I7 3770K Overclocking / Asrock Z77 extreme4

[1b0b1]

Why are you idling @ 4.2ghz with Speedstep on?

It was a wrong setting applied under the power settings where the minimum processor state was set to 100%, it was rectified though

 

[Vectronic]

you can stick with LrgFFT or even Blend for stability tests (though I suggest much longer tests than 5 minutes).

Small FFT/IBT are good for finding out the maximum your temps will/could reach.

I keep mine at 4.2GHz decent overclock (24%), temps are always below TCase, and I can even shut off my CPU fans, and it will still stay below 80C under load... so even in the worst case scenario (fan failures, etc) it won't completely melt down.

 

[1b0b1]

Well, below 67.4 is kinda hard to achieve unless you've got a decent cooler. I achieved 42 multiplier with a maximum of 72c running a small FFT for 20 minutes.

 

[Vectronic]

Our coolers are about the same (especially if you have yours in a Push/Pull setup)... so are our cases.

Fractal Design R4
2x140 (front) + 1x120 (bottom) for intake (all running at 5volts, or about 380RPM)
1x140 (rear) for exhaust (currently 1000 RPM)

It's really up to you what you consider safe temperatures... anything below 80C is fine really... I'm basically setting mine up for durability/longevity... I *could* run it at 4.5GHz or even 4.6GHz but that's way too noisy, and if something fails, way too dangerous.

 

[1b0b1]

Alright, Small FFT didn't crash after 23 minutes and maximum temp reached was 79. I'll try the custom test where the previous voltage crashed on after 10 min

 

[felang]

I like to stability test with OCCT, might not be as demanding as IBT but for a gaming PC it seems to be more than enough, also in my experience it detects errors faster than P95.

 

[1b0b1]

I like to stability test with OCCT, might not be as demanding as IBT but for a gaming PC it seems to be more than enough, also in my experience it detects errors faster than P95.

I'll do a test with it, any specific settings ?

 

[Vectronic]

For a complete stability test, you'd probably want to use Blend, and let it run for 12 to 18 hours...

For quick testing, probably Large FFT for 10 minutes, if that passes, try IBT for 5 runs on Maximum (probably take about 20 minutes)... then fire up the most resource using game you have, play for 10 minutes... if you do any 3D rendering, render a fairly large scene... put on some tunes, toss 50 images into Photoshop, run one of the automated scripts on all of them...etc.

There really isn't a definite way to test... if you pass various short tests, up the voltage by 0.01v, and you're probably good.

If you run into errors with P95, write down where it errored (what FFT, etc) that way you can do targeted custom tests rather than having to wait for it to get to that point again.

 

[1b0b1]

1 worker stopped after 12 minutes , now what.

 

[Vectronic]

it's fun isn't it?... up the voltage, try again...

Keep in mind there are more things to change than just multiplier and core voltage... LLC, your RAM voltage, and timings... have pretty large impacts

Once you get "stable" there's also overvolting or undervolting other things, VTT/PCH/VCCSA... you might even be able to get stable with a higher BCLK using a lower voltage, or a lower BCLK and higher multiplier, etc... there really isn't *that* many things but I'm still finding various tweaks... that you pretty much have to find yourself because no board, CPU, or system is the same.

If you are using Offset, you can also do "weird" things, like set your Turbo Voltage really high +0.300, and your Offset really low -0.300... so you might be able to idle at 0.6v, but under load it will still go up to 1.2v... it widens the voltage range... but is also harder to get stable.

 

[1b0b1]

it's fun isn't it?... up the voltage, try again...

Haha yes totally fun, I actually don't mind, but with this voltage -0.080 and the temps almost reached 80c. This is the maximum I can tolerate. I think i will have to go for 43 multiplier.

Or do you think if i decreased the PLL voltage and increased the Vcore, it might work :sneaky:

 

[Vectronic]

Read my changes to previous message...lol

It's possible, you might even want to overvolt PLL and keep the same Vcore... you'll pretty much get the same temperature, but it might be stable, rather than not stable.

For PLL though, baby-steps, and keep it under 1.9v... it's just a whole bunch of playing around to see what works.

Whats' you're current setup at the moment?.. -0.080 @ 4.4 or 4.5?... but, what's Turbo Voltage at?... is hyper-threading enabled or disabled?... assuming you still have other voltages on "auto" what do they report at?

If you're at 4.5... try 4.4 first rather than 4.3... that 100MHz could be as much as 10C cooler.

 

[felang]

I'll do a test with it, any specific settings ?

Not really, just make sure to use 64 bit and stock test should be fine. Their is also an option to use a more stressful LINPACK based test, but IMO OCCT stock test is fine.

 

[1b0b1]

Read my changes to previous message...lol

It's possible, you might even want to overvolt PLL and keep the same Vcore... you'll pretty much get the same temperature, but it might be stable, rather than not stable.

For PLL though, baby-steps, and keep it under 1.9v... it's just a whole bunch of playing around to see what works.

Whats' you're current setup at the moment?.. -0.080 @ 4.4 or 4.5?... but, what's Turbo Voltage at?... is hyper-threading enabled or disabled?... assuming you still have other voltages on "auto" what do they report at?

If you're at 4.5... try 4.4 first rather than 4.3... that 100MHz could be as much as 10C cooler.

Now I only upped the PLL voltage a little bit and the prime test that failed before till now is working fine.

hyperthreading is enabled, Vcore -0.080v @ 4.4 , turbo voltage: +0.004v. for the rest of the settings I'll check next time I restart.

 

[1b0b1]

Hmmmm, I was watching a video on youtube and it stopped: Shockwave flash crashed. But on prime95 nothing wrong yet.. Could this be an instability or just a coincidence ??

 

[Vectronic]

Thats seems s***ty...lol, I don't know if it's because of HT (i5 doesn't have it) but at least on mine that seems over-volted still.

If you disable HT you might get stable with less voltage (thus less heat), and you probably don't need 8 @ 2.2GHz... 4 @ 4.4 will run quicker for almost everything, but that's up to you.

Could be instability, could just be a glitch with Flash/YouTube... assume instability.

 

[1b0b1]

I am not going to disable hyper-threading, out of the question , btw it seems that everything is okay after upping the PLL voltage a little bit. Prime still running, 30 minutes.

I think I'll do more tests on this settings, maybe it works

 

[1b0b1]

So I found that the best thing is to go back to 42 multiplier with a -0.110v offest and a +0.004 tubro voltage, everything else is on default settings.

This offset lead to the same voltage as the stock one (1.128v~1.136v on full load and 0.7*~0.8* on idle) and everything is working fine.

I'll try to go higher once I get a better cooler.
Thanks for the help anyways.

 

[lehtv]

Looks good. 42 is a good sweet spot for most 120mm coolers, you just need something heavier and bigger to properly handle the heat from increased voltages

I did temporarily run my 3770K at 4.7GHz with HT off. My cooler technically can't handle stress testing it (CPU temp >90C) but I did it anyway to reduce bottle necking in Planetside 2 where temperatures stayed at 70C or so.

 

[1b0b1]

Well, I think I'll raise the bar at little bit but not now knowing that I will most probably never reach high temps on a normal daily use, not even while gaming or rendering. The most important thing is to at least make it stable at the lowest voltage possible.

 

[Idontcare]

Hmmmm, I was watching a video on youtube and it stopped: Shockwave flash crashed. But on prime95 nothing wrong yet.. Could this be an instability or just a coincidence ??

(I know I probably sound like the proverbial broken record about this subject, but I'll post it up again )

There are over 700 instructions in the ISA of your processor.

Stress testing with prime95 or IBT/LinX might, at best, result in you knowing if maybe 100 or so of those instructions are "stable" at your specific OC'ing parameters and conditions.

That means you are blind to the stability prospects of nearly 85% of the instructions your other applications might possibly task the CPU with computing...and it only takes one of those other 600 instructions to be unstable such that your app in question (including your OS) crashes mid-stream.

We rely on stress testers like LinX and Prime95 not because they do a good job in testing all the supported instructions in the ISA but because they heat up the processor in such a way that gives us some insight into the stability of the processor under what the industry calls "accelerated lifetime reliability testing".

If shockwave crashed then it is highly likely that there is a particular instruction called by shockwave that is not stable, even at the lower operating temperatures that come when not stress testing, and stress testing with prime95 is simply not going to catch that for you.

What you can do in this case is run prime95 on low priority so your processor is good and warm, but the OS is not laggy because it is assignd low priority. Then go run watch the youtube video. It will call on the same instruction that is the weaker-linker in your CPU's circuitry and it will have increased odds of crashing because of the elevated operating temperatures thanks to prime95.

Then tune your OC parameters so the video plays without causing stability issues.

 

[1b0b1]

Thanks anyways man your help is much appreciated. I honestly don't count a lot on prime95. What matters the most to me is the applications/games that I am using, if those are stable, then the OC is fine for me.

 

[1b0b1]

Plus, different tests under the same stressing program could some time lead to different results where some are stable and some are not.

 

[Vectronic]

There are over 700 instructions in the ISA of your processor.

I always find it annoying that there isn't any "test everything" tests. I'm sure Intel/AMD/etc have them (surely they don't do it all "theoretically")... someone leak that s**t...lol

 

[Idontcare]

I always find it annoying that there isn't any "test everything" tests. I'm sure Intel/AMD/etc have them (surely they don't do it all "theoretically")... someone leak that s**t...lol

If I could program worth sh*t then I would do it and open source it. My coding skillz aren't up to that challenge though

I've wondered if you could do something like that in a poor-man's way by say running LinX on low-priority and then simultaneously running multiple instances (one or two per core) of AIDA64's "Instruction Latency Dump" routine which queries each instruction in the ISA and reports its latency.

It still wouldn't be good enough though because while AIDA64 would crash if the instruction was so borked as to result in an fault when it was executed but there is no error checking involved so you still wouldn't know if the computed results were correct.