excellent in the chart while raising the graphics card's intake by several degrees, and the machine as a whole is not obviously better off. The reviews worth reading in a case context report both temperatures, or say plainly that they did not.
The same cooler on two different chips
One more reason rankings move between publications: the processor underneath is part of the result in a way that is not just about how many watts it makes.
Where the heat comes from matters as much as how much of it there is. A chip that concentrates its output in a small, dense area is harder to cool than one producing the same total across a larger die, because the bottleneck moves upstream into the processor's own packaging where no cooler can reach. On such a part, every cooler in the test compresses toward the same result and the differences between them shrink.
So a cooler that separated itself clearly from the field on one processor can finish in a cluster on the next generation, without changing in any way. The chart moved because the thing being cooled did.
What a comparable review states
You can tell within a paragraph whether a cooler review is going to be useful:
- The ambient temperature, and preferably a delta rather than an absolute.
- Whether the test was on an open bench or in a case, and which case.
- The power limit the processor was held to, or the package power reached.
- The fan speed policy — maximum, or normalised to a stated noise level.
- How many mounts per cooler, and the spread between them.
An outlet that publishes those five things can be compared against itself across years. An outlet that publishes none of them has given you a picture of one afternoon.
What to do
- Compare within one chart, never between two. The bars in one chart shared a room, a bench and a power limit.
- Prefer the noise-normalised comparison. It answers the question you actually have, which is how quiet the machine will be at a temperature you can accept.
- Fix your own ambient first. A case that runs eight degrees above the room hands eight degrees to every cooler you might buy, and no purchase removes them.
- Ignore differences smaller than a few degrees. They are inside mounting scatter, and you will reproduce neither of them.
Two things we would not buy
A cooler chosen because one review put it two degrees ahead. That gap is smaller than the scatter between two mounts of the same cooler, let alone between two rooms — and you are about to mount it once, yourself, on a different chip in a different case.
And a large loop bought on a maximum-fan-speed chart. That chart's ordering is substantially a ranking of how much noise each cooler was permitted to make, and the machine you take home will be running its fans at a fraction of that. Find the equal-noise comparison before deciding.
How this was put together
The methodology points here are drawn from the cooler testing that states its own conditions: Gamers Nexus's ambient-normalised delta approach, their noise-normalised comparisons and their published mounting-repeatability work; der8auer's measurements on mounting pressure and contact; Igor's Lab's case-versus-bench comparisons; and the power-limit discipline that the better outlets adopted once processors began converting cooling headroom into clock speed rather than into lower temperatures.
The derived point is ours: the five-degree difference that a 21-degree lab and a 26-degree office produce from the same cooler on the same chip — larger than the gap that separates most competing coolers, and absent from any chart that prints an absolute temperature without saying what the room was doing.
