First, in thinking about fan-curves you may intuitively assume that a simple linear relation-ship between heat
(like in the following picture) perfectly protects your processor from any ill-effects:
This is not completely right. There are several side-factors which affect quantity of heat-production and quantity of heat-transportation over time. I'm only going to explain two of them in laymens words:
1) heat-production: while overclocking the processor you linearly raise voltage and/or clocks, but the relationship between heat and voltage and heat and clocks is exponential, which means aproportional, untill a (theoretic) point where very very tiny amounts of clock or voltage will result in a huge amount of additional heat produced. This effect has it's own name: "law of diminishing returns" and is rooted in the way how electrons move through a conductor - you can not change this relationship. You have the possibility to exactly measure your own personal curve by noting the Amperes your processor is consuming at various frequencies and voltages while at 100% utilization. Your fan-curve will be very similar to that one.
2) heat-transportation: this is a relative slow process, it's not an immediate effect, and the smaller the heat-difference between the heat-transporting material and the heat-source, the longer it takes. It means the bigger the difference in temperature (like 100Â° C) the faster heat will be taken away. With a fan-cooling the difference between the temperature of the heat-source (the metal-body on the processor) and the heat-"consumer" (the air) is relative small, this produces a considerable delay in heat-transfer. Additionally air does also move relative slow and can't be moved away instantaneously from the heat-source.
So there is a exponential raise in temperature and an exponential delay of heat-transfer, if both come together strong in the wrong moment you can cook your processor in the timespan the action-reaction-chain need to speedup the fan after the temperature-sensor started to see too much heat.
That's why you should make a fan-profile which takes preceding/anticipating steps to work against the delaying factors. Namely an exponential curve a step ahead of the heat-curve.
In the fan-profile editor you can add additional points to the line we had in the picture above by clicking on the line:
Then you grab the point with the mouse and move it around. The point will stay active (you see the highlighted box around it, the other points stay grey). To delete points, active it by clicking the point and pressing the delete key.
Now you add as much points as you need to make a rough exponential curve:
Now there are two additional things to have in mind to adjust this curve:
1) size of the fan or speed of the fan, in general the capacity of the fan to transport heat away
2) quality of the fan and it's bearings
The first point is related to the horizontal range of the curve. You see in the picture that at 84Â° the fan runs at 59%, if the fan is big it can be enough to compensate a 10Â° raise in a the next second, if the fan is small it may need too long and the processor jumps to 94Â° without any counteraction. So depending on the fan you got you may want to move the entire curve to le left, making the fan be a bit ahead of what may happen.
The second point is related to the vertical range of the curve. You maybe see that at 0Â° the fan is stopped. and then very slowly starts to raise it'S speed. The problem is than fans have a real difficult time to work at very low speeds like 10 cps or a 100. Not only is it electrically difficult, it's also very stressfull to the bearings. Mechanical resistance is a relative big factor with low speeds, but starting to disappear with higher speeds. So in this case you want the fan to rotate at a constant initial speed while being silent, efficient and reduces the stress on the bearings. You want to move the entire curve up, establishing a lower limit for the fan-speed:
I have a 5870 with cooler v1, and I conducted several experiments about how fast can the fan transport heat away and how low do I want to have the temperatures. I tested with various loads and kept in mind that the heat may be produced faster than transported away for a little timespan. Once I was happy with the regular behaviour of my curve (staying below Â°80) I started slowly overclocking the processor and in each step of raising voltage and frequency adjusting the curve as well, it goes like this:
a) have 100% load
b) raise core-frequency, search artifact
c) raise core-voltage, if there are artifacts
d) raise fan-speed at the specific temperature the core is actually at until at 80Â°
This is my final curve after overclocking to 1000MHz/1.3V/~82A:
The processor stays firmly below 80Â° for anything regular (non-FurMarks) and most of the time below 45% fan-speed. Even with FurMarks I have difficulties to really hear the fan. But keep in mind that is just because of how in my
card all of the mentioned factors play together (quality of the processor and the fan as well as temperature difference to ambient). I am lucky that I don't hear my fan.
Here is the final behaviour of my fan-profile in a real-world situation (STALKER CoP), approaching 65A:
If I'd play longer the temperature would converge slightly below 80Â° at approx. 44/45% fan-speed. My ambient temperature is 90Â°F+/32Â°C+ as I'm living near the equator.
Keep in mind your uttmost care is for not having a hot processor, and not for having a silent fan! If your overclock results in a loud fan which annoys you while being at 80Â° you should accept a lower overclock.
Please also keep in mind that you need to find out which is the maximum acceptable temperature of your processor (without trottling), care about the lifetime and subtract 20% and substitute that temperature with the 80Â° in this guide.
It's always better to be overly cautious when it comes to overclocking.
Okay, maybe you proof-read it or edit it. It can contain errors, that's because I don't know better, or maybe because I try to make an aspect overly simplistic. Keep in mind I also want to prevent that people burn up their chips, trying to give cautious suggestions.
Marry Christmas BTW
Edit: added real world measures.
Edit: added ambient temperatures