Overclocking a Core i5 3570K
 

I've just built a new Core i5 3750K based system. It seems to be stable at 4.4GHz with vcore + 0.0v. Core temperatures while torture testing were around 60 degrees. I'm currently torture testing at 4.5GHz with vcore + 0.1v. (It wasn't stable at vcore + 0.05v.) Temperatures are up to about 75 degrees.

If its stable at 4.5GHZ at this or maybe a little higher voltage, then my plan is to pull back on the overclock - to 4.3GHz, say, then resume P-1 testing for GIMPS. I have 2 questions:

1. Will running 24/7 for years with all cores at 100% at these kinds of speeds and temperatures damage or significantly reduce the lifespan of my processor?

2. Defining "stable" to mean 24 hours of torture testing with no errors, will pulling back 200MHz from the highest stable overclock be a sufficient margin for production or should I take it even lower?

[QUOTE=Mr. P-1;327321]1. Will running 24/7 for years with all cores at 100% at these kinds of speeds and temperatures damage or significantly reduce the lifespan of my processor?[/QUOTE]Yes. Read up about electromigration. Although the precise determination of "significantly" is open to debate.[QUOTE=Mr. P-1;327321]2. Defining "stable" to mean 24 hours of torture testing with no errors, will pulling back 200MHz from the highest stable overclock be a sufficient margin for production or should I take it even lower?[/QUOTE]If it was me I would take it all the way back to stock. This has been discussed here before but in summary: You could stand to lose more work through occasional errors than you can make up for with over-clocking. But the exact trade-off point is dependant upon which type of work you intend to do. For P-1 I'm not sure where the trade-off point is but I usually use the metric of one month of runtime per work unit is the maximum work unit I commit for stock machines and reduce the time period accordingly for over-clocked machines.

Edit: Where the meaning of "work unit" is: a minimal computation sequence that must run from beginning to end before any meaningful result is obtained.

Edit2: For runs longer than one month always use ECC enabled hardware.

I disagree with retina.

1) Yes it will reduce the life of your CPU but it is unlikely to reduce it to before the point of obsolescence. I would try to keep temps < 70C if possible. You need to leave room for those very hot days.

2) I find torture tests aren't the best for determining stability. Throw in a bunch of double checks - stagger them so that your computer is tested with 0/1/2/3/4 DCs and 4/3/2/1/0 P-1 tests simultaneously and verify if your DCs match. If you can do 10-12 DCs without any mismatch I'd say you are good to go. There is no hard and fast rule but 200MHz should be sufficient.

Core i5 3[B]75[/B]0k?
Guess you mean either Core i5 3[B]57[/B]0k or Core i7 3[B]77[/B]0k :razz:.

[QUOTE=retina;327322]Yes. Read up about electromigration.[/QUOTE]

From what I can find out, the rate of EM deterioration is dependent upon three factors:

Voltage: I do not intend to overvolt at my production overclock.

Temperature: My temperatures at 4.4GHZ with my installed Frio cooler were about 3 degrees higher than at stock speed with the stock cooler. At 4.3GHZ, I would expect it to run cooler.

Frequency. My understanding is that the rate of EM deterioration varies proportionately with frequency, which I view as not reducing the life of the CPU (measured in CPU ticks), rather compressing it in terms of wallclock time.

[QUOTE=VictordeHolland;327342]Core i5 3[B]75[/B]0k?
Guess you mean either Core i5 3[B]57[/B]0k or Core i7 3[B]77[/B]0k :razz:.[/QUOTE]

Yes.

This is an example of the situation P-1 is trying to gauge.
If a cpu has an error rate of 2% at stock then the necessary work for testing 100 exponents is 204 tests in 204 time.
If you overclock your cpu by 10% and have an error rate of 5% then 210 tests are needed but they will be done in 210/1.1=190.909090... time. This is a saving of about the time of 13 tests at stock.

The trouble is determining the error rate but that would be the same with all cpus. Retina seems to be suggesting underclocking cpus to eliminate errors completely. It is amazing what error percentage it is possible to have before you are behind the production speed of stock. For the example above an error rate of 12.2% breaks even.

What speed increase are you getting Mr. P-1?

[QUOTE=garo;327328] I would try to keep temps < 70C if possible. You need to leave room for those very hot days.[/quote]

I live in Scotland. There aren't very many hot days.

That said, I do intend to monitor the temperature and keep it at about 60, reducing the overclock if necessary as the weather gets warmer.

[quote]2) I find torture tests aren't the best for determining stability.[/quote]

They're good for quickly establishing instability. But I take your point - I should prove, using doublechecks, that my system is stable at x MHz before going into production at x-200 MHz.

But before I do any of that, I'm thinking that I should test it for a week or so at stock speed. That way I will find out it I'm having any problems which are nothing to do with the overclock. For example: at 2.5MHz +0.1v, mprime has segfaulted once. Is that because of the overclock or some other problem?

I've also noticed some peculiarities in the temperature graph: it goes through occasional episodes of square-waving, when every few seconds the temperature will suddenly drop to about 45 degrees for about a second or so, before returning to the 60-70 range. The torture test doesn't fail, and the threads are still getting the usual 98-100% of processor time. So it looks to me as though it's throttling, even though it's nowhere near the temperature at which it should do this.

[QUOTE=henryzz;327351]Retina seems to be suggesting underclocking cpus to eliminate errors completely.[/QUOTE]Not at all. Underclocking will, in most cases, reduce the error rate but I doubt it can ever eliminate errors completely.

[QUOTE=henryzz;327351]This is an example of the situation P-1 is trying to gauge.
If a cpu has an error rate of 2% at stock then the necessary work for testing 100 exponents is 204 tests in 204 time.
If you overclock your cpu by 10% and have an error rate of 5% then 210 tests are needed but they will be done in 210/1.1=190.909090... time. This is a saving of about the time of 13 tests at stock.[/QUOTE]

That's true for DC assignments, where all you lose for a erroneous run is the time you spent on it. It's also true for an LL, though GIMPS may not find out about it (including the possibility of a missed prime) until the DC wave catches up.

For factorisation, (trial or P-1), the calculation is a little different. If an erroneous run fails to find a factor which it should have, the cost to the project isn't that the factorisation has to be repeated. The cost is that two unnecessary LLs get done.

I also think your error rate assumptions are implausible. The overall error rate for the project is 2%. In practice, this means most machines are error free, while a few have a much higher error-rate. Error-prone machines do not hurt the project - and might still contribute if they are sometimes error-free - if they are doing DC checks only. Error-prone machines doing factorisation are likely to harm it.

[quote]The trouble is determining the error rate but that would be the same with all cpus.[/quote]

Not necessarily. Overclocking-related errors are caused by a circuit not completing its calculation by the time its results are read at the next clock pulse. Two corresponding circuits in two cores may have fractionally different electrical properties, causing one to take slightly longer to complete its calculation than the other.

[QUOTE]What speed increase are you getting Mr. P-1?[/QUOTE]

I don't know, as I'm currently booted up non-overclocked and didn't check before. Here are my non-overclocked benchmarks:

[code]Compare your results to other computers at http://www.mersenne.org/report_benchmarks
Intel(R) Core(TM) i5-3570K CPU @ 3.40GHz
CPU speed: 3555.84 MHz, 4 cores
CPU features: Prefetch, SSE, SSE2, SSE4, AVX
L1 cache size: 32 KB
L2 cache size: 256 KB, L3 cache size: 6 MB
L1 cache line size: 64 bytes
L2 cache line size: 64 bytes
TLBS: 64
Prime95 32-bit version 27.9, RdtscTiming=1
Best time for 768K FFT length: 4.902 ms., avg: 4.934 ms.
[Mon Feb 4 19:52:51 2013]
Best time for 896K FFT length: 6.109 ms., avg: 6.177 ms.
Best time for 1024K FFT length: 6.753 ms., avg: 6.768 ms.
Best time for 1280K FFT length: 8.858 ms., avg: 8.875 ms.
Best time for 1536K FFT length: 10.625 ms., avg: 10.658 ms.
Best time for 1792K FFT length: 12.838 ms., avg: 12.862 ms.
Best time for 2048K FFT length: 14.115 ms., avg: 14.136 ms.
Best time for 2560K FFT length: 18.314 ms., avg: 18.358 ms.
Best time for 3072K FFT length: 22.194 ms., avg: 22.247 ms.
Best time for 3584K FFT length: 27.787 ms., avg: 27.809 ms.
Best time for 4096K FFT length: 30.450 ms., avg: 30.487 ms.
Best time for 5120K FFT length: 39.470 ms., avg: 39.566 ms.
Best time for 6144K FFT length: 48.028 ms., avg: 48.044 ms.
Best time for 7168K FFT length: 59.815 ms., avg: 59.862 ms.
Best time for 8192K FFT length: 67.326 ms., avg: 67.416 ms.
Timing FFTs using 2 threads.
Best time for 768K FFT length: 2.567 ms., avg: 2.591 ms.
Best time for 896K FFT length: 3.151 ms., avg: 3.175 ms.
Best time for 1024K FFT length: 3.492 ms., avg: 3.512 ms.
Best time for 1280K FFT length: 4.622 ms., avg: 4.762 ms.
Best time for 1536K FFT length: 5.479 ms., avg: 5.624 ms.
Best time for 1792K FFT length: 6.641 ms., avg: 6.657 ms.
Best time for 2048K FFT length: 7.273 ms., avg: 7.307 ms.
Best time for 2560K FFT length: 9.413 ms., avg: 9.619 ms.
Best time for 3072K FFT length: 11.416 ms., avg: 11.443 ms.
Best time for 3584K FFT length: 14.197 ms., avg: 14.229 ms.
Best time for 4096K FFT length: 15.613 ms., avg: 15.680 ms.
Best time for 5120K FFT length: 20.288 ms., avg: 20.366 ms.
Best time for 6144K FFT length: 24.626 ms., avg: 25.045 ms.
Best time for 7168K FFT length: 30.428 ms., avg: 30.546 ms.
Best time for 8192K FFT length: 34.430 ms., avg: 34.818 ms.
Timing FFTs using 3 threads.
Best time for 768K FFT length: 1.872 ms., avg: 1.896 ms.
Best time for 896K FFT length: 2.181 ms., avg: 2.205 ms.
Best time for 1024K FFT length: 2.449 ms., avg: 2.479 ms.
Best time for 1280K FFT length: 3.259 ms., avg: 3.403 ms.
Best time for 1536K FFT length: 3.861 ms., avg: 3.895 ms.
Best time for 1792K FFT length: 4.647 ms., avg: 4.697 ms.
Best time for 2048K FFT length: 5.149 ms., avg: 5.318 ms.
Best time for 2560K FFT length: 6.679 ms., avg: 6.803 ms.
Best time for 3072K FFT length: 8.107 ms., avg: 8.159 ms.
Best time for 3584K FFT length: 10.041 ms., avg: 10.513 ms.
Best time for 4096K FFT length: 11.063 ms., avg: 11.106 ms.
Best time for 5120K FFT length: 14.333 ms., avg: 14.387 ms.
Best time for 6144K FFT length: 17.335 ms., avg: 17.784 ms.
Best time for 7168K FFT length: 21.178 ms., avg: 21.542 ms.
Best time for 8192K FFT length: 24.145 ms., avg: 24.206 ms.
Timing FFTs using 4 threads.
Best time for 768K FFT length: 1.564 ms., avg: 1.600 ms.
Best time for 896K FFT length: 1.741 ms., avg: 1.791 ms.
Best time for 1024K FFT length: 2.004 ms., avg: 2.050 ms.
Best time for 1280K FFT length: 2.693 ms., avg: 2.885 ms.
Best time for 1536K FFT length: 3.273 ms., avg: 3.321 ms.
Best time for 1792K FFT length: 3.882 ms., avg: 3.946 ms.
Best time for 2048K FFT length: 4.400 ms., avg: 4.465 ms.
Best time for 2560K FFT length: 5.746 ms., avg: 5.821 ms.
Best time for 3072K FFT length: 6.901 ms., avg: 7.006 ms.
Best time for 3584K FFT length: 8.420 ms., avg: 8.531 ms.
Best time for 4096K FFT length: 9.501 ms., avg: 9.764 ms.
Best time for 5120K FFT length: 12.002 ms., avg: 12.036 ms.
Best time for 6144K FFT length: 14.494 ms., avg: 14.546 ms.
Best time for 7168K FFT length: 17.653 ms., avg: 17.715 ms.
Best time for 8192K FFT length: 19.786 ms., avg: 19.878 ms.
Best time for 61 bit trial factors: 2.630 ms.
Best time for 62 bit trial factors: 2.658 ms.
Best time for 63 bit trial factors: 4.385 ms.
Best time for 64 bit trial factors: 4.386 ms.
Best time for 65 bit trial factors: 4.373 ms.
Best time for 66 bit trial factors: 4.341 ms.
Best time for 67 bit trial factors: 4.340 ms.
Best time for 75 bit trial factors: 4.559 ms.
Best time for 76 bit trial factors: 4.557 ms.
Best time for 77 bit trial factors: 4.538 ms.[/code]

[QUOTE=Mr. P-1;327359]I don't know, as I'm currently booted up non-overclocked and didn't check before. Here are my non-overclocked benchmarks:[/QUOTE]

What's the point of OCing if you don't even bother to check if it's giving an appreciable speed increase?

Since you long table is just for one mode and thus does nothing to answer the question, why post it? Do the differential diagnosis, then get back to us.