Dynamic Overclocking of 980X

[aigomorla]

that's the answer you should have given a while ago :thumbsup:

oh sorry.. i been around lawyers too much.

I dont give more info unless asked... bad habit.

 

[Timmah!]

So...today i checked my bios and as a matter of fact i was unable to find a way to manual control the turbo....aside of option to switch it on and off i see nothing like "turbo boost tech config" mentioned in the other thread dedicated to unlocked turbo regarding MSI boards...i went through whole M.I.T section of the BIOS and its not there
so i suppose Gigabyte boards do no really support this feature...tough luck

Anyway messing around with BIOS i was a little bit tempted to bump the multiplier from 25x to 26x or 27x...if i do this and set vcore to normal while leaving EIST and C1E on, i do risk nothing, right? Apart of stability issues of course...or do i need to mess with all those other voltages - qpi, vtt etc..?
And even though i suppose i should switch turbo off, what will happen if i wont? When default multiplier is set to 26x (instead of 25x), will those 2 turbospeedbins move automatically to 27x and 28x?
@ Aigomorla: I know you think its pointless to overclock it only slightly...but lets just say all i want is to bump my cinebench score from 8,9 to 10

 

[coffeejunkee]

Timmah, that should be completely safe. But concerning the unlocked turbo, I just spoke today to an MSI mobo owner who has an 875k and he didnt have the multipliers. So apparently their mobo's do offer them for a cpu that's not really supposed to have them as far as i know. What you could do is ask Gigabyte if they plan on including it in updated bios.

Sorry to go a bit offtopic, I find this Eist/C-state stuff quite interesting.

This isn't entirely correct. EIST or Speedstep allows the processor to change both it's clock and voltage based on the load, on how much work the cpu is doing. It has nothing to do with idling. It will change the multiplier to any number between 12 and the max, and modify the voltage according to the VID table. This wrecks an overclock done with BCLK, because every clock will be different now, so you can't abide to the VID table anymore. But it can still be used if you OC by multiplier, although I don't see much point in this technology to be honest. It's been shown that you actually save power if you run your processor at full speed even at small load percentage, because the task finishes faster.

Sorry, I was trying to keep things simple, maybe not a good idea on Anandtech I mainly mentioned idle because the effect is easy to notice while idling. Now when I take a look at my bios, C1E is described as: "CPU will enter C1 state and reduce core clock to system bus ratio & VID". EIST is described as: "EIST driver manages clock and VID to best serve the thermal, performance and power requirements"
So this is in line with what you are saying. Except when I disable EIST there is no mutliplier lowering whatsoever.

Here's a screenshot with both EIST and C1E enabled:

Now I disable EIST:

Now I enable EIST again and disable C1E:

I was aware that EIST adjusts the multiplier according to load, except I've never actually seen this happen, it's either x9 or x20 with my mobo (or x21/x24 with turbo enabled). At the most EIST seems a little hesitant to drop to the lowest x9 mutliplier sometimes, it hovers a bit around x10/x12 before settling down. I'm also more interested in power savings during idle.

Now a very interesting test would be to see to how much power saving this actually translates to. I would very much like to see the differences between EIST enabled, C1E enabled and C3/C6/C7 states enabled and various combinations of these. Also it would be interesting to see how much the dynamic vcore saves compared to a fixed vcore under various overclocking situations.

C1E, or enhanced halt state, only kicks in when the OS sends the halt state, which means there must be no instructions sent at all. C1E also refers to the VID table to lower both the multiplier to 12x and the voltage accordingly. C1E is the only state you see while your computer is on. C2 and C3 states are stand-by states if I'm not mistaken, C4 is hybernation, and anything further than that is complete shutdown.

Like i have shown above , the multiplier stays the same, only the voltage is lowered. Only when I enable EIST will it drop. Maybe it works different on other motherboards, maybe cpu-z doesnt read the multiplier correctly. Both options would be interesting to know more about.

EIST and C1E are completely independent from each other. Both have a hard time when you overclock by BCLK (but EIST is a lot worse than C1E) because every multiplier will now give a different clock speed then what the VID table expects.

I do understand how the VID tables get messed up, but for me, EIST doesnt really cause problems when running 20x200. I do need a little extra vcore to stay stable with C1E enabled. C3/C6/C7 however is a no go above bclk 185. But offcourse every cpu/mobo is different.

And yes, the normal setting that your gigabyte board has is unique and very useful, because it allows you really drive up your multiplier to the edge of stability without worrying about automatic over voltage. Most boards have only auto or fixed, at which point you might as well turn C1E off because you ain't saving that much power if the voltage doesn't go down.

Yes, this feature should be mandatory.

 

[Timmah!]

Timmah, that should be completely safe. But concerning the unlocked turbo, I just spoke today to an MSI mobo owner who has an 875k and he didnt have the multipliers. So apparently their mobo's do offer them for a cpu that's not really supposed to have them as far as i know. What you could do is ask Gigabyte if they plan on including it in updated bios.

Thank you coffeejunkee.

So i just did - seems succesfully - my first 980x overclock
As i said i only bumped multi from 25 to 26x and set vcore to Normal...did not touch nothing else except that...
so far it seems ok, i managed to get to Win and write this...cpuz still shows default vcore 0,944 when idle and 1,200 under load...though the multi jumps to 27x...
in a way i actually managed what i wanted, dynamic overclock...i just did not set the turbo to 27x directly, but indirectly by setting the default "rated" multi 25x a bin higher...
btw does the chip with turbo and speedstep ON actually run anytime at its rated frequency...i mean 3,33GHz with 25x stock or in my case now 3,46 with 26x?

Now i need to test it for stability...which soft do you recommend? For my former E8400 i used Prime95, but i see there are more possibilities...Linpack or LinX (is it the same?), Intel BurnTest and some more... Which one is the best...i mean i do not need to run it for 12 hours to find out if it is stable or not? Or should i stick to Prime95?
Anyway i am going now to check how it fares in Cinebench....

 

[coffeejunkee]

Because it's always nice to have to correct yourself twice in the same thread:

EIST does in fact work as advertised. The reason I probably didnt notice is because of the aggressive use of Turbo Mode on my mobo. Without Turbo there are occasionally times when EIST sets the multiplier to x17 or x18. With Turbo enabled the multi is usually set to x24 in these situations, so the EIST effect is gone.
I do agree with you (JAG87) regarding the usefullness of the EIST feature. My cpu is plenty fast enough for normal daily use, even if it runs at the lowest x9 multi all the time (forced by power saver scheme). Notebooks might benefit though.
This doesnt matter too much tough, since I'm only interested in lowering the multi during idle, for which I need EIST enabled anyway.

And yes Timmah, with Turbo enabled, your cpu will probably never run at its rated speed. And now to see how high you can get that multiplier, 30 would be nice. (dont be afraid to up some voltages a little, Intel indicates 1.4V vcore as safe limit. Stay below that in bios and you should be completely safe).

As for stress testing, Linx and Intel Burn Test are the same, they both use Linpack. I use Linx since it's faster and makes my cpu hotter than Prime95. 20 runs Linx with all memory is pretty stable in my book. For this small overclock, you could probably get away with 5 runs as well. And monitor your temps ofcourse: http://www.techpowerup.com/downloads/1691/Real_Temp_3.40.html. 100 degrees celsius is max, but preferably stay under 80 (which will only be reached when upping the vcore, i think you should see temps around 60 depending on your cpu cooler).

Edit: I just stumbled upon this slide:

http://benchmarkreviews.com/index.p...sk=view&id=578&Itemid=63&limit=1&limitstart=1

which seems to confirm the availabilty of unlocked turbo multipliers on the 980x. Or else "increase core ratios via turbo" in combination with the black i7 logo wouldnt make much sense. The blue i7 and i5 logo obviously refer to the i7 875k and i5 655k.
Therefore it seems that the mobo is the culprit. What I would do is send Gigabyte a mail asking them to provide this functionality. Since it's clearly available on at least some Asus mobo's as can be seen here:

http://www.xbitlabs.com/articles/cpu/display/core_i7_875k-core_i5_655K_3.html#sect1

Or just be happy with your current overclock on already the fastest cpu available.

 

[Timmah!]

And yes Timmah, with Turbo enabled, your cpu will probably never run at its rated speed. And now to see how high you can get that multiplier, 30 would be nice. (dont be afraid to up some voltages a little, Intel indicates 1.4V vcore as safe limit. Stay below that in bios and you should be completely safe).

As for stress testing, Linx and Intel Burn Test are the same, they both use Linpack. I use Linx since it's faster and makes my cpu hotter than Prime95. 20 runs Linx with all memory is pretty stable in my book. For this small overclock, you could probably get away with 5 runs as well. And monitor your temps ofcourse: http://www.techpowerup.com/downloads/1691/Real_Temp_3.40.html. 100 degrees celsius is max, but preferably stay under 80 (which will only be reached when upping the vcore, i think you should see temps around 60 depending on your cpu cooler).

Edit: I just stumbled upon this slide:

http://benchmarkreviews.com/index.p...sk=view&id=578&Itemid=63&limit=1&limitstart=1

which seems to confirm the availabilty of unlocked turbo multipliers on the 980x. Or else "increase core ratios via turbo" in combination with the black i7 logo wouldnt make much sense. The blue i7 and i5 logo obviously refer to the i7 875k and i5 655k.
Therefore it seems that the mobo is the culprit. What I would do is send Gigabyte a mail asking them to provide this functionality. Since it's clearly available on at least some Asus mobo's as can be seen here:

http://www.xbitlabs.com/articles/cpu/display/core_i7_875k-core_i5_655K_3.html#sect1

Or just be happy with your current overclock on already the fastest cpu available.

Thanks again for the info:thumbsup:

So i dowloaded the Intel BurnTest and succesfully passed 5 runs at maximum stress level (with most RAM available)...so i am really pleased, now my cinebench score is up to 9,23 from 8,87...
however i do not really like the temps read by RealTempGT during the test...most of the time they hovered around 65 degrees Celsius, but occasionaly on one or 2 core they hit 69...so almost 70 degrees at 3,64Ghz...
I do not have the best cooling solution - Zalman CNPS Performa, so i do hesitate a bit now to bump it more...maybe i will try one bin higher but no more...
regarding unlocked turbo, i am bit disappointed that giga´s flagship board (until UD9) does not support it, but i think let it go...i wanted to bump turbo´s multi and i managed to do it (though in a different way as i thought), but the result is the same, so i am fully satisfied...
one last question, through linpack testing RealTemp kept showing CPU´s usage almost constantly at 50 percent....how is that? Should not be CPU loaded at 100 percent? I want to test it that way, cause during rendering i am going to do it will surely be loaded to full....

 

[coffeejunkee]

Well, I wouldnt worry too much about those temperatures, we're talking 6 cores/12 threads here. I know it sounds high but there really is no danger hitting 80 degrees or more with Nehalem cpu's. If there was, Intel wouldnt set the throttle limit at 100 degrees celcius. The real danger is high voltages. If you look at the ln2 overclock guys, they have their cpu's at -85 degrees or colder but they still die from the ridiculous high voltages they use.

What is a little more troubling is the fact Linx only stresses to 50 percent. It clearly stays at 100 percent when I run it. Linx does slow down between runs, so then it would be normal to see less load but if it never hits 100 percent it seems something is wrong. Maybe check with task manager as well or see what Prime95 does.

 

[Timmah!]

Today i decided to bump again one bin higher to 27x (28x turbo) and again succesfully passed 5 runs of IBT...interestingly temps did not increase at all, actually i think maximum temp on any core was lower than before
BTW i set only vcore to normal, all other voltages there i left at auto....is it right that way? Or should i put all of them to normal too? Do they actually rise more than before (at stock speeds), if i didnt change BCLK or any RAM, QPI, uncore multiplier? I mean i did multiplier overclocking, not bclk overclocking...I need to know this, cause i feel lil unsafe after reading horrible stories of dead gulftowns at XS...

Other thing, Linpack... i was looking more precisely at the CPU usage and it seemed to run at full pelt for few seconds at the beginning and at the end of every run, but in between for the most part it hovered around 50 percent...all 12 threadtabs in task manager looked like zig-zag....so i am not sure what to think about this...

Finally so it seems now it is stable at 28x multi in the linpack...but this is with all cores under load...it can however clock itself to 29x, when only one core (or two, not sure here) is needed...and i did not test this situation for stability...what should i do? Might it be unstable despite all cores@28x are ok? Because probably most apps i use (apart from 3dsmax) will rather run the CPU at this speed....

 

[coffeejunkee]

BTW i set only vcore to normal, all other voltages there i left at auto....is it right that way? Or should i put all of them to normal too? Do they actually rise more than before (at stock speeds), if i didnt change BCLK or any RAM, QPI, uncore multiplier? I mean i did multiplier overclocking, not bclk overclocking...I need to know this, cause i feel lil unsafe after reading horrible stories of dead gulftowns at XS...

I haven't experienced any problems using the auto setting for non-changed voltages during overclocking, but I would indeed recommend setting every voltage to normal or its designated value as shown in the bios. Just to make sure your mobo doesnt mess around according to its own free will. The exception offcourse being the ram, if you have a 1600MHz or faster kit you'll probably have to set vdram to 1.64V and dram termination to 0.82V. One potential risk i could see is the mobo increasing the qpi/vtt voltage but that seems unlikely when doing multiplier oc-ing.

Other thing, Linpack... i was looking more precisely at the CPU usage and it seemed to run at full pelt for few seconds at the beginning and at the end of every run, but in between for the most part it hovered around 50 percent...all 12 threadtabs in task manager looked like zig-zag....so i am not sure what to think about this...

Well i'm not sure what to make of this. Have you checked if this happens on stock settings as well? And when running Prime95 does taskmanager show all cores at full load? One possiblity that came to my mind is turbo throttling: http://www.xtremesystems.org/forums/showthread.php?t=225450. But in that case, it just shouldnt use the turbo multiplier, not run at only 50% load. Maybe see what happens when you set minimum and maximum processor state to 100% in Windows power options? I would also run 3D mark vantage to see if your cpu score is normal. Maybe I'm looking into it too much, but again, when I run Linx, all 4 cores are loaded 100% all the time until the run finishes.

Finally so it seems now it is stable at 28x multi in the linpack...but this is with all cores under load...it can however clock itself to 29x, when only one core (or two, not sure here) is needed...and i did not test this situation for stability...what should i do? Might it be unstable despite all cores@28x are ok? Because probably most apps i use (apart from 3dsmax) will rather run the CPU at this speed....

Set number of threads to 1 in Prime95 or Linx. A Superpi 32M run might be usefull too, albeit in my experience it's more prone to memory instability.

 

[Timmah!]

@cofeejunkee
Well, i guess i will try Vantage...do you know what is the score for 980x at 3,78GHz? so i could compare...
BTW i did not try Linpack on default clock, so i cant tell and now i am too lazy to downclock it back Anyway the Cinebench score was rising as expected...was 8,87 at default, 9,32 at 3,64 and is 9,54 now....so this looks as expected and CPU is loaded to full during the test....
Personally i have a feeling its working as it should despite this little problem, so far i havent got opportunity to do something productive with it (i mean some rendering), but when it comes, i will certainly pay attention to CPU usage and performance...
Regarding voltages and RAM, i have 6x2GB 1333 Kingston HyperX sticks at default CL9 latency...not sure about voltage now, but i did not mess with it, so it should be at default...there is a XMP profile, which should take the RAM to CL7, but i hesitate to use it, cause i do not have matched kit and this was actually meant for 6GB, not 12GB....

 

[coffeejunkee]

Those Cinebench scores seem alright to me, I only get 3.76 at stock

3D Mark score should be around 32k: http://service.futuremark.com/resultComparison.action?compareResultId=2391977&compareResultType=19

Concerning ram: the XMP profile will probably use 1.65V. I dont have much experience with xmp though and I've seen it work differently with different processors (changing the bclk too fex). If you havent changed anything your ram will probably run at 1066MHz (and require 1.65 for 1333). But I can't really tell for sure without a product code. 1333MHz might actually be asking too much with 12GB. Maybe you will also need to increase qpi/vtt in that case.

 

[Timmah!]

Yeah, those scores are fine, i saw screens around internet of other gulftown getting same numbers.
Just checked cpuz and it shows my RAM running at 675 Mhz with 2:10 multi and 9-9-9-24 latency...and i really did not touch it in BIOS...i am actually surprised, i thought it runs at 1066 by default as well, at least i remember reading about gulftowns long before they were launched that they are still going to officialy support 1066Mhz like Bloomfield, despite Lynnfield CPU´s supporting up to 1333 already...so i suppose they changed it...
I could try the XMP or manually try to increase the speed or decrease latency, but as you said, that will probably require more qpi/vtt voltage and for now i would like to avoid that....
BTW thanks for those vantage numbers, tomorrow i will download vantage and test it...

 

[coffeejunkee]

That's interesting indeed. According to Intel it should run at 533MHz using 2:6 ratio. And like you, I understood this from the reviews as well. Maybe Gigabyte decided it wont hurt to go a little faster.

Are you sure you dont have 1600MHz ram btw?

 

[aigomorla]

whats really funny is your comparing a standard i5 cpu against a 980X cpu.

Whats even more funny is the 980X is faster clock per clock then all the other current lineup minus westmere-EP.

Whats also even more funny is your listening to someone who doesnt own a gulftown, and trying to follow him debug on his i5.

i5 is NOT a i7 980X.
The i5 turbo mode is nothing like the 980X.
The 980X is in its own catigory in intel cpu's because its a HEXCORE not a QUAD or DUAL CORE.
With a VERY LARGE cache supply of 12MEGS, not 8MEGS, or 4 MEGS...

Im really lost in why ur trying to debug your oc off values on an i5.

You cant even play with VTT on an i5 while ur suposed to play with VTT on a 980X.
The MAX Power Draw values between the 2 platforms is even different.

Socket 1156 vs Socket 1366.

At least get a 920 user to help ya out.
Because he knows what VTT is.

CoffeeJunkie... the 980X has a max safe voltage value of these:

Vcore: 1.35
VTT: 1.325
QPI: 1.3
DRAM: 1.66

Why is dram so high on an i7? because we run tri channel.
As you go higher in speed + Capacity, you will notice you will need more VTT + DRAM voltage.

Vs. a 1156 max:

Vcore: 1.35
VTT: 1.275 <--- which is why i dont like 1156.
Dram: 1.55-1.6v <--- u shouldnt need more then 1.55 for dual channel.

Unless you played with a 980X, its not wise to have him compare his OC with yours.
As the two chips are totally different down to the platform.

 

[coffeejunkee]

Relax mate. I'm not comparing anything, just trying to help. His original question about turbo-multipliers hasn't been resolved. Apart from that, he wanted to know how to make his vcore and mutliplier drop during idle, and that's what I explained.

The only reason i'm in this thread is basically because I seem to be one of the few people who overclock their cpu's and are also interested in power efficiency. Instead of the many people that will tell you to: "just disable eist you want your cpu running at full speed all the time."

 

[aigomorla]

Relax mate. I'm not comparing anything, just trying to help. His original question about turbo-multipliers hasn't been resolved. Apart from that, he wanted to know how to make his vcore and mutliplier drop during idle, and that's what I explained.

The only reason i'm in this thread is basically because I seem to be one of the few people who overclock their cpu's and are also interested in power efficiency. Instead of the many people that will tell you to: "just disable eist you want your cpu running at full speed all the time."

oh i wasnt saying anything up to that point.

You were actually doing a great job.

it was until he asked you to start comparing vantage scores and whatnot.

Thats where the problem happened.

 

[Timmah!]

whats really funny is your comparing a standard i5 cpu against a 980X cpu.

Whats even more funny is the 980X is faster clock per clock then all the other current lineup minus westmere-EP.

Whats also even more funny is your listening to someone who doesnt own a gulftown, and trying to follow him debug on his i5.

i5 is NOT a i7 980X.
The i5 turbo mode is nothing like the 980X.
The 980X is in its own catigory in intel cpu's because its a HEXCORE not a QUAD or DUAL CORE.
With a VERY LARGE cache supply of 12MEGS, not 8MEGS, or 4 MEGS...

Im really lost in why ur trying to debug your oc off values on an i5.

You cant even play with VTT on an i5 while ur suposed to play with VTT on a 980X.
The MAX Power Draw values between the 2 platforms is even different.

Socket 1156 vs Socket 1366.

At least get a 920 user to help ya out.
Because he knows what VTT is.

CoffeeJunkie... the 980X has a max safe voltage value of these:

Vcore: 1.35
VTT: 1.325
QPI: 1.3
DRAM: 1.66

Why is dram so high on an i7? because we run tri channel.
As you go higher in speed + Capacity, you will notice you will need more VTT + DRAM voltage.

Vs. a 1156 max:

Vcore: 1.35
VTT: 1.275 <--- which is why i dont like 1156.
Dram: 1.55-1.6v <--- u shouldnt need more then 1.55 for dual channel.

Unless you played with a 980X, its not wise to have him compare his OC with yours.
As the two chips are totally different down to the platform.

Well, i actually appreciate coffeejunkee´s help, he was certainly very helpful...btw if you believe that advices he gave me, are for some reason wrong, then i will be grateful to you, if you correct him...
regarding "getting someone at least with 920 to help", just tell me how...i made a few questions but i have no influence on people, if someone does not want to answer them, what can i do....you alone with gulftown could answer them anytime, instead now you wonder, why i do listen to him...
BTW those voltages are interesting, but in my case irrelevant, cause i do not want to bump any of them...so far i managed it, cause i run now at 3,78 with default vcore...do i need to "play with VTT" when i only bumped multi from 25x to 27x? I left it on auto....()

 

[coffeejunkee]

What do you mean? I linked to the score of a stock i7 980X...or is there something wrong there?

Concerning qpi/vtt, you're right I was a bit careless there. In typical mobo manufacturers fashion this setting has multiple names across brands/chipsets, preferably with as little or vague explanation as possible. It's called qpi/vtt on GB P55 mobo's but I believe Intel themselves just calls it vuncore iirc. Afaik know Bloomfield/Gulftown has an uncore as well.

Now Lynnfield doesnt have a qpi like Bloomfield so the Gigabyte name is confusing. I didnt mean to imply that you should raise qpi and vtt, only the latter. Which might be necessary when running 12GB of ram, especially when overclocked it. But no need to touch it when everything works fine.

btw if you believe that advices he gave me, are for some reason wrong, then i will be grateful to you, if you correct him...

so will I

 

[Timmah!]

What do you mean? I linked to the score of a stock i7 980X...or is there something wrong there?

Concerning qpi/vtt, you're right I was a bit careless there. In typical mobo manufacturers fashion this setting has multiple names across brands/chipsets, preferably with as little or vague explanation as possible. It's called qpi/vtt on GB P55 mobo's but I believe Intel themselves just calls it vuncore iirc. Afaik know Bloomfield/Gulftown has an uncore as well.

Now Lynnfield doesnt have a qpi like Bloomfield so the Gigabyte name is confusing. I didnt mean to imply that you should raise qpi and vtt, only the latter. Which might be necessary when running 12GB of ram, especially when overclocked it. But no need to touch it when everything works fine.

Do not worry...i would rather run my chip at stockspeeds, than touching those voltages
From what i read IMC seems to me like the most vulnerable part of the chip, so i am not going to mess around with VTT, especially when my whole previous overclocking "career" is only about getting E8400 to 3,6 GHz
not on 950EUR chip anyway

 

[aigomorla]

What do you mean? I linked to the score of a stock i7 980X...or is there something wrong there?

oh im terribly sorry, i thought you were comparing scores with a 3.7 i5.

I misread your post..

And yes the IMC on gulftown is a glass slipper, and on the A0's it was notorious for dying for unknown reasons.
(i have a dead one to prove it.. lol)

B0's were fixed to be a bit more stable, and B1's are the full retail versions most people have.

B1's also have a beef'd up IMC like Xeons, which makes them a tad bit faster clock per clock.

Its not something ur eyes will notice, but number wise its there.