prime small FFTs = ok. blend = 124 BSOD in 2 minutes

[the_stigma]

Hey all

I have a question regarding stability testing for a 2500K on a P8Z68-V pro (basically same stuff as the normal P68's).

I'm shooting for prime stable daily use setup, and my CPU seems to easily go to 4.7, maybe even higher) but above 4.6 I'm having a hell of a time trying to get the blend stress-test to work. small FFTs work fine and can run a long time with no issues, but blend CONSISTENTLY BSOD's the system 2-3 minutes into the test (always with 124 error).

From my own cursory research I think this indicates that the CPU can run fine, but that there might be something in its link to memory that is failing.

What sort of settings should I be looking at to get blend stable also? I'm intermediate-level with some understanding of what each mobo setting does, but I'd like some feeedback from the veteran on what the typical candidates are for this sort of consistent failure-pattern.
vccio voltage? DDR voltage? I dunno...

For reference I am running stock memory speeds using some very standard 1333 Kingston memory specced to 1.5V. Very standard stuff.

Thanks in advance!
-Stigma

 

[.vodka]

124 on Sandy is a vcore related BSOD. You'll find that unless your 2500k is a better than average overclocker, from 4.5GHz onwards the voltage it needs to be stable starts to get out of control and with it temperatures, and that's where the good chips stand out from the rest.

To dial in the exact voltage needed for stability at these clocks, use fixed voltage first then go towards offsets + different LLC settings to enjoy ~1v idle and what's required under load.

This guide pretty much applies to the entire Asus P67/Z68 range, it's full of interesting data.


My 2500k (on a Sabertooth P67) is good for up to 4.5GHz, 1.31v at ultra high LLC (75%) and a given negative offset, P95 AVX stable in all of its variants since day one (~3 years) cooled with a TRUE Spirit 140 Power- It even works fine with the deeper C states dropping idle CPU power to <5w. 4.6GHz and above break these and on top of that stability gets much harder to achieve. It'll do 4.7GHz LinX 6144MB stable, at least 5 passes at 1.36v, it even did 3 hours of P95 AVX blend without a single problem at these settings but it'd BSOD on web browsing for example, or some nonsense like that. Probably what you're experiencing right now...

 

[the_stigma]

No, my BSOD is directly related to the blend testing. I'm not really going into regular-use testing until I have it prime stable. I hear some users had issues with 124 BSODs when in mostly idle, but that's related to not using a manual voltage and stuff going wrong when it jumps between low-power and high-power states.

For my testing I'm using manual voltage so far. I'll deal with any other issues later - and for that matter I'd be ok with using manual voltage all the way as this is a gaming rig anyway, and a few more watts more power in idle isn't really any issue for me.

I'm not using extreme voltage to get to the 4.7 where I am now. with my current load-line calibration at extreme (making the vdroop into a very small vboost) my setting is 1.415 and actual is 1.424 under maximum load prime-smallFFTs. That's a voltage that should be fine given good cooling (which I have). I don't plan on going higher than this, but I don't think I've it any crazy vcore wall...

I know that 124 CAN certainly be vcore related, but then why am I stable in small-FFTs which as far as I know stresses the CPU itself significantly more than blend?

If anyone else has any ideas on what to try for this I'd really appreciate the help

-Stigma

 

[.vodka]

Well since you're already up there and have the cooling required, try 1.45v. This is what I meant with the voltage requirements for a given OC, you think you have what it's needed but in fact it's sometimes higher. Every chip is different.*

Also try a little VCCIO, it should help the memory controller if it's unstable (blend stresses the memory subsystem in particular, and you're running stock memory, so it's the CPU end that could use the push here). I've also read people experimenting with PLL voltage, lowering it from its stock 1.8v and achieving stability.

At higher multipliers (47+) PLL overvoltage (not PLL voltage, they're similarly named) is another setting that could be of use. It's mostly used when your CPU doesn't POST with a given multiplier to allow a higher OC. It sets things straight at these multipliers as they are at lower ones.


It gets sketchy above 4.5GHz. I gave my example just as that, an example. I'd like to hear other ideas on this, too..



edit: Since we're discussing the memory here. If you're using two memory modules, have you installed them on the preferred slots for two modules operation? It's stated in the manual, page 2-5, 2-11, slots A2 and B2 (blue slots)


edit2: *As another example I can provide from experience, my 2500k boots w8.1 at 4.5GHz and will happily chew through ten passes of linpack (AVX by default), 6144MB on a 2x4GB setup (kingston, 1333 9-9-9-24 @ 1866 10-10-10-24 1T 1.5v), at 1.25v. The moment you run P95 blend or try to boot linux, it'll BSOD 124. Hard. vcore up to 1.31v fixes both. It gets weird sometimes...

 

[the_stigma]

I'm going to run an over-nighter small FFT stability test first and foremost to ensure that current settings are solid for that task (I've run over an hour already, but 8hrs beats 1hr for making sure) - then I'll give vccio another try I guess. I have tried messing around with it a little already, and it didn't seem to do much.

Don't think I'll go to 1.45v. Even if it was stable there I wouldn't be comfortable with it for daily use. I suppose it could maybe tell me if vcore is related to the problem if its suddenly stable if i give it loads more juice though... I will have to consider it.

This post:
http://www.tomshardware.co.uk/forum/270754-29-blend-test-fail
Seem to indicate that it might be memory that can't handle it? That seems weird to me though, as the memory itself is untouched. It has to be the CPUs link to memory that is failing right?

Thanks for the help so far. I'm sort of aware of those setting you mentioned, but it's tricky if I don't know more specifically what to look at for this issue. Multiple variables like that makes for an exponential amount of combinations to test...

EDIT: yea, 2 sticks are in the blue slots.

-Stigma

 

[soccerballtux]

divide the problem in half.

so, memtest

blend is CPU and RAM. CPU seems fine? check RAM.

you just divided the problem in half.

 

[the_stigma]

divide the problem in half.

so, memtest

blend is CPU and RAM. CPU seems fine? check RAM.

you just divided the problem in half.

While under overclock as a stability test, or to remove the possibility of damaged modules?

-Stigma

 

[Mushkins]

While under overclock as a stability test, or to remove the possibility of damaged modules?

-Stigma

Take it one step at a time. Make sure the modules work properly at all first, then worry about whether or not they work properly during a stress/stability test.

Does the blend test still give you error 124 with everything at stock?

 

[flexy]

The small FFT test is really not so much a "stability" test but more a test to test your max. CPU temps. (It's normal that you can maybe run small FFTs for hours on end but crash in large FFTs or doing a memory check).

Even as a max. temp test, the small OCCT FFT test is not 100% accurate since it doesn't use AVX instructions like linpack of the latest Prime95. (correct me if I am wrong)...but it's still a good, quick check to see how hot your CPU could get under far-fetched circusmtances (There is no need by the way to run small FFT more than 10 mins since max CPU temps are usually reached in this time).

Testing blend or large is much better to test overall stability since, as the name implies, it uses more memory and not all data fits in the CPU cache.

Stability also means that your entire sub-system like "bus", memory etc. must be stable, and in addition to this a test must also check how good your CPU transitions from little load to high load etc. (This is why OCCT and Linpack does those tests in repeating bursts...little load...high load...little load...high load...etc...rather than just testing 100% at full load all the time).

TLDR: Use OCCT Small just for a quick idea about max temps but use large for actual stability testing.

If it is non stable w/ blend or large...you have to do the usual for your CPU which is normally increase VCore voltage and VInput voltage.

Here is a very comprehensive guide about Haswell overclocking:

http://www.simforums.com/forums/haswell-48ghz-on-air-building-a-haswell-system_topic46180.html

Edit: Even if you're not OCing your memory as you state you should make sure it's actually free from errors, so do a memtest of your memory first at *stock* CPU speed. Download from here: http://www.memtest.org, burn to DVD or USB and let it run at least one pass to see that it's fine.

Then go and test OCCT small for 10mins-15mins to check whether your temps are ok, and then do blend or large for a longer actual stability test. If you increase voltage, rinse and repeat, check temps w/ small and then do a longer blend or large test.

 

[the_stigma]

Thanks flexy - im running a memtest on stock just to be absolutely sure first before I proceed. I'm 99% certain there is no issue since this is not a new system and I've stability tested my lighter overlock (on a shitty cooler) before, but it doesn't hurt to make doubly sure. Gonna let it run 3-5 passes or so.

This is a Sandy Bridge 2500K by the way (yea I know, its old, but still surprisingly relevant which is why a new CPU has been sort of low on my hardware-priority list).

-Stigma

 

[the_stigma]

Small update:

Ran memtest (5 pass) at stock. All ok. I think we can regard memory as problem free.

Then tested the tip from the tomshardware post about lowering DDR3 frequency. It had no bearing on stability whatsoever as far as I can tell. Also gave a shot to boost the VCCIO a bit - but from everything I read this is something that you don't need to touch until you start getting into 2000+ mhz modules, so unsurprisingly it had no effect for me.

Gave it a shot to run the 47x100 with as much vcore as was comfortable testing with - about 1.460 (actual@load). Still crashes within 1-2 minutes, and it behaves essentially the same as it has been from 1.425 (actual@load) and up. While I wouldn't start truly sweating until I aproached the big 1.5, this is starting to get so far away from the sort of voltages I want to keep below that I question the worth of even testing any higher than this.

I've decided to find a blend-stable value for 46x just so that I have something to compare to. 1.405v fails fairly quickly (10 minutes ish) but 1.410 (1.416-1.424 actual@load) is now stable for an hour and a half (running as we speak), so either this is blend stable or I am probably no more than 0.005v away from the stable threshold.

Worth noting that 46x seemed to be aproaching small-FFT stable (my log indicates I stopped testing at 40mins) around 1.375 Vcore though - yet it apparently requires a lot more juice for blend. This is a surprise to me - I always thought small-FFT put more direct strain on the CPU. You really just need a lot more Vcore for blend? ... Either that or there is some aspect I am just missing here.

Given this (preliminary) data I guess it doesn't seem so unlikely anymore that my problem might just be that I've started to hit the bad part of the Vcore curve for my silicone. If 46x needed +0.035v to get from small-FFT to blend stability, I assume that at 47x the gap would probably be as big if not bigger. I was aproaching small-FFT stability at 1.420'ish (for47x100) - so add 0.035 to that and we are talking about 1.55++ for blend stability (not to mention actual will be 0.005-0.010v higher with my settings) - and that is just too much for me. I want a day-to-day stable OC, not a temporary benchmark speed.

I think I will abandon 47x for now and focus on (1) finding a rock solid 46x that can blend overnight 8+ hrs so i have a good bse to work from, and then (2) further exploring the space between 46x and 47x via some very very gentle BLCK overclocking. I know BLCK is bad to touch, but considering I can cover the 100mhz gap without even going up to 102 it should be fine. If I'm lucky the curve doesn't start to turn nasty until it's getting close to 47x. We will see

But of course, if anyone has any further insights into how to boost the part of the CPU that blend seems to stress so much - let me know and I will give it another shot

EDIT: One important question. I've always used prime95 as my stability tester - but once I THINK I'm using stable values, what other stability tests are worth checking with? I'm not asking what the harshest synthetic test is (because I think prime95 is unrealistic enough as-is) but just "different" so that I can stress it a few different ways. I'm aware of Intel burn test (but haven't used). That's essentially an easy interface for linpax right?

-Stigma

 

[sm625]

I'm a big time skeptic of memtest. I think its a worthless test personally. It is just not random enough. If you look at the way DDR works, it works most reliably when you are constantly hammering away on it like memtest does. so its no surprise that memtest wont catch many problems. It is better to fill it in blocks (not all at once) and then let the cpu go into a lower power state, and then come back and check it.

Run some games. You will eventually find a game that crashes you rather quickly.

 

[the_stigma]

I'm not really suspecting any defective memory in any case (these two sticks have been in regular use for a few years and not given me any trouble as far as I am aware).

Of course my memory COULD be defective despite me not noticing anything wrong AND 5-pass memtest failed to detect anything wrong with them anyway, but let's not run off on a wild tangent here I think its the CPUs interface to memory that is not able to keep up - not anything wrong with the memory itself. This is just further reeinforced by the fact that as I mentioned in the previous post I can run blend at 46x with an ok voltage (just more than small FTTs). If the memory was kaputt it would hardly matter what I set my Vcore to

-Stigma

 

[flexy]

Small update:

(...) yet it apparently requires a lot more juice for blend. This is a surprise to me - I always thought small-FFT put more direct strain on the CPU. You really just need a lot more Vcore for blend? ... Either that or there is some aspect I am just missing here.


EDIT: One important question. I've always used prime95 as my stability tester - but once I THINK I'm using stable values, what other stability tests are worth checking with? I'm not asking what the harshest synthetic test is (because I think prime95 is unrealistic enough as-is) but just "different" so that I can stress it a few different ways. I'm aware of Intel burn test (but haven't used). That's essentially an easy interface for linpax right?

-Stigma

Well I tried to explain it to you above, maybe read my reply again
Small FFT stability doesn't really mean anything BUT to see whether your temps are still ok.

When you do real stability testing with blend or large, THIS is where it will show whether your system is actually stable. So it's expected that you would need to up Vcore once you run blend (or large) and see that you're not stable. (Note: You need more V with large/blend since the CPU has to shove data across the bus, read from memory etc. while small FFT happens all within the CPU, so it's simply less strain on the CPU/system and needs less voltage)

* You said you tested your mem at STOCK 5x so we can relatively safely assume your mem is ok so I wouldn't mess with this right now UNLESS you find the stable point of your CPU, then can you think about tweaking your mem.

* As for prime (especially the latest) and basically all synthetic benchmarks, nowadays they're ALL extremely far-fetched so I really don't see these tests as "valuable" anymore, unlike in the past, say 5-10 years ago. Sorry, I have to say it's incredible silly say to run Linpack, or prime or whatever synthetic benchmark which might loop some AVX calculations, getting temps (at least for my Haswell) up to 95C or higher.

Those are extremely far-fetched "tests" which have nothing to do with reality whatsoever. As a rule: An overclocked system can ALWAYS show instability at some point anyway. You can run the most extreme burn test for 5 hours...and it crashes at hour 5 and 1 minute. So what?

For more "real life" stability testing I found ASUS "Real Bench 2" actually pretty good (the stress test), it's a highly demanding stress-test but it's not far-out silly like the latest prime or Linpack with AVX instructions that cause stress and temps which you would never see on your PC, not even with heavy gaming, rendering, encoding etc.

That being said, for a quick overall stability-test I still like OCCT large, say for an hour or two.

 

[Deders]

I've had errors appear on faulty sticks in Windows Memory Diagnostic that didn't show up in memtest, worth a try. Also have you adjusted your QPI/VTT voltage?

 

[rgallant]

just saying
on my 3770k when I first started to test the range of it's oc , I down clocked my ram to 1866 so only cpu stalls would show and same as you , the system would fail only when I ran blend .
drove me nuts for a week ,then for the first time I turned the xmp on for the ram and it never failed blend again @ 2400 lol.
not sure if that ram has xmp if it does and not turned on try that.

 

[the_stigma]

I've had errors appear on faulty sticks in Windows Memory Diagnostic that didn't show up in memtest, worth a try. Also have you adjusted your QPI/VTT voltage?

I don't think I've seen settings for those values on my motherboard (remember this is Sandy Bridge)

-Stigma

 

[the_stigma]

Well I tried to explain it to you above, maybe read my reply again
Small FFT stability doesn't really mean anything BUT to see whether your temps are still ok.

Thanks, I think my understanding of the function was just fundamentally wrong. I envisioned that all data fitting into L2 would stress it more (because it never had to wait for anything to continue operations), not less (because it didn't have to work with the memory interface as well).

I got 6hrs prime blend stable @ 46x100 overnight now, but it took 1.420Vcore (something like 1.425-1.430 actual). That's within reason for me, and 4,6Ghz is fine I suppose. From here I will just play around a bit and see if there is an even more ideal point in the curve. If I don't have to sacrifice too much speed to reach 1.400Vcore then that seems like a comfortable place to be for 24/7.

I will check out realbench2 and OCCT as you suggested. How does OCCT large roughly compare to Prime blend in terms of how harsh it is?

-Stigma

 

[Deders]

I don't think I've seen settings for those values on my motherboard (remember this is Sandy Bridge)

-Stigma

Maybe it's called the uncore voltage or something, what voltage settings do you have?

 

[the_stigma]

just saying
on my 3770k when I first started to test the range of it's oc , I down clocked my ram to 1866 so only cpu stalls would show and same as you , the system would fail only when I ran blend .
drove me nuts for a week ,then for the first time I turned the xmp on for the ram and it never failed blend again @ 2400 lol.
not sure if that ram has xmp if it does and not turned on try that.

Nah, my memory for sure doesn't have an XMP profile. It's Kingston value - so best left at stock =P I'm not sure if this generation of motherboard implements XMP in any case (I can't remember having seen it - but then again I haven't messed around with memory, so I might just have missed it).

As far as I understand, XMP is basically just a preconfigured OC-profile for the memory, so I'm not sure why it worked to resolve your problem, but perhaps you had insufficient voltage autodetected for the memory and the XMP profile specified a higher voltage, thus actually making your memory stable despite a higher speed. Just speculation ...

-Stigma

 

[the_stigma]

Maybe it's called the uncore voltage or something, what voltage settings do you have?

Better i just link you the manual than trying to list everything:
http://dlcdnet.asus.com/pub/ASUS/mb/LGA1155/P8Z68-V_PRO/E6696_P8Z68-V_PRO.pdf
see page 73

A lot of things like uncore I don't think really exists anymore on these (relatively) newer systems. I don't think I've seen any motherboard with voltage settings for it on Sandy Bridge at least.

-Stigma

 

[Deders]

It's the VCCIO voltage, try raising it but make sure it's not more than 0.5v either side of your memory voltage or it may damage the controller.

 

[soccerballtux]

Take it one step at a time. Make sure the modules work properly at all first, then worry about whether or not they work properly during a stress/stability test.

Does the blend test still give you error 124 with everything at stock?

yes. you keep dividing the problem in half until you've isolated the variables, then check each variable

 

[the_stigma]

yes. you keep dividing the problem in half until you've isolated the variables, then check each variable

Memory checks out (5-pass memtest at stock, best I know to test it at least)

No trouble blending at stock (or at lower multiplier / higher Vcores)
So it does seem to me that it is the CPU that is having trouble - but the question is just if there is anything other to do about it other than lowering speed / increasing Vcore.

I suppose at this point I've "solved" the issue of instability running blend (by just throwing some more Vcore at it), but my Vcore is a little high'ish to get stability at 46x (1.425 @ extreme LLC, small Vboost at load - or ca 1.435 @ Uhigh LLC, roughly 0 Vdroop at load). If there is anything left to try to achieve better results then I'm very open to suggestions. Otherwise I'll have to settle and accept that my silicone is probably just average-to-belowaverage. Shame since my VID info seemed better than average.

-Stigma

 

[the_stigma]

It's the VCCIO voltage, try raising it but make sure it's not more than 0.5v either side of your memory voltage or it may damage the controller.

I do have VCCIO - which from what I read "boosts the signal from CPU to the memory" (in very laymans terms), so based on that it seems like it could be relevant, but I also read about it that stock values should be more than sufficient up to 2000++ Mhz memory (and I'm only running 1333Mhz), so I wouldn't think it should be necessary.

That said I have only barely tried fiddling with that value - and only up as much as to 1.1 (from 1.05 stock). The range for stable operation apparently isn't very wide on this, and I read that no more than 1.2 is recommended for extreme memory bandwith overclocks.

I'm a little sceptical that this will help me - but suppose I could give it another whirl and see if blend is stable at a notch or two less Vcore than what I have established it needs to complete an 6-hour overnight (if it has no effect it will probably crash in less than an hour, so not so difficult to test). Do you have any recommendations on how high to test setting it though?

-Stigma