Ok, one of my cards has a fan on it that is dying, but it has given me a chance to observe a behavior in cgminer that, I believe, needs tweaking, when it comes to auto-fan and auto-gpu.
I notice repeatedly, that the temps of my card are fluctuating a lot, and this is why...
I'm using the following options (I was using hysteresis 4, changed it to 2 a few days ago thinking it might do better at holding the temp on card #2 set to 2 instead, didn't work out that way)
"auto-fan" : true,
"auto-gpu" : true,
"gpu-threads" : "1",
"gpu-engine" : "600-825,600-825",
"gpu-fan" : "0-85,0-85",
"gpu-memdiff" : "200,200",
"intensity" : "9,5",
"temp-hysteresis" : "2",
"temp-target" : "70,70",
"temp-overheat" : "80,80",
"temp-cutoff" : "90,90"
What ends up happening, is the #2 card, immediately sets up at 825 engine, 50 fan, upon startup.
Ok... would rather have it start out of the gate at default with 85 fan, but no big deal, it'll find its sweet spot in a few minutes, right? wrong...
There is some lag time between the heat being generated, and the sensor picking it up, so my temp is constantly fluctuating at least 5 degrees.
First, since the temp hasn't risen yet, cgminer starts stepping down the fan. Usually with bigger steps when the temp is below 68. At the same time, the clock is already at 825, and is building up heat. By the time the sensors pick it up, it is raising the fan speed quickly up to 85, at which point the temp is usually over 72-73, and then starts dropping the clock speed step by step to generate less heat.
So far so good. It will make its way, most of the time, all the way down to 600 before the temp starts to come down. The fan speed gets reduced along the way as expected... but when my temp finally drops below 68 degrees, instead of gradually bringing the clock speed back up 1 step at a time like I would expect, it slams it back up to full in one go, from 600-650 directly to 825 again.
Since the temp is still low, the fan speed continues to be reduced, until finally the heat being generated, reaches the sensor, which then starts ramping up the fan again, quickly, as the sensor's temp is rising fast, and the cycle continues, over and over and over again.
These temp fluctuations can't be good at all for the card. Granted, the fan on that card isn't operating as efficiently as it used to, but it is still running, it just doesn't spin as freely as it should. This just gave me a chance to see auto-fan and auto-gpu do its thing over a long period of time.
I have observed my card's temp fluctuate like this for days now, never finding a sweet-spot to settle into, which I would expect it should, since the card does cool down when the engine is at a certain point.
My suggestion would be a change in behavior to auto-fan and auto-gpu...
for auto-fan: allow the fan speed to rise as quickly as it needs to, but never lower it by more than a single step at a time.
for auto-gpu: allow the engine speed to fall as quickly as it needs to, but never raise it by more than a single step at a time.
It should never go directly to the maximum overclock speed, but get there gradually.
Making them adjust more slowly, in the direction that would raise the temp of the card, gives the sensors on the card more time to pick up the heat being built up, since heat isn't an instant indicator, there is some lag time for heat to propogate, but don't limit how fast the card can react when the temp needs to be cut down.
The rationale is, that while the sensor may think the card's chips are at 80-85 degrees that instant, those chips may actually be at 90-95, but that the heat just hasn't reached the sensor yet, giving the chips even more of a pronounced heat/cool cycle than they should.
-- Smoov
