Why do CPUs at negative temperatures perform better?

[Azsen]

I will be setting up a phase change vapochill system in the next week or so and I heard that when your CPU is at negative temperatures it actually performs faster?

1) So why and how do CPUs perform better at negative temperatures?

2) How much better do they perform? 10%, 20%, 30%+?

Thanks

 

[Calin]

The transistors that the microprocessors are made of are affected by many kinds of noise. One of those noises is the thermal noise (that increase with the temperature). The CPU will simply be more resistant to errors created by the thermal noise. Remember that if you manage to keep the processor's surface at -20C, you are reducing its internal temperature by some 60C. Compare that to the limit of functioning (some -200C to some +120C) and you get a bit of a improvement.
One more thing - colder transistors work slower than hot ones (the number of "mobilities" decreases with temperature). However, the idea is to have processor working good and as fast as possible, not as fast as possible but with errors.

How much better do they perform? You will find sooner or later - I have no idea. However, if you look for air cooled max overclockings and vapochill cooled max overclockings, you will get an idea

 

[dmens]

Lower temperature increases current and in effect decreases path delays.

 

[pm]

CPU's work better at lower temperatures because wire RC flight time improves because the resistance of the wires improves, and more significantly because the mobility of carriers improves at lower temperatures.

On die wire resistance (near 25C) decreases by about 0.4% per degree Celsius improving the RC wire delay. CMOS transistors speed by at lower temperatures due to increased carrier mobility. The increase in mobility leads to improved saturation velocity which increased current driving capability. There are three factors, called scattering mechanisms, that influence mobility with decreasing temperature: phonon, surface and Coulombic scattering. Phonon scattering is the dominant scattering mechanism at room temperature and is caused by lattice vibration. At approximately 80K, Coulombic scattering caused by impurities becomes more dominant as the effect from phonon scattering becomes less significant, and surface scattering from charge traps also starts to have an impact on mobility. nFETs improve slightly faster compared to pFETs due to the changes in temperature dependent mobility between electron and holes.

There are two concerns with lower temperature operation: increased noise, analog circuit operation and high-speed circuit issues. Noise increases with decreased temperature because the switching delays are decreased (from lower RC and faster FETs) leading to higher frequency transients. These signals can more effectively couple across wires due to the reduced resistance. Analog circuitry may not be properly characterized for extremely low temperature operation. Also, circuits that are designed to require a certain delay - such as pulse latches - can see the delay decrease to the point where they induce a failure.

 

[f95toli]

I guess someone should point out that there is no such thing as a "negative temperature", the lowest possible temperature is 0 Kelvin (room temperature is about 295 Kelvin).
The point is that there is nothing "special" about 0 degree Celsius (or Fahrenheit) in this case.

 

[imported_whatever]

It should be noted that at the same clock etc, the performance will be the same. It will just let you clock higher.

 

[iwantanewcomputer]

whoa whoa whoa now, we've got contradictions here.

pm said "mobility of carriers improves at lower temperatures"

calin said "number of "mobilities" decreases with temperature"

i don't think the number of mobilities changes, this is just electrons or wholes right? thermodynamically it would seem that higher T would increase mobility cause the electrons have more energy, thus jump from one atom to another easier

 

[Calin]

I will accept what pm said - it was a looooooooo.....ooong time since I have ahd any kind of contact with semiconductors (but I still think the sheer number of mobilities decrease). However, the current capacity is equal to their number multiplied by their "speed", and their "speed" increases with the inverted resistance.
I am sure at higher temperatures transistors have lower "resistance" - this is the reason why you can't put two transistors in parallel (the hot one will conduct more current, becaming hotter and hotter)

 

[f95toli]

At the temperatures you can reach with a vapochill system the number of charge carriers won't change much compared to room temperature, you need to go to much lower temperatures (below 100K at least) before you see this effect (which was discussed in a recent thread).
However, also the mobility changes with temperature so the real behaviour is actually quite complicated. From what I understand the performace of a CPU should increase as you lower the temperature down to about 100K or so, after that the performance should start to drop.

 

[pm]

Originally posted by: iwantanewcomputer
whoa whoa whoa now, we've got contradictions here.

pm said "mobility of carriers improves at lower temperatures"

calin said "number of "mobilities" decreases with temperature"

i don't think the number of mobilities changes, this is just electrons or wholes right? thermodynamically it would seem that higher T would increase mobility cause the electrons have more energy, thus jump from one atom to another easier

I consider this difference to be more a matter of semantics. The underlying answer is the same however you phrase it and I think we are both right.

I did a search on Google on:
"phonon scattering" temperature mobility

and this turned up a number of links.

The first one, and one of the most readable is:
http://ece-www.colorado.edu/~bart/book/transpor.htm

2.8.4, 2.8.5 and 2.8.6 are relevant to this discussion. The graph at the bottom is interesting as well.

 

[white]

we're talking about many different things here. silicon and the metal interconnects behave differently when increasing or decreasing temperature. in a metal, lower temperatures will increase electron mobility. in a semiconductor, it will increase, but eventually decrease at very low temperatures (<100K). its general shape is an upside-down 'U'. lattice scattering is responsible for the decrease in mobility at high temperatures whereas ionized impurities (dopants) are responsible for scattering at very low temperatures. in the temperature range you're working with, you're probably still on the part of the curve that will increase when you decrease the temperature.

carrier freeze-out is an issue when dealing with deep space applications due to the cold temperature, but it shouldn't be an issue here unless cooling with liquid nitrogen or something colder.

it will likely run faster, but not by an appreciable amount.

 

[Azsen]

Thanks for the responses guys. I understand Calin's answer. PM can you sorta dumb your answer down a bit for me please? I don't have a degree in electrical engineering to understand all that terminology. Ok this is highly techinical though...

I guess someone should point out that there is no such thing as a "negative temperature", the lowest possible temperature is 0 Kelvin (room temperature is about 295 Kelvin).
The point is that there is nothing "special" about 0 degree Celsius (or Fahrenheit) in this case.

I think the thing about 0°C being special is it is the point where water freezes. How that affects CPUs is what I want to know, maybe it doesn't?

Is it possible CPUs have built in wait states in them to control heat, and when that heat is gone it eliminates the wait states so it improves the speed?

 

[BriGy86]

all i know is that the higher the temp on a circuit is the less current it supports, i think because when the temp is higher the actual integrety of the circuit is less (more prone to breakage) but i could be wrong

all in all a colder computer is better (minus condensation

 

[MobiusPizza]

no. The higher temperature would liberate more conducting electrons from semiconductors. They conduct better when temperature is high.

I will be setting up a phase change vapochill system in the next week or so and I heard that when your CPU is at negative temperatures it actually performs faster?

I don't know whether that's correct. Where do you read that from? People chill processors just to overclock them much more easily.
I doubt the performance increase if there is any to be high for processors with same clock.

 

[pm]

Thanks for the responses guys. I understand Calin's answer. PM can you sorta dumb your answer down a bit for me please? I don't have a degree in electrical engineering to understand all that terminology. Ok this is highly techinical though...

I didn't exactly agree with the teminology that Calin was using in his answer, so my post was a bit in reply to his more than anything. I agreed with the conclusion that he was making, but not the mechanism responsible for the improvement, so I dropped into 'highly technical' mode to talk about what is actually responsible for the improvment. I don't think that there's a really good way to "dumb down" a conversation about about scattering mechanisms.

I think the thing about 0°C being special is it is the point where water freezes. How that affects CPUs is what I want to know, maybe it doesn't?

It doesn't. Aside from condensation as someone noted.

Is it possible CPUs have built in wait states in them to control heat, and when that heat is gone it eliminates the wait states so it improves the speed?

No. CPU's don't actually run faster at colder temperatures if the clock frequency isn't increased. As Annihilator noted, cooling is used to allow the CPU to be overclocked farther. There aren't temperature dependent wait states except in the unusual circumstance of when the CPU is in thermal-throttling mode (if the CPU supports this capability).

no. The higher temperature would liberate more conducting electrons from semiconductors. They conduct better when temperature is high.

This is true in an instrinsic semiconductor (ie. if you have a chunk of pure silicon). In highly doped semiconductors (ie. the ones that IC's are made with), the improvement from the ability of carriers to move through the lattice overwhelms the small decrease in carriers resulting in a net improvement in conduction at lower temperatures.

 

[Azsen]

Well just for the hell of it I will do some before and after benchmarks to see. I'll probably run say PC Mark 04 on air cooling at the same speed 2 times and average the results, then run PC Mark 04 two times at the same speed under VapoChill cooling to see if there is any difference.

 

[Spencer278]

Originally posted by: pm

I think the thing about 0°C being special is it is the point where water freezes. How that affects CPUs is what I want to know, maybe it doesn't?

It doesn't. Aside from condensation as someone noted.

Just wanted to point out that water doesn't condensate at 0°C but it condensates at the dew point which can be higher or lower then 0°C depending on the amount of water vapor in the air. You dont want anything inside your computer case exposed to air that is colder then the dew point.

 

[Indigopeacock]

I am just guessing.... higher temp, more excitability, more friction, more heat, more energy lost, more noise, so more unstable. Is that right?

 

[Azsen]

Well I benched before and after the VapoChill install at stock speeds with PC Mark:

3.2Ghz air cooling:
PCMark score: 5217
CPU score: 4942
Memory score: 5086

3.2Ghz VapoChill cooling:
PCMark score: 5209
CPU score: 4947
Memory score: 5113

So not much difference at all. It would seem with the Vapochill it improved on the CPU score and memory score but lowered the overall PCMark score? Haha what the... Then again those results are sorta within the margin of error.

 

[f95toli]

Not surprising. Unless you overclock your system nothing should (and can not since the speed is set by the clock) change.

 

[wchou]

what is negative temperature?

 

[Gannon]

Originally posted by: wchou
what is negative temperature?

He just meant in the "common" temperature terms. i.e. -30 degrees C or negative something farenheit.

 

[Loki726]

as others have said, most modern processors are based on synchronous systems, in which the speed will only depend on the logic implemented and the speed at which the system is clocked at.

That being said, I haven't taken any courses on ansynchronous system design (not until next fall at least) so I have no idea how speed is determined in such a system. Can someone here fill me in on how a processor based on an asyncronous design would work? I know that there have been some asynchronous designs but none of my professors have gone into them in any kind of detail other than mentioning that they are relatively hard to implement. Is it that there are several clocks that operate independent of one another, or does the next state start when the previous one is completed (like in a ripple carrier), or is it something else?. Also, would the speed of such a system depend directly on the speed of the individual circuit elements?