Liquid Cooling Myths

[rge]

A cool CPU will generate more heat than a hot CPU (assuming the same clock, running the same application, blablabla)

At 0C ambient (outside) my i950 at 4.4ghz, 1.4 vcore loads at 50C and cpu uses and hence dissipates ~160W as heat (read by DES chip), and at 25C ambient but o/w same settings the cpu loads at ~75C and uses and hence dissipates ~180W heat using same loading program, seen same on multiple mobos and cpus. Cooler cpus consume less power, and since all consumed power in cpu is released as heat, I dont get your statement.

And as an aside, unlike in metals (conductors) where resistance increases as heat increases (from increased unwanted collisions from increased heat energy), in semiconductors and silicon and insulators in general, resistance decreases as heat increases. Here and scroll down, see negative Temp coefficients in germanium, silicon. Insulators have tightly bound matrix and dont conduct electricity well, so added energy via heat is beneficial to allow more free movement, ie conductance ie, less resistance. Same energy in metal, which already has good conductance, allows "too much movement" and start getting unwanted, interfering collisions.

But reason cool cpus generate less heat has to do with leakage. Hotter cpus leak more, hence more power consumption.

[Manicdan]

when the 480 first came it out it showed a 1W drop for every 1.5C drop in temps

water cooling is nice in how it does reduce total power consumption by lowering temps, however this might not be noticed by the majority due to 2 things
first is that we also overclock more because we have more cooling power
and second is that we have to factor in pump and additional fans consuming power aswell.

[rge]

It is interesting to think about in terms of coolant; if coolant A cools better than coolant B would that mean that there is more heat being released by the cooler chip?

By definition yes.

So, the CPU temp would actually be higher for the better coolant..wacky.

NO. If more heat went to coolant, less heat is in cpu (conservation of energy) hence better coolant equals lower temps. Read next part for why.

Or, would the better coolant be able to take care of the extra heat that is being generated by the cooler chip, thus keeping the CPU temp the same (relatively speaking)

The coolant is the transfer medium. In a fantasy world where no heat is dissipated, then the coolant in essence becomes part of the chip, hence in such a fantasy land temps would be same.

But in real life, better transfer medium = lower cpu temps, because heat is also dissipated faster with a better transport medium via coolant system, ie rad/fans etc.

[stephenswiftech]

The current does not decrease (perhaps it can a bit) it is rather wasted as heat. As the CPU heats up you'll see the current demand rise as well.




Tiborrr's test shows this well

where did the extra 30w go when the CPU was warmer?, did the cpu convert the 30w to more calculations? I think the 30w got converted to Heat!

Just keep things simple: a silicon has a higher electrical resistance at high temperature than at low temperature. the silicon is still supplied the same voltage no matter what its temperature is. Because U = R.I. if you maintain U the same value. as R rises, I decreases. P (electrical power through a resistor i.e. heat loss) equals U.U/R, in other words as R rises, P decreases. And as R decreases, P increases.

If this wasn't true, this mechanics wouldn't converge which would be absurd. (if what you suggest was the truth, then cooling would get easier and easier as the temperature goes down. If the heat load of the component you're cooling was going down as you decreases its temperature, then you would need less and less cooling effort which is absurd). If you've observed the opposite, then you are probably right but this isn't caused by the higher silicon impedance .


Might be something in tiborr's data that must have affected the results. But looking at power draw isn't the best way to evidence this phenomena as power draw depends on a LOT of things. If you want to evidence it, you need to limit the number of variables.

My data is from an actual resistive silicone which was monitored in current, voltage and internal temperatures.

Don't take it the wrong way, we are still talking about incremental numbers, like I said <1% heat for 10C difference.

[Church]

So, better cooled CPU uses slightly less power, worse/hotter one - more. It's more then offset (then mentioned 1%) though with power needed to bring better cooling (eg. pump & extra fans for rads). But in general, similar HW at similar overclocks dissipates close enough amount of heat, so in most cases users fanatically defending that with one cooling way room gets significantly colder or warmer then with another probably blatantly lie to themselves and others because of some empathically/placebo imagined 'how things seemingly work' reasons, or by errors caused imprecise way of 'measuring' difference with no testing methodics employed to eliminate other factors skewing results (eg. room temps are not measured but felt (and no care if 'thermometering' user has dressed differently), no care for seasonal temp differences, no closed room, no constant load, no care to directions hot air gets blown to in relation to user and how concentrated it is).

[MrBean]

There is also the question about what I consider acceptable noise level and I really doubt that when I get the extra 6950 and put both under water that the 360 will be enough with noise well under 30dB during load. But I might be wrong still, I am going to try it out though when the parts arrive!

Look at my project-log, and you will be amazed what can be done with a radiator. I have a single 240 Feser, and running 2x E5430 Harpertowns (quads) at 3.2gig, as well as a 5970 in the same loop, controlling it via a TMS-200, and NB 120's fans mounted - and I can tell you, the system is quiet - not silent, but really quiet, non-intrusive.

Been running like this for nearly 2 years, and Melbourne can get up to 45 deg C in mid-summer. Your idea though to test for yourself is good, that is the best advice I can ever give anyone. As you have mentioned, if your case can accept a large rad, and you have one why not - I have a Feser quad in my Wife's system, running an i7 920 (at over 4Ghz) with another 5970, just because I had the rad - my point simply was that there are to many on forums nowadays that states this size rad is needed, when indeed not so - not everyone's system can accept a quad, or even 360, then it would be good to know what will and what won't work, as opposed to what the 'general' trend is.

Of course, as many has mentioned, 'E-peen' seems to be more and more important nowadays

But, no worries Sheila, all good, appreciate your input.

That's a great point as well as, what is realistic for 'full load'. If I run prime95 and UniEngine's Heaven demo (vsinc off) side by side, my kill-a-watt tells me I'm using well over 300 watts, but no real world usage including video gaming with folding@home running in the background even comes close to that.

Precisely the point I was trying to make.

I run a process-simulator, as well as gaming and various other task in the fore/background, and rarely ever does the TMS-200 jack up the fan- and /or pump speeds. Running a dual quad, yes, even one as power-hungry as my Harpertown setup, allows for very low utilization across the cores, with pretty low heat-loads.

Of course you get 'em guys running OCCTPT or some other such program to prove a point, although all they're doing are (prematurely) killing their CPU's, as CPU's were NEVER designed to deal with those loads for extended periods in practice - and they will rarely ever encounter loads programs such as OCCTPT subject them to.

edit: Programs such as OCCTPT leads to enormous acceleration of electron-migration, and now wonder we have quite many 'premature' sudden deaths of various CPU's on forums such as these - just do a search, and see for yourself - and I am pretty sure quite a few were subjected to these 'stability=testers', while under pretty heavy overclocked conditions, i.e high Vcores. /edit

As you have seen, even Prime95 will not exceed 70-75% of the rated wattage-capacity of your CPU - for example, on my E3 1245 mini-itx Xeon setup, the chip is rated at 95W power-draw, yet, running 8 instances of Prime95 Torture-test only results in a 69W power-draw.

Yet, all 8 cores (4 physical, 4 hyper-thread) indicates 100% load in task manager.

Quite interesting.

[stephenswiftech]

So, better cooled CPU uses slightly less power, worse/hotter one - more. It's more then offset (then mentioned 1%) though with power needed to bring better cooling (eg. pump & extra fans for rads). But in general, similar HW at similar overclocks dissipates close enough amount of heat, so in most cases users fanatically defending that with one cooling way room gets significantly colder or warmer then with another probably blatantly lie to themselves and others because of some empathically/placebo imagined 'how things seemingly work' reasons, or by errors caused imprecise way of 'measuring' difference with no testing methodics employed to eliminate other factors skewing results (eg. room temps are not measured but felt (and no care if 'thermometering' user has dressed differently), no care for seasonal temp differences, no closed room, no constant load, no care to directions hot air gets blown to in relation to user and how concentrated it is).

slightly more power... not less. I have data between 20 and 40C, pretty much linear in this range.

[rge]

@churchy, yes better cooled cpus have slightly lower power consumption, at higher temps slightly higher power consumption via leakage. There are multiple research papers from chip design and research teams that have measured such, even developed mathematical formulas for measuring the increase in dissipated power at different temps, especially increasing leakage at higher temps as power density increases from die shrinks ie going to 32nm. searching 2 seconds.

Power dissipation of a chip is closely related to and is affected by the temperature of the chip. For example, dynamic power dissipation increases the temperature of the chip, which in turn increases leakage power dissipation. This is due to a non-linear dependence of leakage power on temperature. So, it is essential for any power estimation methodology to include temperature in its estimates.

Also in that paper, notice leakage power at 90 nm is very small, but ranges 20-40% (actually another paper) on 32 nm.

And yes, it is too small for anyone to notice at same normal room temp. Water cooling and air cooling going to put out roughly same heat in a room.

[CrazyNutz]

slightly more power... not less. I have data between 20 and 40C, pretty much linear in this range.

There are 2 examples (rge, and tiborrr) in this thread that show the opposite of what your claiming. The examples show less power usage for cooler CPU's.

Resistance increases in silicon when temp increases. Whenever current is limited with resistance it will produce heat.

The extra power usage is due in part of higher leakage as well.

[CrazyNutz]

dbl post

[stephenswiftech]

There are 2 examples (rge, and tiborrr) in this thread that show the opposite of what your claiming. The examples show less power usage for cooler CPU's.

Resistance increases in silicon when temp increases. Whenever current is limited with resistance it will produce heat.

The extra power usage is due in part of higher leakage as well.

Well I've got precise and local measurements (not overall power draw measurements which depend on many other things that are not controlled) and the theory (electricity 101) to back it up.

It's maybe not as easy to understand as I thought it was but I can try again: you decrease temperature, you decrease resistance. lower resistance means more current going through it. If more current is going through it means your power source has to deliver more current (= more power). Another example: take a simple power source, 2 wires and a resistor cooled by natural convection, no heat sink, nothing. Everything is fine, ambient is 20C and the resistor is maybe 25C. Now, the ambient rises 5C. So does the resistor (30C). What you are saying is, ok the resistance increases and so does the electrical power... but if the electrical power increases then the heat also increases which increases the resistor temperature again.. and so on. Your theory cannot be possible because it does NOT converge or does not tend to an equilibrium. Instead, if the ambient increase yes the resistance increases but as it does since the voltage has not change, then current has to drop. as it drops it find another equilibrium. the consequence of this is that the electrical power (and the heat output) have slightly dropped...

I can't find any better example. but that doesn't change the actual fact: a cool CPU does not use more heat or dissipate more heat than a hot CPU, it's precisely the opposite. Go back to the basic equation mentioned above that's all it boils down to.

Like I said, the measurements users have done show otherwise most likely because of other things - if you don't control them, you just can't claim they are not the cause of what's been observed.

[defect9]

well, uh, damn this thread exploded.

[stephenswiftech]

well, uh, damn this thread exploded.

ahah well you guys wanted to bust myths right? there u got a nice one ahah

[rge]

There are 2 examples (rge, and tiborrr) in this thread that show the opposite of what your claiming. The examples show less power usage for cooler CPU's.

Resistance increases in silicon when temp increases. Whenever current is limited with resistance it will produce heat.

The extra power usage is due in part of higher leakage as well.

As does every researcher and chip maker in the industry show....but i tire of arguing with what is common knowledge from mounds of reputable research:

With today’s high-end processors dissipating 10’s of Watts, the CPU may become overheated. Overheating leads to reduced reliability at the least, and CPU damage and system failure at the extreme. Overheating is not often a show-stopping problem in desktop systems (Intel Prescott CPU’s excepted). However, it is of great concern in commercial systems such as server farms. By adjusting the temperature of such systems, or maintaining them at suitably low temperatures, reliability should increase markedly. Also, power consumption should decrease, and system efficiency should increase.

http://www.cs.virginia.edu/~skadron/tacs05/uht.pdf

[tiborrr]

The reason for high power draw at higher die temperatures is due to current leakage at the gate.

CPU =! resistor

[CrazyNutz]

Proof positive that resistance in a CPU will cause more power consumption, and thus more heat. Be it long narrow wire or more resistive (from increased temp) wire.

An additional form of power consumption became significant in the 1990s as wires on chip became narrower and the long wires became more resistive. CMOS gates at the end of those resistive wires see slow input transitions. During the middle of these transitions, both the NMOS and PMOS logic networks are partially conductive, and current flows directly from Vdd to VSS. The power thus used is called crowbar power. Careful design which avoids weakly driven long skinny wires has ameliorated this effect, and crowbar power is nearly always substantially smaller than switching power.

Source: http://en.wikipedia.org/wiki/CMOS

Whether it be resistance, or leakage (I agree both play a part) cooler CPU's do not generate more heat, nor do they use more power than when they are hotter.
Hotter CPU's are operating less efficiently, and therefore they will generate more heat, and consume more power.

We can go on for ever, I'll just agree to disagree

[Tom128]

Remember that one time we were talking about water cooling myths? That was fun

[Church]

Hmm, i'll probably will side with tiborrr on this one. For a bit different reasons, - simply because cpus these days have everything dynamic. Frequencies, voltages, power. Eg. can one be 100% sure to Fully know at what algorithms these work for all CPUs, when they now can temporary self overclock way overspec but still keeping within TDP and power usage? And does someone 100% know how power regulation works on Any motherboard out there including voltage droop, power saving features and alike advanced features that may work or not or be buggy on some? It's as tyborrr said, (modern) cpu is not resistor, and without fully knowing advanced logic behind autoregulation one can only roughly measure specific hardware behavior at specific conditions.

[MrBean]

Hmm, i'll probably will side with tiborrr on this one. For a bit different reasons, - simply because cpus these days have everything dynamic. Frequencies, voltages, power. Eg. can one be 100% sure to Fully know at what algorithms these work for all CPUs, when they now can temporary self overclock way overspec but still keeping within TDP and power usage? And does someone 100% know how power regulation works on Any motherboard out there including voltage droop, power saving features and alike advanced features that may work or not or be buggy on some? It's as tyborrr said, (modern) cpu is not resistor, and without fully knowing advanced logic behind autoregulation one can only roughly measure specific hardware behavior at specific conditions.

Touche. Well said, and 100% in agreement

[Utnorris]

I am curious, has anyone done a test on how flow rates effect temps with current blocks? I.e. .5Gpm versus 1.5Gpm with blocks made within the last year. I am just curious as I don't see any difference when I crank my flow up to 1.25Gpm versus .56Gpm or so. IIRC it mattered when blocks were really restrictive, but with the newer blocks being so fine tuned now, do you still really see a "big" difference in temps with increased flow and I am talking like a 5c difference, not 1c? A lot has changed since the original tests were done, cpu's, blocks, etc., so I am curious if the 1Gpm-1.5Gpm sweet spot still applies.

[bds71]

i know of a material which will REALLY throw a wrench into all of your equations: graphene can't wait to see what this "little" guy will do!! (and the discussions trying to prove/disprove theories should be a hoot to see ) note: for those that don't know what this little miracle is - it is basically a single atomic layer sheet of carbon (similar to graphite, but not). it's properties deal with nano-scale properties (which seem to defy physics, but actually start relying more on quantum physics for its properties). probably won't see this for another 10 yrs or so, but who knows....this little guy is gonna change EVERYTHING!! just throw out the old rule books, because this guy breaks all the rules. "cool" stuff indeed.

[Church]

Utnorris: all skinneelab cpu block tests have charts with cooling dependence from flow. Vets advising 1gpm and telling that flow above nets almost nothing base that on those tests as well.

[eXa]

Stephen. Im pretty shure you are wrong on this one.
All where i look ive seen less wattage with cooler chips. The GTX480 was extreme in this sence and even watercooling it vs aircooling would decrease the wattage by quite an amount. Also ask the Ln2 guys, i think it was Chew* that stated that even with 2v on the cpu it didnt really put out more wattage than on stock voltage and roomtemp.

[NaeKuh]

And of course people like Naekuh who want a <1C delta !

Yeah, then there's Naekuh But, he's on a planet of his own, hehe. I am talkin the general populace here, that excludes him.

Soz Naekuh, had to.....

u bastards!!

i stand to pee, and i bleed RED BLOOD!!! dammit! im as male human as they get...
^ THIS IS NOT A XS MYTH!!!!! or thats what i keep telling skinnee!!!

Stephen. Im pretty shure you are wrong on this one.
All where i look ive seen less wattage with cooler chips. The GTX480 was extreme in this sence and even watercooling it vs aircooling would decrease the wattage by quite an amount. Also ask the Ln2 guys, i think it was Chew* that stated that even with 2v on the cpu it didnt really put out more wattage than on stock voltage and roomtemp.

actually u both are correct...

Putting on my serious hat...

What your saying is that by watercooling the die itself draws less power.
Well that has to do with overall interference from heat and materials. ie.. the super conducting metal we all look for.

As things get hotter, more energy is lost via HEAT.. adur... remember physics.. nothing is for free.
This is where ur seeing all the voltage leak... its HEAT.. and a LOT OF IT.

Now in chew's case... and dont get me wrong.. im not anywhere near chew's level in knowledge.. i think its because a physical limitation inside the arch which prevented the cpu from scaling any higher.

Once again physics... u can never get something for free. and you must always play the game.. and no u can never cheat!

[Utnorris]

Utnorris: all skinneelab cpu block tests have charts with cooling dependence from flow. Vets advising 1gpm and telling that flow above nets almost nothing base that on those tests as well.

My question or point was that while flow made a big difference on older blocks like the Storm or the original EK Supreme, this was due to their restrictive nature and that with newer blocks developed in the last year or so flow does not make a huge difference. If you look at Skinnee's test results here:

http://skinneelabs.com/2011-cpublock-compilation/2/

Going from .5Gpm to 1Gpm only gives you a 2 degree difference and going from .5Gpm to 1.5Gpm might net you at most 3 degrees. In the real world of things this will not make a difference, unless you just happen to be right on the line of being too high in temperatures, which would be solved by more radiators or better airflow. With the Storm or the old Supreme a high flow could net you almost ten degrees, that is big. Too me, saying flow is important is more myth than reality. Is it important, some, but not as much as it was a few years ago due to the improvements in water blocks. Deciding on whether to spend money on a second pump for better flow versus an additional rad, I would say that the additional rad would have way more benefit unless you already have more than enough rads for the system.