[Graphs] What Happens Around 3.3 GHz That Sends Temperatures Skyrocketing?

[know of fence]

We have seen overclocks hit a wall at around 5 GHz, but really it's just the tail end of rather extreme cooling solutions. I was surprised to find out that how we get there isn't just a steady climb. Meanwhile Intel's ultra mobile devices turbo up to around 3 GHz which, as it turns out, may not be just an arbitrary number.

So I probed and recorded the full voltage range of my 22nm Haswell-Refresh Pentium Anniversary Edition G3258. But really it is representative of all Haswell CPUs, just move it around a little or bend it a bit if you are on a stock cooler.

At first glance this is just a steady increase, meaning ramping up the cpu frequency in 100 MHz steps always requires Vcore to be raised. But a closer look reveals that there are two pieces to this graph, one that is almost completely linear and a second one that visibly curves upwards.
I tried to highlight the point where the curvatures kicks off, but it's kind of hard to pinpoint in this representation.

So to make this threshold more pronounced I plotted the voltage difference between two multipliers (AKA differetial or delta). Unfortunately this also makes measurement inaccuracies more pronounced. Also added some temperature (Tmax) data.

Because it is an numeric approximation of a derivation of the original voltage curve, linear parts become flat and parabolic-ally curved parts become linear.

More importantly though, the two pieces of the voltage delta show very different trends. Stock frequencies up to 3.2 GHz show that it takes on average ~ 17 mV to reach the next step. This voltage basically compensates for increased frequency, as temperatures only slowly creep upwards. Beyond 3.3 GHz however voltage jumps increase, to the point where most of the voltage serves to stabilize and compensate for the now much higher and faster climibing temps.

CPUs really hit a wall around 3.3 GHz not 5 GHz, about where the two trend lines intersect. What happens then is that desktop makers keep bashing their heads against this thermal wall with increasingly ludicrous cooling solutions and diminishing returns that allow for a rather unimpressive ~1.5 GHz of headway!

What I like to know is: What do you think happens inside the CPU around 3.3 GHz that sets off the sudden rise of temps?
Does this threshold or wall move when going from one node process to the next? If it does, which way did it shift for 14nm Skylake?

 

[sm625]

To me it seems like 3.8GHz is the point where it really breaks out and pulls what the greenies dont like to call a hockey stick.

As for why it happens, I believe it happens because the square wave that you envision occuring inside the transistors actually turns into a sine wave at higher frequencies. At a certain frequency, the width of the bottom of the square wave, which represents the time where the transistor is completely off, is reduced to zero. So the period where the transistor is consuming zero power is reduced to almost nothing, resulting in an increase in power in a linear fashion. As long as even a tiny portion of the cycle results in the transistor spending some time in the off state, the function is linear. But when you go past that critical frequency, the transistor is never truly off. At this point you need higher voltage to make the sine wave larger which lets the transistor actually hit its off state. This function is nonlinear because the voltage is also increasing. Here is a picture to illustrate:

Note, those numbers are completely made up. I have no idea what the eye patterns for haswell actually look like, but they should look something like this.

 

[dark zero]

Supposedly the límit is at 4.7 or.4.8 Ghz to above.. That is where the voltage gets increase dramatically for some Mhz more

 

[Dresdenboy]

To fully understand the temperature increase it's necessary to look at the thermal conductivity of the cooling system. This should explain the delta between expected temperature based on known voltage and frequency to the observed temperature.

The necessary voltage hockey stick is caused by clock skew and degradation.

 

[Ketchup]

Not to get too technical here, as others can do it much better than I, but this is normal CPU behavior from what I can see. CPU architecture, no matter how far along it is, always has limits. Increased voltage, decreased temperatures, and so-called "binned" chips can push these limits, but you are still left with what the architecture can do.

Voltage will help push temperature up, but what really pushes this up is frequency. And when you put the two together, the temperature increase is magnified. Ever since I started overclocking with the original Athlon, I have observed that as soon as you hit a frequency that requires more voltage than stock, the temp numbers increase exponentially. And over the past 15 years, I have never seen a CPU that doesn't abide by the exact same behavior.

 

[know of fence]

To me it seems like 3.8GHz is the point where it really breaks out and pulls what the greenies dont like to call a hockey stick.

As for why it happens, I believe it happens because the square wave that you envision occuring inside the transistors actually turns into a sine wave at higher frequencies. At a certain frequency, the width of the bottom of the square wave, which represents the time where the transistor is completely off, is reduced to zero. So the period where the transistor is consuming zero power is reduced to almost nothing, resulting in an increase in power in a linear fashion. As long as even a tiny portion of the cycle results in the transistor spending some time in the off state, the function is linear. But when you go past that critical frequency, the transistor is never truly off. At this point you need higher voltage to make the sine wave larger which lets the transistor actually hit its off state. This function is nonlinear because the voltage is also increasing. Here is a picture to illustrate:

Note, those numbers are completely made up. I have no idea what the eye patterns for haswell actually look like, but they should look something like this.

That's an explanation I wasn't expecting, which is a good thing. But it implies that this threshold isn't actually movable or at least not movable by much.

As far as the actual number there is a difference between something kicking in and kicking into full effect. At 3.8 GHz the voltage jump (35 mV) is already twice the average, which would be full effect. However I was looking for the point at which these additional voltage increases are beginning. I can shift the data points between red and blue, if I start the red graph at 3.8, lke you suggest then the trendlines intersect at 3.6. But I was looking for a point where these two match, the data as jumpy as it is, points to 3.3 GHz.

 

[KingFatty]

Are you able to isolate the temperature as a factor?

The problem is increasing temperature will increase the power consumption of the chip.

So maybe overlay the temperature on the data you have, to see if there is a threshold temperature where the current increases due to the increased temperature.

Or, you can try to choose a particular temperature, and then hold the chip at that temperature when testing each step on the chart.

It could be that you are just seeing the affect of increasing temperatures causing increased current consumption, because temps do not seem to be accounted for in your first chart, so we don't know what the temps were doing there?

Or, as seen in your 2nd chart, when the temps increased, that resulted in the expected increase of power consumption, so really the first chart is just another way to visualize the temperature indirectly via the effect that the temperature had on current consumption.

 

[dmens]

AFAIK, once you hit a certain speed, gate delay does not scale well with voltage. Wire delay does not really scale nicely with voltage.

When you overclock and try to crunch a pipeline with a design point of some gate/wire delay to go faster, the ROI on voltage bumps decreases as you go higher. And since voltage has a squared relation to power, your temps skyrocket.