[Q] How and/or why does cooling a chip make it draw less power?
Posted by Droid_pro@reddit | hardware | View on Reddit | 33 comments
I've always noticed this aspect of hardware particularly while reading CPU delidding/direct-die results and have wondered why this happens. I would say ELI5 but I'd love to get technical if necessary.
Thanks and apologies if this kind of post isn't allowed here. Just figured it's time I learn and what better way to ask than a community of people who probably know more than I do?
Doormatty@reddit
Lowered temperature = Lower resistance
Lower resistance = Less power
gondezee@reddit
More resistance = less power. These are voltage driven, not current driven.
Droid_pro@reddit (OP)
How? My intuition leads me to believe that it would be the other way around. And what do you mean by voltage vs current driven?
gondezee@reddit
Semiconductors (generally, in normal practice, at normal temperatures, with normal materials) have a negative temperature coefficient. As temp increases the resistance decreases. Current is a function of voltage over resistance, so as the supply rail voltage is fixed with a decreasing resistance, current increases. Power is voltage times current, so current increasing means more power. Power is heat, and so the cycle continues. This is called thermal runaway.
PJ796@reddit
BJTs and diodes are NTC
MOSFETs, like those found in most digital logic ICs, are PTC and increase their Rds(ON) resistance as temperature goes up while Rds(OFF) gets lower
gondezee@reddit
Yup yup you’re right. Was thinking in linear region, not as switch.
PJ796@reddit
In the linear region a MOSFET doesn't have a drain-source resistance. It just draws a constant current irrespective of voltage
gondezee@reddit
There’s an inversion point ZTC, where at a low enough VGS, the VGSth has a negative temp coefficient effect where more current will be drawn at a fixed VGS at higher temps than lower temps.
PJ796@reddit
Only below the ZTC point will it draw more, which is pretty hard to get to when the MOSFET is barely conducting, at least without a dedicated heater, as for CMOS logic it requires that it both gets really hot and is supplied with a voltage near Vgs(th) as that's where ZTC usually is.
gondezee@reddit
Yea not logic. Admittedly I don’t work in semis or vlsi, I’m in board and product level so mostly interconnect and analog/dc land, chasing SI and EMC gremlins around. We’ve seen our fair share of fried FETs in PSUs and step down converters, mostly EOS due to transients (dear readers: please don’t run cat5 outside kthnx), but have had some thermal induced failures too. Really though our biggest issue with heat is baking the electrolytic in your caps, basically killing their filtering ability and inducing bigger transients. Heat is primarily generated by the power FETs though usually placed close to bulk electrolytics, and the FETs are what I see cratered with the sometimes ballooned caps. From this I did conflate my Si band gap theory here, thanks for challenging that, been brushing up on my FETs during meetings this morning.
Droid_pro@reddit (OP)
Ah. That makes sense. Thank you
Pristine-Woodpecker@reddit
MOSFETs depends on gate currents exceeding a threshold, whereas BJT essentially amplify the base current by a constant factor.
CMOS design essentially (in the ideal sense) never has any current flowing through it (which is why resistance is not as relevant as you might think). It just switches from one off state to another off state. It's the "imperfect" leakages that cause it to use power.
Visotoniki@reddit
More resistance more power loss in the form of heat, which in turn increases the resistance, if the cpu wants 1.1v its gonna get 1.1v, core voltage cant be left to random voltage drops based on resistances.
gondezee@reddit
The voltage rail is fixed, current is combined leakage plus drain source current when the gate is open. Both the leakage and drain source current increase with temperature due to increased carrier mobility.
danfay222@reddit
In semiconductors resistance generally decreases as temperature increases
Doormatty@reddit
TIL! I had no idea this was a thing!
Thanks for teaching me something!!
Strazdas1@reddit
Yes, its not intuitive but its a thing. This is why theoretically a chip could perform faster when hotter all else being equal. In practice its not really something youll observe as there are other downsides to being hot.
jizzajam@reddit
Fwiw I believe that was the one of the first observations that started semiconductor research
PJ796@reddit
BJTs and diodes are NTC
MOSFETs, like those found in most digital logic ICs, are PTC
SomeGuysFarm@reddit
I've never thought about this particular problem, but:
The supply voltages are (essentially constant). At least a trivial analysis would indicate that lowering resistance would increase current. Power goes as I\^2, so unless there is some kind of constant-current limiter involved, it seems like power should go up?
It's been over 3 decades since the little bit of chip design I did in college, so I've essentially no clue what's implemented in modern CPU designs and am asking to learn something.,
Affectionate-Memory4@reddit
That's going to be a big old "it depends" nowadays. Semiconductor resistance goes down as temps go up, but conductors do the opposite. Which one dominates your chip could even vary depending on the temperature range it is currently in, as they do not have to be perfectly linear RT functions.
You could in theory have a chip that is more efficient at 50C than 30C for example, because the semiconductor resistance value has gone down more than the conductor resistance has risen. You could also see the opposite thing happen on a different chip.
Qesa@reddit
It largely depends on where you want current flowing. Typically you want current going through wires and activate semiconductors, while you don't want it going through inactive semiconductors and insulators. The higher temperature results in more leakage due to decreased resistance in inactive semiconductors and insulators, meanwhile the higher resistance in wires manifests as decreasing the maximum clock speed attainable at a given voltage (due to voltage loss in the wires). The latter decreases power as well as performance, but at the same perf the higher voltage needed means more power too
Affectionate-Memory4@reddit
I agree entirely. I only mention that theoretical chip as an interesting thought experiment, but in any practical setup, it's going to behave like you're describing here.
DonTaddeo@reddit
The transistors work better as switching devices (lower on resistance) at low temperatures. Thus, all things being equal you can achieve higher clock speeds if the cooling of the CPU is improved. Alternatively, you can use lower supply voltages to reduce power consumption.
Also, as others have pointed out, leakage currents decrease as well and that reduces power consumption.
TheNiebuhr@reddit
In addition to other answers, in gpus the power delivery also suffer more leakages when hotter.
danfay222@reddit
It’s been a while since I’ve done semiconductor physics so I’m a bit fuzzy, but iirc the dominant effect is that leakage current rises pretty significantly as temperature increase, so dropping the temperature cuts your wasted power consumption.
Droid_pro@reddit (OP)
I see. A lot of sources are saying internal resistance is also a factor. You'd say that leakage plays a larger role in increasing power draw at higher temps?
Pristine-Woodpecker@reddit
I'd say so yeah. Leakage current is (IIRC) exponential with temperature, so it quickly will overwhelm resistance as temperature goes up.
Higher temperature also means a higher required voltage to achieve the same gate switching speed, and Power=CV\^2f, so there you have another quadratic increase in your power usage as temperature goes up. (You can also not increase voltage, and see your software crash as the circuit does not meet timing requirements...)
Droid_pro@reddit (OP)
Fascinating. I should have taken more transistor theory classes.
danfay222@reddit
Semiconductor and conductor resistances tend to go opposite directions with temperature, so I’m honestly not sure what the net effect would be.
Droid_pro@reddit (OP)
Ah that makes sense. Thank you. It's amazing that we've figured any of this out tbh
Visotoniki@reddit
From physics point of view shit is more efficient when cool, less resistance if I remember correctly but when it comes down to pc hardware most hardware nowdays overclocks itself based on thermals, so better thermals might translate to higher clock rates which usually are achieved by consuming more power.
Doormatty@reddit
Cooling a chip in no way makes it draw any less power.