Why Intel CPU's run at 95°C and why AMD's should, also

[mtcn77]

The original study is from AMD's forum and it is a response to a hung up query.
Noctua, AMD and GN's technology journalists have attended to the question of why AMD runs hot. I'm going to prove it on a third party case study.
You can access individual company takes on the question in the hereby links:

1. Noctua:
2. AMD@reddit:
3. GamersNexus:

You can go look at individual threads to go into the specifics, but the crux of the matter is that cpu heat density is limited from sufficient heat conduction by the thermal paste performance.

• AMD Community Forum

Spoiler: "AMD Community Forum" thread post

After applying metal liquid between the cooler and IHS things improved by almost 20°C in the range of 65-80°, that was when the CPU draws from 80 to 120watts, (80° became 60 at 120watts!) but improved only by 3°C when the wattage jumps over 150watts, (90° instead of 93-94°), also frequencies improved accordingly.

• Under 55°C it sits at 4350MHz all cores and occasionally some cores jump to 4.4
• @65°C it draws 80Watts and sits at 4150MHz all cores
• @75°C it draws 90Watts and sits at 4100MHz all cores
• @85°C it draws 100-110Watts and sits at 4050MHz all cores
• @90°C it draws 150-160Watts and sits at 3980MHz all cores
• @94°C+ almost 170watts, throttling rules and it sits at 3900MHz all cores.

Our envoy takes us into the mystical journey of why 250w-spec tower air cooler is necessitated to run at high heat in order to cool the cpu properly. Gamersnexus is the go to guide here. As ironic as it is, how TDP is calculated is totally different to what we have been used to;

Spoiler: AMD's take to GN:

AMD defines HSF θca (°C/W) as: The minimum °C per Watt rating of the heatsink to achieve rated performance.

Its internal definition, for comparison, says “the minimum required heatsink resistance necessary to maintain the case temperature within specification for the thermal design power (TDP) and assumptions for the external ambient temperature and system temperature rise (Tsys).”

“Theta CA”

The HSF is what stands between the CPU and the surrounding air, and θca is the thermal resistance between the CPU and the air, so HSF θca is the thermal resistance of the heatsink. Lower is actually better here, not higher, so AMD’s phrasing has some interpretive gray areas. AMD’s reviewer document should instead read “maximum” instead of “minimum,” so it should be the, quote, “maximum *C per Watt rating of the heatsink,” as lower is better and so maximum would be the last value permissible for rated performance before becoming insufficient for rated performance. When we reached out, AMD clarified that “you can interpret the original copy to mean ‘the [minimum standard]’ where lower values produce superior results.”

As it stands, the under recognized limit to cooling performance, the heat conductance through the IHS, is the dominant overclocking performance determinant.
AMD keeps reference to '°C/W', thermal resistivity, while 'W/°C'(thermal conductivity) - how much temperature gradient you can maintain between the ihs and heatsink plates - maintains how much heat capacity you can specify for the TDP of their cpus. This is inverse to what AMD has been saying, you get improved cooling by first maintaining a high gradient, not a lower temperature.

Spoiler: Noctua's response@reddit:

From Noctua:

Due to the small size of the CPU-die, the heat density (W/mm²) of this chip is very high. For example, a 120W heatload at a chip-size of 74mm² results in a heat-density of 1.62W/mm², whereas the same heatload on an older Ryzen processor with a chip-size of 212mm² gives a heat-density of just 0.57W/mm².

I'm pretty sure AMD did enough stress test to determine the CPUs are going to be fine.

TL;DR: It's just harder to cool because the heat builds up so fast, the heat spreader can't extract the heat from the die fast enough.

7nm dense architectures can only extract enough heat from the cpu die integrated heatsink, once temperature reaches safety limits.
The difference of liquid metal is, it provides the best conductance to the heatsink plate under normal operating temperatures.
Nonconductive low tier pastes can only operate within the same conductance at the highest TjMax. The cpu is making the best of available thermal gradient when liquid metal is applied, jumping from 80w>120w while cooling below 80°C>65°C.
Since AMD, GN, or Noctua hasn't mentioned in anyway how the Zen 2 architecture would be served best either with nickel plated air coolers that work with liquid metal in stock settings, or aio coolers that work cooler than 90°C needs the same liquid metal application, I think this requires a reevaluation. We don't see any observable difference to overclocking these cpus under normal conditions because the ihs is the conduction limit and the sole solution being upgrading to liquid metal interface materials.

 

[Deleted member 185158]

you get improved cooling by first maintaining a high gradient, not a lower temperature.

It eliminates big "spikes" in temps. That's a given.
The issue is, people rely on the SenseMi quite a bit and run higher voltages than normal.

Ideally these chips should be run 70c or less, not 70c or more with high gradient. The cpu High Temp Alert signal to the board is 70c. You konw this while running a stock system and when the temp reaches this figure, the cpu fan is at 100%.
The TDP figure that's advertised is for max P-state load. Not max turbo load. That's why we see 160w pulls with PBO (example wattage) Also, the TDP figure can have a 5% swing.
Also, running almost any x86 architecture hotter causes more leakage, thus needs more voltage. Up and Up we go in temps. Such a bad way to "cool" a processor.

Plate to Plate contact.....

The surfaces are roughly flat. Not exactly lapped flat. So having a decent paste and using as little as possible would be a great way to go. This helps heat transfer evenly across the surfaces while most IHS plates are slightly concave in the middle. Any one that's ever lapped any AMD IHS Plate know this first hand while the outer edges, the nickle plating is removed first.
And speaking of Nickle plating, it's very poor in conductivity compared to Copper. This is another reason lapping is a good thing, we don't see it much, people are just lazy.

Nobody thinks of cold plate size and mass. I've done some testing in this area with lidded and de-lidded Zen +. I know this isn't Zen 2, but it's the cold plate concept I want to address, not the cpu type as to keep the talk strictly about cooling, not the Cpu itself.
Most cold plates on AIO and tower coolers are very small and thin. This is not a good approach if we are trying to achieve this "high gradient" while you don't have enough mass to maintain a good temperature. This heat is stored before removal. It's not an instant action. You only see the CORE TEMP or perhaps the motherboard temp. But I don't see people running around probing the heatsink plates..... This is an issue. How do expect to maintain a high gradient if you can't or don't monitor your cold plate temps?? It just make any sense to talk about having a higher temp gradient with most under par cooler.

Now back to what I had quoted....

Not a lower temperature is interesting. I found a lower temperature reduced leakage by a very considerable amount.
For example, Try running 4Ghz(static) under 1.2v @ 90c, It's just not gonna happen. You want to be closer to 10c for that. Not 90c.

Efficiency gets lost in heat and high voltage. I do understand the concept of reducing temp swings and spike having a higher temp gradient, but 90c is just too F'n hot. Address your cooling issues.....

 

[mtcn77]

What I bring the plot to is, as conduction is improving as the cpu warms up, you can transmit more heat and actually improve cooling by the consequentially higher thermal transmittance.
There is a pttl setting, but my guess is, it isn't connected(just runs at base clock) to pb2.
If it is, then great - AMD=Intel. However, if not, that means you cannot overclock through sense m.i. pb2-xfr without liquid metal interface.
What else, if it is, then there is the problem of undervolting at high temperature - we know for sure sense m.i. don't apply a poole & frenkel rectifying undervolt at the same speed bins as temperature elevates. Then, you need liquid metal for sure, otherwise the cpu won't overclock past 80w due to temperatures issues. Eventhough it is safe, it can work at 2x TDP by a simple trick, thus AMD should provide necessary sense m.i. directions at such high TjMax by tapering voltage and temperature parameters not to burn the cpu while keeping TIM practical life in use, running within inches of liquid metal's performance frame.

If it can output the same high heat at a slightly lower voltage while overclocking which we know is possible looking at undervolt study results of ryzen that are wildly different than pb2 voltage bin selections between 1.325v-1.4v, its temperature maintenance will prove itself by the gradient's now higher role in the same tdp equation. Personally, to pull that off at 90°C, someone needs to study how much undervolting offset is present at stock. I'm not certain how the sense m.i. software can be made aware of transmittance(we know they use an exponential moving average of °C in order to keep a steady tdp profile), but if it includes 'instantaneous power use in comparison to gradient temperature' that will give a close cut view of how much headroom there is.

 

[Deleted member 185158]

Stability comes with cooler temps, not hotter temps. In this aspect, AMD = Intel.

This is an example of what I was talking about earlier with running Cold vs Hot and leakage.

Do Note the Cpu v-core is 1.188v @ 4ghz , 300mhz past base clock and static clocked to the stock max all core boost state for SenseMi. (actually 1/4 multi higher)

So From this experience, It is very, super, really really, super duper to the 10th power, Difficult to agree with running a cpu hotter vs colder.

..... I agree to disagree while I totally understand your direction and point of view.... however I have evidence that it's not a good way to run a cpu.

 

[eidairaman1]

Stability comes with cooler temps, not hotter temps. In this aspect, AMD = Intel.

This is an example of what I was talking about earlier with running Cold vs Hot and leakage.

Do Note the Cpu v-core is 1.188v @ 4ghz , 300mhz past base clock and static clocked to the stock max all core boost state for SenseMi. (actually 1/4 multi higher)

So From this experience, It is very, super, really really, super duper to the 10th power, Difficult to agree with running a cpu hotter vs colder.

..... I agree to disagree while I totally understand your direction and point of view.... however I have evidence that it's not a good way to run a cpu.

View attachment 143781

The A6-3650 I just cleaned recently did not linke running at 78C, max for it is 72C.

 

[Deleted member 185158]

The A6-3650 I just cleaned recently did not linke running at 78C, max for it is 72C.

Luckily the Ryzen chips have a much higher max operating temp. But 90c is throttle for Zen/+ and 95c for Zen 2. Basically the no go zone. Gotta have headroom for ambient temp swings.

 

[eidairaman1]

My 8350 could handle 75 without shutting down so idk

 

[Deleted member 185158]

My 8350 could handle 75 without shutting down so idk

Thermtrip for FX 8350 is 110c. Max operating temp is a throttle point.
If the board shuts down... It may or may not have just saved the chip.

My 8350 could handle 75 without shutting down so idk

Also I had an Opteron 165 that would handle 80 to 90c. It did degrade fast though as I was running is 3.2ghz.

 

[Vya Domus]

Heat rarely kills a CPU under normal circumstances if ever, by that I mean no overclocking record attempts.

It's the voltage that gets it, more specifically the voltages that have to be applied when running at higher clocks, that's the real reason degradation occurs. You can run a stock CPU right at it's temperature shutdown limit 24/7 and it'll likely be fine even after decades. Run in at plus 30% more voltage, it probably wont be.

 

[mtcn77]

..... I agree to disagree while I totally understand your direction and point of view.... however I have evidence that it's not a good way to run a cpu.

View attachment 143781

I cannot just insert another paragraph. The gradient plays a more pronounced factor here than stated in AMD's thermal design point calculation. AMD measures resistance as a die cast value. It is not. Let us see what it is;

Spoiler: "GamersNexus" are the technology journalists

AMD expands and says the following of its formula: “The TDP formula is straightforward: TDP (Watts) = (tCase°C - tAmbient°C)/(HSF θca)”

AMD says this in its guide: Using the established TDP formula, we can compute an example in the form of the 105W AMD Ryzen™ 9 3900X: (61.8-42)/0.189 = 104.76 TDP, [with] tCase°C [as] 61.8°C optimal temperature for processor lid.”

This is the thermal resistance series formula. What is correlated is the gradient and what is inversely related is the resistances of individual plate materials. If we increase the gradient, the resistance drops by default instantaneously;
Thanks, 'thermtest'

Spoiler: Thermtest to the rescue

The heat flow, or boundary temperatures of the system, can also be determined when an object’s resistance is known. In series, the heat flux through a composite material is considered constant, and the different series are equivalent to:

R = R_{1} + R_{2}R=R1+R2

When the temperatures on each side of the composite material are known (T_{L}TLand T_{R}TR), the heat transfer rate is expressed as:

\dot{Q}=\frac{T_{L} {-} T_{R}}{R} = \frac{T_{L} {-} T_{R}}{R_{1} {+} R_{2}}Q˙=RTL−TR=R1+R2TL−TR

All this mumbo jumbo just means, you can keep resistances, add them and when you put their sum in the denominator, it still is outweighed by the gradient difference at the nominator field of the fraction.
What we can do differently is keep an artificially high gradient, using less heat in the process, playing around with the voltage. True dat, I still don't understand what you mean by your cinebench postie, shrimpbrime, but consider this;
The solid piece of copper slab can only conduct 400W/m*°C, TIMs are 10, liquid metals are 100, so TIM on copper is 0.1025 m*°C/W while liquid metal on copper is 0.0125 m*°C/W and unless you keep the gradient steady, maximum per ambient, the cooler just won't get passively hot enough until air convection takes it away. I recon, once heat generation stops, a hot cooler will cool its coldplate faster than its coldplate could heat it. So, it will be more heat & voltage stable as the cooler warms up in the artificial heat gradient stage.

 

[eidairaman1]

Mtcn77 wheres your system specs?

 

[mtcn77]

I'm not using my system as of this moment. I have a spare 2400g setup that I don't use.

I had it all filled in over at OCN.

 

[eidairaman1]

This is not ocn though

 

[mtcn77]

There is a similar perk mentioned over at here: https://physics.stackexchange.com/q...tivity-provide-a-smaller-temperature-gradient
Let's go back to Noctua's fly by;

Spoiler: Noctua@reddit:

From Noctua:

Due to the small size of the CPU-die, the heat density (W/mm²) of this chip is very high. For example, a 120W heatload at a chip-size of 74mm² results in a heat-density of 1.62W/mm², whereas the same heatload on an older Ryzen processor with a chip-size of 212mm² gives a heat-density of just 0.57W/mm².

There is a constant reference of cpu die area which brings us into crossroads with thermal transmittance.

Spoiler: "Thermal transmittance":

Φ = A × U × (T1 - T2)

• A is area m².
• U is transmittance W/m²*°C.
• (T1 - T2) is the gradient.
• Φ is watt as in heat transfer.

You see, the only changes to the ihs is always the same. Temperature gradient between the plates play a huge role in this.

This is not ocn though

And I'm not banned here for being pushed around by nvidia loyalist moderators.

 

[Bones]

Heat rarely kills a CPU under normal circumstances if ever, by that I mean no overclocking record attempts.

You are correct here, under most circumstances as described it's not an issue to worry about and I seriously doubt any of us would run a chip for decades before upgrading anyway making it a non-issue.

It's the voltage that gets it, more specifically the voltages that have to be applied when running at higher clocks, that's the real reason degradation occurs. You can run a stock CPU right at it's temperature shutdown limit 24/7 and it'll likely be fine even after decades. Run in at plus 30% more voltage, it probably wont be.

I must disagree with this.
The defined thermal limit for a chip is what it can tolerate and still live without dying on the spot in short order, does not mean it will not be affected and this thermal limit was never meant to "Say" it's always 100% OK at that temp sustained.

Heat has always been an enemy of electronics period, it's the nature of it and while it isn't required to cool one down to subzero temps throughout it's useable life it's better to keep it away from it's defined thermal limit if possible.

An aircraft is one example of this.
Although during it's operational life it may never actually be stressed to it's defined limits it will eventually suffer from material fatigue even though it had never endured stresses at or around 100% of it's rating each and everytime it flew for 100% of the time it was up in the air.

Same thing if you take a small piece of wire and start bending it over and over again, it will eventually break and this can be felt as you keep bending it - Becomes easier to bend as you go until it just breaks in two.
Even though you didn't place enough stress to break it right from the start (Which would have been at or exceeding 100% of it's strength) it was enough over time to affect it and like all things it will suffer from the effects of this continued stressing.
Heat can and does "Work" on a chip too except it's not going to literally break in half, instead it's functionality is affected over time and like everything else it will fail one day but in this case failure will occur sooner.

In short:
Running it at the limit all the time, while doable if you want just isn't a good ideal.

 

[mtcn77]

Heat has always been an enemy of electronics period, it's the nature of it

Yes, we call it poole frenkel effect. As it heats up, it needs less charge for that band conduction. A seperate voltage profile can work around that, I assume.
I don't even want AMD to change their exponential moving average temperature algorithm. It is just that, lower thermal resistance of the heatsink at a higher temperature will instantly catapult their tdp designation by 2x higher budget. Rather than calculate its return to the mean temperature, instead its return to gradient temperature in response to current heat(make shift transmittance formula) that would switch its cooler estimation by a much clearer approximation. Sense M.I. can be made cooler aware if it variates voltage in response to keep the same thermal transmittance between the cpu and the cooler plate. If it gets hot, granted, you are dealing with a potential hot aio water cooler situation waiting it to cool down for a long time, however our intention was never to expect it to cool the cpu faster than the cooler reached steady state.
Poole frenkel here works by lowering voltage requirements as a safety net against electromigration and sense m.i. crossects its highest bin with what poole frenkel voltage limit at that temperature is.

OK, you don't have to increase voltage to heat up; you can introduce hysteresis, too, delaying the onset of cooler's acceleration slope. That should be safer and not require seperate voltage bins.

 

[EarthDog]

must disagree with this.
The defined thermal limit for a chip is what it can tolerate and still live without dying on the spot in short order, does not mean it will not be affected and this thermal limit was never meant to "Say" it's always 100% OK at that temp sustained.

Cpus have two thresholds... where it throttles and where it shuts down. So long as the cpu isnt throttling to protect itself that means intel/amd are ok with those temperatures. If you are throttling and hit thermal shutdown, that is too hot. In our overclocking endeavors over the years, we found occasionally some chips can lose stability a bit before that. This is why we usually say, for Intel chips with 100C throttling point (105C+ shutdown) to keep it under 90c. For headroom under throttling point. I'd gladly pound and have pounded chips in this manner.

While i understand the underlying point of your wire analogy, metal fatigue by mechanical movement and temperature with silicon is quite different.

 

[Vya Domus]

I must disagree with this.
The defined thermal limit for a chip is what it can tolerate and still live without dying on the spot in short order, does not mean it will not be affected and this thermal limit was never meant to "Say" it's always 100% OK at that temp sustained.

Heat has always been an enemy of electronics period, it's the nature of it and while it isn't required to cool one down to subzero temps throughout it's useable life it's better to keep it away from it's defined thermal limit if possible.

The only real way a CPU degrades is through electromigration, that process is accelerated by larger voltage drops across conductors. Hence when you overclock a CPU and apply higher voltages that will be by far the main reason it degrades over time, incidentally when that happens the heat output also increases, but that's the side effect. The thermal limit on processors is imposed more for stability purposes rather than fear of damage. I mean think about, if it's supposed to be the maximum temperature after which the chip dies on the spot then that means it really should die once it hits that temperature.

But they don't, think of the plethora of laptops with crappy cooling that cause shutdowns or run pretty much right at the limit of thermal shutdown, despite that the chips pretty much never die. It's exceedingly difficult to kill a chip, you either have to apply insane voltages to it or disable all safeties and run it without a cooler. And you need really, really high temperatures to damage it beyond working order. In other words it's unrealistic to expect that something like this would happen.

Voltage and temperature can both kill semiconductors but in different, slower or faster ways. You're probably thinking of electronics in general where temperature can lead to damage in a much more nuanced and obvious way across time but with IC's due to the way they are used it doesn't really occur in the same way.

 

[Zach_01]

Cpus have two thresholds... where it throttles and where it shuts down. So long as the cpu isnt throttling to protect itself that means intel/amd are ok with those temperatures. If you are throttling and hit thermal shutdown, that is too hot.

The only real way a CPU degrades is through electromigration, that process is accelerated by larger voltage drops across conductors. Hence when you overclock a CPU and apply higher voltages that will be by far the main reason it degrades over time. The thermal limit on processors is imposed more for stability purposes rather than fear of damage. I mean think about, if it's supposed to be the maximum temperature after which the chip dies on the spot then that means it really should die once it hits that temperature.

But they don't, think of the plethora of laptops with crappy cooling that cause shutdowns or run pretty much right at the limit of thermal shutdown, despite that the chips pretty much never die. It's exceedingly difficult to kill a chip, you either have to apply insane voltages to it or disable all safeties and run it without a cooler, that's pretty much the only way you can kill it with high temperatures.

As ZEN2 for all I understand, Its ok to operate close/under throttling limit of 95C for long periods of time, BUT when in stock settings and not OCed. Only then SenseMI or silicon FITness controller can adjust clock and voltage according to temp to preserve silicon longevity and avoid electromigration. Because even tho throttle begins at 95C the cut down of boosting and voltage starts way back. Running CPU (all core loads) with temps 90C and 60C would have a difference of around 250~300MHz and difference for voltage too.

------------------------
On OP topic

While higher CPU temps keep temp gradient higher, hence higher heat transfer, I dont think its wise to deliberately keep CPU temps high to achieve high heat transfer. The further cooling down of cooler's or block's cold plate is the right way to do it and when this hits the limit, of the given system (air, water, water chiller... etc) the only other wise choice is to reduce thermal resistance between CPU and cooler if and when its possible...
...hence TIM.

 

[mtcn77]

Graphene nanoribbons illuminate at 2500°C if I recall correct. I think we will have a different class of categorization for these nanotechnology devices that work at their fundamental limits.

 

[EarthDog]

I dont think its wise to deliberately keep CPU temps high to achieve high heat transfer.

Agreed. This idea seems quite off to me. I'm running it as cool as things allow for my settings. No way would way should we run warmer for better heat transfer.

Everything equals out. Someone earlier mentioned block thickness or something helps... and while that is true, a larger copper block will hold more heat than a smaller, one, the larger one gets saturated as well and then its down to the metal properties again. This is akin to adding more water to a water loop thinking it will be cooler. Once things hit the saturation point, it is what it is. It doesn't change the temps, just longer to reach the equilibrium.

The problem with today's processors is the tiny die and density trying to get the heat out.

 

[moproblems99]

This is another reason lapping is a good thing, we don't see it much, people are just lazy.

Keep in mind that some plates (blocks, etc) are purposefully convex because the tightening process brings them flat. Lap the block and you have just done f'ed that up.

 

[Deleted member 185158]

There is a similar perk mentioned over at here: https://physics.stackexchange.com/q...tivity-provide-a-smaller-temperature-gradient
Let's go back to Noctua's fly by;
There is a constant reference of cpu die area which brings us into crossroads with thermal transmittance.
You see, the only changes to the ihs is always the same. Temperature gradient between the plates play a huge role in this.


And I'm not banned here for being pushed around by nvidia loyalist moderators.

The cpu die area is split into pieces on Ryzen 3000 chips. Zen+ and previous only had a single die.
Which die area is Noctua referencing? the I/O chip?

I liked the point you brought up earlier about Tim 10 Wm2K vs LM 100 Wm2K thermal transmittance.

You'd think the solder I removed and replaced with TIM would had made a HUGE difference. Nah, not like you think.

I'm running a De-lidded 2700x, solder replaced TIM with the stock cooler and it changed the effects of the processor none.

It's not the IHS plate, or the TIM / LM..... It's the density of the transistor count on small die. This is where the "temp spikes" come from.

 

[mtcn77]

I've also given up on the voltage modulation idea. Currently, it is heat vs sense m.i. cooler fan modulation. If there has ever been any doubt, we never advise testing above silicon FITness thresholds during any time. I just need some more clarity on how sense m.i. can pace cooler speed with heat buildup from the negative inverse plane, triggering just when cpu temperature gradient is about to hit throttle perimetry.
Do we go by cooler steadystate measurements? There has to be a linear estimation to fit this into a binary domain to wreak maximum havoc to case ambients. Not safe for atx cases with psus on top.

 

[Deleted member 185158]

I've also given up on the voltage modulation idea. Currently, it is heat vs sense m.i. cooler fan modulation. If there has ever been any doubt, we never advise testing above silicon FITness thresholds during any time. I just need some more clarity on how sense m.i. can pace cooler speed with heat buildup from the negative inverse plane, triggering just when cpu temperature gradient is about to hit throttle perimetry.
Do we go by cooler steadystate measurements? There has to be a linear estimation to fit this into a binary domain to wreak maximun havoc in case ambients. Not safe for atx cases with psus on top.

Start with the Max wattage draw and convert that into BTU first.

Example, 160w is 545 BTU/hr. This should help your mathematics during thermal resistance and displacement.

Not all the cpu heat is dissipated to the heat sink. A great deal of it yes, but some is dissipated through the board as well. (this needs to be accounted for.)