Why does heat impede higher CPU clock frequencies?

[Antoneo]

When people say, cooling solutions prevent damage to the chip, what type of damage are we talking about? What do fried cores denote, are the gates inside being destroyed?

I've been wondering because a lot of times, a better cooler (allows more heat to be transferred off the CPU) would allow higher frequenices, just why is this true? Does heat have some effect on electron flow? Is it a resistor effect?

 

[TuxDave]

In a nutshell, yes, the gates/devices have a maximum operating temperature before the they start breaking down.

You have the heat/frequency relationship backwards tho. Lower temperatures do not mean faster circuits, at least in a significant manner. It's more like faster circuits generate more heat and if you can't remove the heat, the core will break down.

 

[Painkiller]

When increasing frequency in say the copper interconnects, resistance increases as well. The increase in resistance causes attenuation of the signal (voltage drop). Thus increasing the voltage for example helps with overclocking. The increases in resistance also causes more heat as power = I^2*R, I being current and R resistance. As temperature rises, this also generally increases the copper interconnect's resistance compounding the problem.

Trung

 

[imgod2u]

Actually, I believe impedence decreases in copper as temperature increases. Moreover, resistance is more or less constant. However, looking at the power formula: P = IV (which is equivalent to P = I^2 * R), you see that voltage directly affects power.

 

[TuxDave]

How about a better equation.

P=C*V^2*frequency (per gate)

Higher frequency --> higher power --> more heat

 

[Lionstl]

IF you treat it as a capcitor, then
energy=(1/2)*C*V^2
and power=(energy/time)=energy*frequecy as reciprocal of time is frequency.

So P=(1/2)*C*V^2*frequency

Like TuxDave wrote "Higher frequency --> higher power --> more heat"

 

[kcthomas]

Originally posted by: TuxDave
the core will break down.

is this refering to dielectric breakdown? The oxide between the gate and the channel prevents electrons from passing through. i think as the temperature rises, dielectric breakdown occurs which is where the oxide which is usually suppose to block electrons, allows electrons to pass. I dont know if this screws up the transistor permanantly though.

 

[Megamixman]

Its actually the elctrons being excited to the point that they can jump the dielectric itself. oxides usualy have smaller temperature range then the copper interconnects or substrate itslef, meaning that most IC's are limited to their dielectric's interms of operating temperatures. Thats the reason why when you get too cold the dielectric can cease to function and act as the limiting material when the silicon substrate can clearly still act as a semiconductor.

 

[TuxDave]

Originally posted by: kcthomas

Originally posted by: TuxDave
the core will break down.

is this refering to dielectric breakdown? The oxide between the gate and the channel prevents electrons from passing through. i think as the temperature rises, dielectric breakdown occurs which is where the oxide which is usually suppose to block electrons, allows electrons to pass. I dont know if this screws up the transistor permanantly though.

If the oxide breaks down, a transistor wouldn't function as it should anymore. Current would be tunneling through the gate oxide at such large levels that most dynamic and perhaps some static gates would never function.

 

[Gioron]

There are a couple effects that contribute to the problem, its not just one thing that does it.

First off, as said before, chips can be modeled as a capacitive load (well, for the most part anyway, its not a perfect model) and in a capacitive load the power dissipated in the load increases as the square of the voltage and linearly with frequency. So as you pump up the frequency and voltage, the heat that is dissapated by the heatsink will increase. This aggrivates the other problems that limit the chip but in and of itself isn't a limiting factor, since chips stop working well before they turn into molten blobs of silicon. If there were no other factors, then upping the voltage would just increase the heat output and make it stop working sooner, but we all know that voltage increases can make things more stable at higher clock speeds so this isn't the only factor.

If I'm remembering my class on chip design correctly, however, the main limit on chips is the rise time of the transistor. Hmm, I just realized I don't have my textbook handy, so I can't give an exact description. From memory: As you increase the frequency, the voltage needs to rise from a 0 to a 1 fast enough that it is stable before the next clock tick hits. We like to model transistors as simple 0s and 1s, but there is also a "who the heck knows" state in the middle where the output isn't certain to have switched states. So if the voltage doesn't rise fast enough to be definitely a 0 or a 1 before the next tick, you start getting errors in your calculations and it doesn't take long before this corrupts whatever was attempting to run on an unreliable circuit (which usually means your OS does something like find out that fase really equals true and commits suicide). Increasing the voltage makes the rise snd fall times faster, which means you can get past this problematical state quicker, which means the chip is more reliable at higher frequencies. However, rise time is not just a property of voltage it also involves the resistance of the wires and the silicon, so as the heat increases the resistance increases and the rise time starts slowing again.

So... increasing the clock frequency makes increased heat which increases the resistance of the silicon which slows the rise time which makes some of the signals not complete their 0 to 1 or 1 to 0 transitions in time. Raising the voltage quickens the rise time which might allow you to raise the clock frequency higher, but that just starts the cycle again. Giving it better cooling will lower the resistance of the silicon which quickens the rise time which makes the signals much more reliable and lets you clock it higher without increasing the voltage.

Thats all from memory of a class 3 years ago, but I think its at least in the ballpark.

 

[zephyrprime]

Originally posted by: Gioron
So... increasing the clock frequency makes increased heat which increases the resistance of the silicon which slows the rise time which makes some of the signals not complete their 0 to 1 or 1 to 0 transitions in time. Raising the voltage quickens the rise time which might allow you to raise the clock frequency higher, but that just starts the cycle again. Giving it better cooling will lower the resistance of the silicon which quickens the rise time which makes the signals much more reliable and lets you clock it higher without increasing the voltage.

Thats all from memory of a class 3 years ago, but I think its at least in the ballpark.

Sounds nice but resistence decreases in a semiconductor as temperature increases. Text
And besides, isn't a transistor's rise time is more due to the speed of changing electrical fields and the mobility of electrons than it is to simple resistance?

Isn't the reason that heat limits an overclock simply because random voltage fluctuations increase with increasing temperature which eventually allows for electrons to flow across the transisitor's channel even if the gate is off?

 

[kcthomas]

Originally posted by: TuxDave

Originally posted by: kcthomas

Originally posted by: TuxDave
the core will break down.

is this refering to dielectric breakdown? The oxide between the gate and the channel prevents electrons from passing through. i think as the temperature rises, dielectric breakdown occurs which is where the oxide which is usually suppose to block electrons, allows electrons to pass. I dont know if this screws up the transistor permanantly though.

If the oxide breaks down, a transistor wouldn't function as it should anymore. Current would be tunneling through the gate oxide at such large levels that most dynamic and perhaps some static gates would never function.

if dielectric breakdown occurs is it permanent or just temporary? does it physically change the oxide so that is not an oxide any more?

 

[TuxDave]

Originally posted by: kcthomas

Originally posted by: TuxDave

Originally posted by: kcthomas

Originally posted by: TuxDave
the core will break down.

is this refering to dielectric breakdown? The oxide between the gate and the channel prevents electrons from passing through. i think as the temperature rises, dielectric breakdown occurs which is where the oxide which is usually suppose to block electrons, allows electrons to pass. I dont know if this screws up the transistor permanantly though.

If the oxide breaks down, a transistor wouldn't function as it should anymore. Current would be tunneling through the gate oxide at such large levels that most dynamic and perhaps some static gates would never function.

if dielectric breakdown occurs is it permanent or just temporary? does it physically change the oxide so that is not an oxide any more?

If I recall correctly, we categorize breakdown as soft and hard breakdowns. If the dielectric has a hard breakdown meaning that the oxide has been severely overstressed via temperature or voltage, then the transistor is basically gone since the oxide no longer retains its insulating behavior. If you overdrive the transistor slightly or overclock it by raising the core voltage, the transistor will exhibit a soft breakdown meaning that it will function but the increased stress will degrade the transistor performance at a faster rate.

 

[TuxDave]

Originally posted by: zephyrprime

Originally posted by: Gioron
So... increasing the clock frequency makes increased heat which increases the resistance of the silicon which slows the rise time which makes some of the signals not complete their 0 to 1 or 1 to 0 transitions in time. Raising the voltage quickens the rise time which might allow you to raise the clock frequency higher, but that just starts the cycle again. Giving it better cooling will lower the resistance of the silicon which quickens the rise time which makes the signals much more reliable and lets you clock it higher without increasing the voltage.

Thats all from memory of a class 3 years ago, but I think its at least in the ballpark.

Sounds nice but resistence decreases in a semiconductor as temperature increases. Text
And besides, isn't a transistor's rise time is more due to the speed of changing electrical fields and the mobility of electrons than it is to simple resistance?

Isn't the reason that heat limits an overclock simply because random voltage fluctuations increase with increasing temperature which eventually allows for electrons to flow across the transisitor's channel even if the gate is off?

From a circuit perspective, a transistor's rise time is influenced by

1) The load the transistor will drive (smaller --> faster)
2) The voltage supply of the circuit (higher --> faster)
3) The transistor which has a higher ft (smaller --> faster)

I really don't know or haven't heard of resistance being a prime limit to our circuits today. The transistor, its operating voltage and the wire capacitance in the end limit how fast we can theoretically go. The power usage determines how fast we can realistically go.

Overclocking leads to increased heat because we're both increasing its clock frequency and voltage to get that extra performance out of the processor. It could fail because simply you can't get the transistors to go that fast or it could fail because you exceed some operating temperature which will lead to soft or hard breakdowns involving the gate oxide or reverse junction currents or simply the wires will melt away via electro migration.

 

[Gioron]

Ok, you're right about the resistance of silicon, I must've remembered that part of the explanation wrong. I'm still fairly sure the explanation I saw before stated that it was a rise time problem, but I'm no longer sure what the temperature related part of the equation was.

 

[tart666]

Originally posted by: zephyrprime

Originally posted by: Gioron
So... increasing the clock frequency makes increased heat which increases the resistance of the silicon which slows the rise time which makes some of the signals not complete their 0 to 1 or 1 to 0 transitions in time. Raising the voltage quickens the rise time which might allow you to raise the clock frequency higher, but that just starts the cycle again. Giving it better cooling will lower the resistance of the silicon which quickens the rise time which makes the signals much more reliable and lets you clock it higher without increasing the voltage.

Thats all from memory of a class 3 years ago, but I think its at least in the ballpark.

Sounds nice but resistence decreases in a semiconductor as temperature increases. Text
[..]

This also sounds nice, but is only true for a semiconductor with no electric field applied. With a CMOS transistor in the ON state, the channel acts more like a regular conductor and increases its resistance with temperature. OTOH, all the channels in the OFF state, do reduce their resistance and increase their already not incosiderable leakage. Adding all of this up, the difference betwee, ON and OFF gets muddled, and numerous soft errors occur.

 

[TuxDave]

Originally posted by: Gioron
Ok, you're right about the resistance of silicon, I must've remembered that part of the explanation wrong. I'm still fairly sure the explanation I saw before stated that it was a rise time problem, but I'm no longer sure what the temperature related part of the equation was.

You are right in that regard. If the fastest rise time/switching time/delay of the transistors is not fast enough, then you can't run the processor much faster. So yeah, that is another constraint in a CPU clock frequency and that's why even if you cool the whole thing waaaaay down, you still won't get infinite clock frequencies.

 

[Gioron]

Originally posted by: TuxDave

Originally posted by: Gioron
Ok, you're right about the resistance of silicon, I must've remembered that part of the explanation wrong. I'm still fairly sure the explanation I saw before stated that it was a rise time problem, but I'm no longer sure what the temperature related part of the equation was.

You are right in that regard. If the fastest rise time/switching time/delay of the transistors is not fast enough, then you can't run the processor much faster. So yeah, that is another constraint in a CPU clock frequency and that's why even if you cool the whole thing waaaaay down, you still won't get infinite clock frequencies.

Yah, but I'd swear that my teacher said something about rise times getting slower as heat increased. Too bad I can't find the notes or the textbook for that course.

 

[kcthomas]

Originally posted by: Gioron

Originally posted by: TuxDave

Originally posted by: Gioron
Ok, you're right about the resistance of silicon, I must've remembered that part of the explanation wrong. I'm still fairly sure the explanation I saw before stated that it was a rise time problem, but I'm no longer sure what the temperature related part of the equation was.

You are right in that regard. If the fastest rise time/switching time/delay of the transistors is not fast enough, then you can't run the processor much faster. So yeah, that is another constraint in a CPU clock frequency and that's why even if you cool the whole thing waaaaay down, you still won't get infinite clock frequencies.

Yah, but I'd swear that my teacher said something about rise times getting slower as heat increased. Too bad I can't find the notes or the textbook for that course.

as some of you have said, if it is true that the resistance increases with temperature, then the rise time will also increase. The time it takes for a a capacitor to charge/discharge is Capacitance*Resistance the capacitor sees.

 

[BEL6772]

I think the effects of heat on speed have been pretty well discussed. One aspect of the original question was how heat damages the chip. There are several bad things that can happen to a chip that gets too hot.
The chip is attached to its package with a solder, which can melt. There are also discrete components soldered onto most modern processor packages that can become de-soldered. Temperatures required for this type of failure are protected against by thermal cut-off circuits, so the odds of experiencing this are fairly remote for modern systems.

Another thing that happens is a little less dramatic. Part of what makes a piece of silicon work is the impurities (ions) implanted into it. These are 'placed' at fairly exact concentrations and locations in order to allow us control over the flow of electrons. The natural tendency of these concentrated ions is to diffuse through the silicon crystal structure to achieve a state of lower concentration. The more heat there is in the crystal, the more mobile these ions are. As these ions drift, transistor performance is degraded. If the heat level is moderately out-of-spec, you can take years off the life of a processor because of ion drift. In fact, one of the ways IC manufacturers simulate a lifetime of use is by running their chips at elevated temperatures.

 

[white]

Originally posted by: Antoneo
When people say, cooling solutions prevent damage to the chip, what type of damage are we talking about? What do fried cores denote, are the gates inside being destroyed?

I've been wondering because a lot of times, a better cooler (allows more heat to be transferred off the CPU) would allow higher frequenices, just why is this true? Does heat have some effect on electron flow? Is it a resistor effect?

Semiconductors have the unique feature that allows us to control the flow of electrons. At different temperature ranges they behave differently. At low temperatures dopants are frozen out and are unable to take part in conduction. At high temperatures we lose control of the flow of electrons and it results in an always open transistor.

I'm not sure what is technically a "fried" core. This has to be an irreversible process so it can't be due to the change in material properties at elevated temperatures that change back at lowered temperatures. It would have to be a permanent change, perhaps electromigration somewhere in the package. We've observed fried cores when we forget to put a HSF on or if it's improperly mounted. It generally doesn't take long until the user smells the smoke. It isn't silicon burning because that didn't melt. Neither did the oxide. I could envision various parts of the package melting such as the underfill material (some sort of polymer) or the package substrate [FR4, BT (fire resistant 4, bismaleimide triazine)], although this isn't supposed to catch fire. I guess if the temperature went up enough, it could start to flow yet not catch fire, and this would be enough to cause a permanent irreversible change that would cause the CPU to stop working.

As for the solder, typical lead-tin solder melts at 183°C, so if we accidentally operated a CPU that hot then the solder balls in the package would start melting. Typical lead-free solders would melt a little higher than that. In terms of thermal stress, it would actually decrease as we approached the melting temperature since there is more thermal stress further away from the melting temp.

While this is a consideration at such extreme temperatures, I think the first cause of failure would be loss of controllability since it'll act like an intrinsic semiconductor. Conductivity is proportional to the number of carriers and the mobility. At higher temperatures, hole and electron mobilities decrease due to lattice scattering, but the increase in number of carriers increases exponentially and dominates resulting in an increase in conductivity. Eventually it gets to the point where the semiconductor behaves like a conductor. We could never make our CPU's out of copper or aluminum or any other metal since we can't control its conduction. So it's the change in behavior of the silicon that is the first step that causes the CPU to cease operation. But it is some other permanent change that will result in a fried CPU.