Relationship between cpu/core temps, vcore, and stability

[DrMrLordX]

This is kind of a silly question, but I've been wondering about the relationship between cpu temps, vcore, and stability. As we all know, when overclocking, raising clock speed requires increased vcore which can drastically increase temperatures. The increased temps can require even further increases to vcore to assure stability. This is why we all try so hard to keep our processors cool, since it allows us to achieve stability at any given clock speed with less vcore than would be required with inferior cooling.

But, how does this work in the opposite direction? If I have a processor at stock clock speeds and I attempt to undervolt it, eventually, stability will be negatively affected if I reduce vcore too far. But if I begin cooling my processor to sub-ambient temperatures, will I get more headroom (or should I say tailroom?) with undervolting? The interesting thing to consider here is that, when overclocking, typically the net amount of power dissipated by the processor and consumed by the cooling solution(s) increases as we increase clock speed. However, if we undervolt while using aggressive sub-ambient cooling methods, then the amount of power required to operate the processor and the cooling system will both decline as the processor is further undervolted. Considering how power-hungry sub-ambient cooling solutions can be, though, I doubt you'd be able to run a heavily undervolted processor at lower net power usage with phase change cooling or TECs than you could run it at stock vcore with the stock HSF.

Anyway, are there any folks out there with phase units that can chime in on this? Can you undervolt your CPU more at stock speeds with the chip running at -40C or lower than you can with the chip running at or near room temperature?

 

[Idontcare]

It's like you made this post just for me

Here's the curve for minimum required voltage versus clockspeed for my B3 stepping Kentsfield QX6700 under phase-cooling (-40C):
http://i272.photobucket.com/al...ableVcoreversusGHz.jpg

Look at stock clockspeed, 2.67GHz, at -40C I could get away with 1.06 Vcore at full-load small FFT and be 24hr stable. Compared to air-cooling at 40C with lapped IHS and lapped Tuniq the same chip needs 1.29V to be small-FFT stable.

However, running with reduced Vcore at -40C versus higher Vcore at +40C does not produce a net power savings at the wall for the computing system. The vapochill LS I used for all my phase cooling pulled ~150W from the wall (kill-a-watt) at these lower clocks, the PSU also pulled another 150W from the wall. Overclocking to 4GHz resulted in the phase unit pulling ~295W from the wall and the PSU pulled nearly 370W from the wall.

Reducing the Vcore from 1.29V to 1.06V reduced the CPU's power-consumption by ~33% (some 50W, which is significant) but the power-consumption penalty for turning on that phase system is just too much (about 3x the power saved at the CPU, or 150W) so in the end I netted about 100W more power consumption with my phase + under-volted CPU versus just running the CPU undervolted (but at much higher volts) on air.

If you think about it, the phase unit (any active heat transfer system) is going to need to be extremely efficient (>70%) in order for the power-consumption of the heat transfer system to outweigh the reduction in power consumption by the CPU. I imagine the fan on the HSF is probably about as efficient as it gets, far more efficient than a compressor unit with two additional fans (my phase unit).

I haven't thought about it deeply enough, but I'm sure there is some facet of thermodynamics involved here that precludes us from moving heat around in such a way as to get more work done and produce less net heat (less power consumption).

But for what it is worth I can't touch clockspeeds above 3.3GHz with this QX6700 on air (despite lapped IHS and lapped tuniq) whereas with phase cooling I can go all the way 4GHz and still be small FFT stable. The CPU might go higher still but the power-consumption outstrips the cooling capacity of my phase system and I get runaway thermal overloading. But that's another story.

 

[DrMrLordX]

Now that is interesting data.

It is not surprising that the power consumption of your VapoChill LS outmatched your savings from a healthy undervolt. There are a few odd reasons why I began thinking about this:

Stacked TECs might be able to reach temperatures far below -40C, but at a terrible penalty unless the heat load was brought sufficiently low (less than 10W). I do not think TECs are really "the answer" here though. Still, your observations are extremely interesting! A reduction of about 80 degrees Celcius allowed you to achieve stability at stock clocks while reducing vcore by .23V, which is significant, especially considering the processor in question. A more modern quad built on a smaller process might yield more interesting results, especially one with a nice, low vid. Still, driving a modern quad into sub-10W power consumption levels would be very difficult, since there would be (or should be) diminishing returns on power savings as vcore continued to drop.

Or . . . to put it another way, I think that if you brought temps down low enough, you might be able to undervolt enough to get your dissipated heat levels low enough that they might be manageable with some form of exotic cooling, though I do not think that TECs would be the answer, especially considering how much heat quads put out at stock and how cold you'd have to get your chip to get voltages low enough for the heat load to be manageable with a stack of TECs.

The obvious problem is continuing to lower temperature while maintaining the relative efficiency of a phase change unit (like your VapoChill). TECs are obviously not the solution, and liquid nitrogen . . . let's not go there right now. But, assuming you could continue to drive your minimum voltage (and processor heat load) down at lower temperatures, could one not apply cascade phase change cooling to achieve those lower temperatures and more drastic undervolts, assuming the motherboard supported such low vcore settings? Think of it this way: most commercial phase change units are configured to deal with heat loads of, oh, I don't know, 150-200W or more. A unit designed to handle smaller heat loads - such as those produced by a quad running at some ridiculously low vcore like .5v or lower - would probably not need to pull as much from the wall. Even a double or triple-stage unit could probably be built to run on less power than the aforementioned VapoChill LS were it designed with a small (50W or less) heat load in mind. At some point, power savings might emerge, though one must wonder: at what cost? Cascade phase change is not cheap.

In the long run, most of my inquiries are indeed silly, because that is an awful lot of work to figure out how to use exotic cooling to run a processor at its stock speeds. What should be of greatest interest to anyone curious about this subject is a technological development that caught my eye recently: superconducting transistors.

Superconducting transistors (http://www.newscientist.com/ar...es-pc-revolution.html), while interesting, admittedly will only operate at temperatures as low as .3K which makes them terribly impractical, and not even within the realm of exotic cooling known and loved by PC junkies worldwide. If they could get one of these lurvely transistors working under LN2 cooling, then you have an interesting proposition: a chip made of such transistors would almost certainly have a ridiculously low rate of power consumption due to the intrinsically low resistance of superconductors. Er, I think. Well anyway it'd probably be low. Low enough that keeping the chip cooled to -200C or so might not be so difficult. You might have a 1W processor that could blow the doors off anything we have now in terms of raw achievable clock speed. Granted, chips made from such transistors are a ways off, and will be very far off if everyone sits on their hands waiting for room-temperature superconducting transistors to become the norm.

The point to this entire mental exercise is to realize that, when coping with sufficiently low heat loads, exotic cooling such as triple (or quadruple, or quintuple) stage cascade cooling; LN2; or even (gasp) TECs can become viable commercial products at least for a small niche if not the general public. Just put it in a box (the PC case or otherwise) and make sure it doesn't make too much noise or cost too much, and there you go. Obviously this will not be a hit with the netbook crowd but what can you do about that?

If anyone could build a CPU based on these transistors and get the running at something above .3K then they very well should do it, especially if they could get the heat load to 3W or less during "normal" operation. I would think putting together a quadruple or quintuple-stage phase system to deal with a 3W (or lower) heat load might not be such a difficult or expensive thing (compared to an equivalent unit designed to handle heat loads in excess of 200W). Hell, a heat load that low might be manageable by way of stacked TECs.

But hey, I'm just rambling. I'd love to see how low a Nano or ULV Penryn could go, voltage-wise, under phase.

 

[Idontcare]

You would do well to read this: http://en.wikipedia.org/wiki/Threshold_voltage

Basically temperature changes the threshold voltage. Lower temperature means lower threshold voltage, which means you can set Vcc lower and still operate the IC in a stable fashion.

However there is a point of diminishing returns where lowering temps further and further does not yield a notable reduction in threshold voltage. At that temperature, and for lower temperatures, the power consumption of the IC is basically fixed.

If you note the basic form of the threshold equation is Vtn = Vto + Y*(A+B) where B is the temperature dependent part.

So even if temperature goes to zero (thus making B go to zero in the equation above) the threshold voltage will still asymptotically approach Vto + Y*A as temperatures are reduced. Hence lower and lower temperature, regardless of efficiency of reaching those temperatures, won't result in lower operating voltages and thus will not result in lowered power-consumption.

Sure ohm's law further helps reduce some joule heating contributions to total power-consumption (that is what your superconductors are doing) but they don't stop leakage and they don't change the fact that switching a transistor requires a voltage change at some location in the circuit (the gate in this case) and that drives power consumption because of capacitance alone even if resistance is zero.

There's no free lunch if you have to go sub-ambient with your cooling. Again I'm pretty sure this is buried somewhere in the thermodynamics of the total system that you can't get more from less when the total system is properly defined and taken into account. If you do more work you must create more entropy, yaddi yadda.

 

[DrMrLordX]

Thank you for the in-depth response. That goes a long way towards highlighting why even a chip comprised of superconducting transistors would be a pain-in-the-arse to work with should it require temperatures of -196C or lower, and why a sub-ambient cooling solution would not, in fact, be "free lunch".

I suppose it's a good thing that there is a minimum threshold voltage that further cooling can not reduce beyond a certain point, or else the gate could never switch the transistor off if things got too cold.