Simple question: As a CPU gets hotter during use, does the efficiency go up or down? By efficiency I mean the computations done per heat produced.

If efficiency goes down with increasing temperatures , it seems like it would overheat from the positive feedback. High temp causes more heat to be produced, raising the temperature still higher.

If efficiency goes up with temperature, then ideal conditions would be as hot as practical. You'd need less heat dissipation overall and it would be much easier to move heat from the smoking CPU.

 

I believe generally CMOS circuits (which I believe all modern CPUs are) gets less efficient as the temperature goes up (MOSFET ON resistance has a positive temperature coefficient).
From that it could overheat if not properly cooled, but they typically have a fan cooled heat-sink to keep their temperature below the critical point.

On my HP desktop PC I notice that the fan speed increases when the computer is doing computational intensive applications, as apparently they speed up the core for such calculations, which dissipates more power.

 

And it would follow that the maximum clocking rate would be reduced at high temperatures. Or put another way, you could clock it faster when cooler. The amount of power dissipated is a function of capacitance, voltage and frequency, which are only an indirect function of temperature as mentioned above.

 

So at a given clock rate, running a constant task, do we agree that the amount of heat being generated actually increases with temperature? That means the cooling system has to be very aggressive to counteract the positive feedback of the "heater" (the CPU).

 

So at a given clock rate, running a constant task, do we agree that the amount of heat being generated actually increases with temperature? That means the cooling system has to be very aggressive to counteract the positive feedback of the "heater" (the CPU).

I'm under the impression that such is the case. It's called runaway heating, and it increases exponentially if it's not aggressively taken care of. I've seen that phenomena manifest in hydraulic systems, but it wouldn't surprise me if it applied to electronic circuits too.

 

If the energy per clock cycle is 1/(C(V^2)) and power dissipation and therefore heat generated is equal to clock frequency times the energy per clock cycle, I don't see how changing the temperature can have an effect on heat being generated.

What mechanism are you considering as a possible cause for heat generation to increase?

 

I was wondering about the temperature coefficient in semiconductors. A CPU is a boatload of transistors. If they all have their resistance go up or down in response to a temperature change, this should affect heat formation. Hold all else equal (clock, CPU load, etc.).

 

If the energy per clock cycle is 1/(C(V^2)) and power dissipation and therefore heat generated is equal to clock frequency times the energy per clock cycle, I don't see how changing the temperature can have an effect on heat being generated.

What mechanism are you considering as a possible cause for heat generation to increase?

Well, among other things, I guess it has something to do that, in a chip, the electronic pathways are distributed throughout its volume, and it only has its outer surface (or area) as its only means of dissipation.

 

Leakage currents are a major factor in modern processors.

http://www.ruf.rice.edu/~mobile/elec518/readings/DevicesAndCircuits/kim03leakage.pdf

 

As NS noted, leakage currents are a large factor in power dissipation and that would go up with temperature.
The other large power dissipator is charging and discharging of the various parasitic capacitances at the high clock rate, and that likely doesn't change much with temperature.
The change in ON resistance with temperature actually would have little effect on the total power dissipation, only on the maximum switching speed.

 

Yes, we can agree that dissipation increases with temperature.

 

Yes, we can agree that dissipation increases with temperature.

I'm not sure if you're joking about the level of agreement here? I mean, dissipation obviously increases and we can apparently all agree on that since it is self-evident.

If you were not making a joke, are you weighing in on the side of heat production – joules per second – increasing with temperature?

 

In post #6 I wondered what would make power dissipation increase with temperature. In post #9 nsaspook reminded us about leakage current and since we know that leakage current increases with temperature, with rising temperature power dissipation will increase if other factors are held constant. Crutschow nicely summarizes the factors contributing to power dissipation in post #11. Being in agreement, I withdraw my skepticism.

 

Hold all else equal (clock, CPU load, etc.).

That is the answer to your question in post #1. There is no "positive feedback" mechanism between the CPU die temperature and the rate at which it processes instructions. The clock frequency sets the CPU "speed", nothing else. Some systems measure the die temperature and alter the clock frequency to reduce heat, but that is an externally applied control loop and can have a simple or complex transfer function.

In a general sense, yes, computations done per heat produced is a very real concern in systems designed for low power, or high reliability, or rugged environment applications. This is why all modern microprocessors have extensive thermal models available.

ak

 

Here's a very interesting article, relevant to what's being discussed in this thread:

https://www.newscientist.com/articl...achine-learning-boom-means-we-need-new-chips/

 

In a general sense, yes, computations done per heat produced is a very real concern in systems designed for low power, or high reliability, or rugged environment applications. This is why all modern microprocessors have extensive thermal models available.

I can't tease out which side you're on. Does more heat get produced when a CPU is operated at a higher temperature? Yes, no change, or no it goes down. Again, all else equal.

 

To boil it down; if leakage current increases with temperature in a CMOS transistor, and if leakage current is the major cause of heat in a CMOS based CPU, then with everything else kept equal the rate of heat production in a CMOS based CPU will increase as its temperature increases.

 

Here's a very interesting article, relevant to what's being discussed in this thread:

https://www.newscientist.com/articl...achine-learning-boom-means-we-need-new-chips/

I keep hearing that silicon is a dead-end. Eventually something will beat silicon based computing but almost everyone of these 'new' discoveries have been rolled back into silicon based manufacturing and extended its life time for another 20 years.

https://cosmosmagazine.com/technology/why-silicon-computers-rule

 

... with everything else kept equal the rate of heat production in a CMOS based CPU will increase as its temperature increases.

That seems to be the consensus.

 

I keep hearing that silicon is a dead-end. Eventually something will beat silicon based computing but almost everyone of these 'new' discoveries have been rolled back into silicon based manufacturing and extended its life time for another 20 years.

https://cosmosmagazine.com/technology/why-silicon-computers-rule

I'm guessing that the industry is trying to avoid the switch to other materials, since the cost of developing new equipment and processes for the new materials, and replacing the old technology, would be substantial.