Linpack 11 temps - when torture just isn't enough

[JimmiG]

This is AIDA64's FPU-only stress test, also known as the "CPU killer" due to the excessive heat it produces, running on my 4770K @ 4.3 GHz (HT disabled):



It tops out at about 82-83C. Pretty hot, eh? However I think my temps are reasonable given the overclock and cooler I have. They're a bit higher with HT enabled, but still fine. It's a moderate, safe overclock with a very good air cooler (Phanteks PH-TC14PE).

Today I decided I really hate my CPU and wanted to really torture it, so I downloaded Linpack v 11.0.5, which supports AVX2.


The first part of lower temps in the graph are with FPU+CPU checked in Aida64, commonly known as the "balanced" stress test. The bump in temps is when I unchecked CPU, running only the FPU test, as above. I then stopped the built-in tests and fired up the Linpack benchmark (runme_xeon64.bat)



What is this I don't even...

Here is the statistics page:

I think I accidentally my whole CPU.

You can download the program from here:
http://software.intel.com/en-us/articles/intel-linpack-benchmark-download-license-agreement

A few notes:

-It might *kill* your CPU! If you're already in the mid 80's or above with the FPU-only stress test, I would strongly advise against running it at all. Even if you have super high-end cooling, those "spikes" might not give your cooler enough time to even do its job.

-If you're on "Auto" voltage rather than fixed, don't run it. We already know that AVX v1 increases the VCore way beyond what it normally is. Who knows what this application will do to your VCore?

-If you have an i7, you need to disable HT for the included .bat file to work. Honestly I wouldn't even want to run it with HT on, since that can add 5-8C to the temps.

-If anyone has the time, it would be very interesting to see the results of a 4670 or 4770 running at stock speeds with the stock cooler. I'm really curious if Intel really dimensioned the tiny stock cooler for this kind of load. It's an extreme case for sure, but in the end it's just a normal program running on your computer, and AVX2 is advertised as one of the major new features of Haswell.

 

[inf64]

I loled at how funny you worded your topic . Yeah, linpack is basically CPU "killer", I have never ever seen CPU temps so high and my poor X4 OCed Piledriver throttling to 1400Mhz occasionally as I saw it in linpack. It just humiliates CPUs...

 

[crashtech]

For some reason this reminded me of the Airbus A300 rudder. You can fully deflect it either way you want in flight, but if you decide to rapidly and fully deflect it from left to right a few times at speed, the tailplane and rudder depart the aircraft.

 

[TuxDave]

-It might *kill* your CPU! If you're already in the mid 80's or above with the FPU-only stress test, I would strongly advise against running it at all. Even if you have super high-end cooling, those "spikes" might not give your cooler enough time to even do its job.

I don't know about it killing your CPU. Your heatsink is only part of the equation for saving your CPU. The on die throttling is the other part that's SUPPOSED to save you.

 

[BallaTheFeared]

I think it's unlikely you'll kill it from just heat at that voltage, the self throttle works quite well with Haswell.

It's all relative though.

 

[Yuriman]

Sounds like you need to delid and cold-weld your bare-die to a waterblock.

 

[Idontcare]

I don't know about it killing your CPU. Your heatsink is only part of the equation for saving your CPU. The on die throttling is the other part that's SUPPOSED to save you.

The on-die throttling is built with a threshold value that is only valid if you are operating the CPU at stock clocks and stock volts.

Taking your over-volted and over-clocked CPU up to its TJmax throttling point is most certainly inducing accelerated aging and degradation to your CPU.

 

[BallaTheFeared]

Does clock speed in such a scenario make any real difference? I wouldn't think it would.

1.21v is just a shade over stock for a lot of chips it seems, probably actually quite close to what stock chips will boost vid to for AVX code.


Not saying he isn't accelerating the electron degradation, but it seems unlikely it's noticeable over what Intel decided to ship at stock. They seem to throttle with the stock cooler and overvolt higher - at least I've seen one so far do as such.

 

[Idontcare]

Does clock speed in such a scenario make any real difference? I wouldn't think it would.

1.21v is just a shade over stock for a lot of chips it seems, probably actually quite close to what stock chips will boost vid to for AVX code.


Not saying he isn't accelerating the electron degradation, but it seems unlikely it's noticeable over what Intel decided to ship at stock. They seem to throttle with the stock cooler and overvolt higher - at least I've seen one so far do as such.

Clockspeed absolutely makes a real difference. Clockspeed determines the amps (current flow), as does voltage, both of which factor into an equation that determines the rate of degradation which itself is exponentially dependent on temperature, voltage, and current flux.

That said, buy the tuning plan and don't worry about it. Or delid (invalidates the tuning plan warranty), reduce your peak temperatures by 20C and likewise don't worry about.

 

[ehume]

The on-die throttling is built with a threshold value that is only valid if you are operating the CPU at stock clocks and stock volts.

Taking your over-volted and over-clocked CPU up to its TJmax throttling point is most certainly inducing accelerated aging and degradation to your CPU.

Clockspeed absolutely makes a real difference. Clockspeed determines the amps (current flow), as does voltage, both of which factor into an equation that determines the rate of degradation which itself is exponentially dependent on temperature, voltage, and current flux.

That said, buy the tuning plan and don't worry about it. Or delid (invalidates the tuning plan warranty), reduce your peak temperatures by 20C and likewise don't worry about.

I wish I had known that before I started OC'ing my 4770k. I didn't dwell in throttle land, but I made repeated visits.

I'm glad now that I settled for an OC that can handle Linpack 11.0.5's AVX2 with temps that are "merely" in the 80's.

 

[TuxDave]

The on-die throttling is built with a threshold value that is only valid if you are operating the CPU at stock clocks and stock volts.

Taking your over-volted and over-clocked CPU up to its TJmax throttling point is most certainly inducing accelerated aging and degradation to your CPU.

Woops, missed the OC part of in the OP.

 

[BallaTheFeared]

Clockspeed absolutely makes a real difference. Clockspeed determines the amps (current flow), as does voltage, both of which factor into an equation that determines the rate of degradation which itself is exponentially dependent on temperature, voltage, and current flux.

That said, buy the tuning plan and don't worry about it. Or delid (invalidates the tuning plan warranty), reduce your peak temperatures by 20C and likewise don't worry about.

Sorry for the density of my skull, but can you explain how clock speed directly effects that?

I measure a 10w difference with the same voltage at 4.8GHz as I did at 4GHz.

Temp difference was roughly 2C.

Why would it matter if he was at 4.3GHz or 3.6GHz if his voltage was the same, he would still most likely throttle. The voltage he's giving it isn't that far outside of shipping spec due to the .1v bump from AVX code. A sample that ships at 1.1v stock (Definitely within the realm of possibility) will likely boost voltage to 1.2v in AVX for instance.

Clock speed would seem to have the most trivial effect, outside of it's direct dependence on voltage to hit it. In this case 1.2v isn't that high, so I think what I'm saying is there is little reason to say there would be a wealth of additional electron migration taking place over stock and what Intel deemed acceptable.

That said, we still don't know the shelf life on these chips. I'm still seeing 13 year old Pentium 4's and older processors by way of repairs, the paste isn't even paste anymore, they run hot, and have a ton of dust build up sometimes that is nothing short of sad. And yet they still run stock no problem, thirteen years later. :thumbsup:

 

[Idontcare]

Sorry for the density of my skull, but can you explain how clock speed directly effects that?

Clockspeed directly affects the amount of current flowing through the device, both active and leakage.

All current is damaging, regardless the voltage or temperature.

However the damage is factored into the device lifetime when the operating voltage, clockspeed, and TJmax are all defined and spec'ed.

Elevating the clockspeed, even with stock voltage and no temperature increase, reduces the lifetime of the chip in non-linear fashion because the degradation mechanisms for current-induced damage is itself non-linear (it varies exponentially with the total amperage flowing through the area of the chip that is being damaged).

Elevating the clockspeed while also elevating the temperature and/or voltage merely compounds the deleterious enhancement of the rate of degradation.

 

[BallaTheFeared]

Hehe, feel so dumb...

Where is the extra amperage coming from? If the cpu was pulling more amps, wouldn't that be reflected in the power draw?

Voltage x Amps = Watts

No?

 

[ShintaiDK]

Essentially higher frequency=more amps.

 

[BallaTheFeared]

10w more is what I was able to measure via a kill-a-watt at the same voltage from 4GHz to 4.8GHz, obviously if I honed 4GHz down I'd be below shipping voltage at around 0.987v, but for the point of my inquiry we keep them the same.

So my question is if 1.2v is accepable shipping votage for Intel, how much of a difference is 3.8GHz vs 4.3GHz going to actually make when it comes to electron migration if the amp change is so small going from 4GHz to 4.8GHz?

 

[Ben90]

You can run Linpack without a CPU cooler and it wont damage it, much less kill it. My 920 was running at 5* away from TJMax for pretty much until I bought this Haswell system.

 

[Idontcare]

10w more is what I was able to measure via a kill-a-watt at the same voltage from 4GHz to 4.8GHz, obviously if I honed 4GHz down I'd be below shipping voltage at around 0.987v, but for the point of my inquiry we keep them the same.

So my question is if 1.2v is accepable shipping votage for Intel, how much of a difference is 3.8GHz vs 4.3GHz going to actually make when it comes to electron migration if the amp change is so small going from 4GHz to 4.8GHz?

Power consumption is fairly straightforward, dynamic power consumption increases linearly with clockspeed (variable "GHz" below):

The amp change from 4GHz to 4.8GHz will be 4.8/4 = 20%. At a minimum, so to will be the decrease in lifetime for the chip.

Just because the absolute value of the watts is small that doesn't mean the impact to lifetime reliability is small.

 

[oneofusjustin]

how do you add threads to linpack? i got it running but i wanna test 8 threads not 4

 

[BallaTheFeared]

I was hoping for something I could understand, but due to my lack of technical knowledge I'm not sure anything other than "just cause" would suffice.

My qualm has been how much, in your initial post it seemed like you inferred a rather dramatic decrease in lifespan. It would seem like very little given the low voltage and low clock speed (compared to stock) of the OP.

Taking a stock cpu and running it at the throttle point vs 4.3GHz at the same voltage, quantify how much you think that will accelerate stock vs OC @ same vcore.

20%? Less? I know it's probably an impossible question, same as answering how much life is actually available to start with... But that is really where my question stems from.

 

[ShintaiDK]

how do you add threads to linpack? i got it running but i wanna test 8 threads not 4

I added it to the LinX GUI interface. You can get it here:
http://shintai.ambition.cz/linx.zip

If it complains, try copy the DLL file into the Windows\System32

Then you can use the interface to set whatever you like easily.

64bit only! (Due to the DLL)

 

[Idontcare]

I was hoping for something I could understand, but due to my lack of technical knowledge I'm not sure anything other than "just cause" would suffice.

My qualm has been how much, in your initial post it seemed like you inferred a rather dramatic decrease in lifespan. It would seem like very little given the low voltage and low clock speed (compared to stock) of the OP.

Taking a stock cpu and running it at the throttle point vs 4.3GHz at the same voltage, quantify how much you think that will accelerate stock vs OC @ same vcore.

20%? Less? I know it's probably an impossible question, same as answering how much life is actually available to start with... But that is really where my question stems from.

It will be 20% less, at a minimum, but likely worse depending on a few factors that strongly depend on the specific materials used in the device itself.

Electrons flowing through a material is kind of like radiation flowing through your body. The flux is proportional to your likelihood of dying from radiation poisoning, but it is not a linear proportionality. A little radiation just means an increase in the chances of getting cancer 20yrs from now. A little more radiation raises the odds, a little more raises the odds of cancer 20yrs from AND suddenly increases the odds of getting a different type of cancer 5yrs from now.

Raise the radiation dosage more still and you start increasing your odds of dying next week, not from cancer but from the effects of radiation poisoning (you kill your blood cell factories in your bone marrow).

It is like this with current flow in your CPU. A little increase is bad, but not going to kill it quick. But it isn't linearly proportional.

Personally I wouldn't worry about it. Keep away from the high temps is what I am getting at. That is where you make things exponentially bad (as in the mathematical kind of exponential).

 

[sm625]

Sorry for the density of my skull, but can you explain how clock speed directly effects that?

I measure a 10w difference with the same voltage at 4.8GHz as I did at 4GHz.

There are thousands of individual capacitances in a CPU, if not millions or maybe even billions. I dont know how many but the point is that they all add up into one very large capacitance. (A capacitor is ideally open when the voltage across it is steady.) Every single logic level transition allows current to flow through that combined capacitance. So the more transitions you have per second, the more current you get. That's why its so important to at least clock-gate any part of the cpu that isnt needed.

 

[Daedalus685]

Electrons flowing through a material is kind of like radiation flowing through your body. The flux is proportional to your likelihood of dying from radiation poisoning, but it is not a linear proportionality. A little radiation just means an increase in the chances of getting cancer 20yrs from now. A little more radiation raises the odds, a little more raises the odds of cancer 20yrs from AND suddenly increases the odds of getting a different type of cancer 5yrs from now.

Raise the radiation dosage more still and you start increasing your odds of dying next week, not from cancer but from the effects of radiation poisoning (you kill your blood cell factories in your bone marrow).

Must resist urge to rant about why linear no threshold is a terrible model for stochastic radiation hazards... :sneaky:

 

[crashtech]

Must resist urge to rant about why linear no threshold is a terrible model for stochastic radiation hazards... :sneaky:

Shush, it's the "consensus" hypothesis.