Intel Haswell Overclocking Clubhouse.

[Womper]

Why are you using Prime 95 for stress testing. I thought it didn't officially support Haswell quite yet and the AVX2 instruction?

Haswell can do older AVX as well as AVX2. Prime 95 just doesn't exercise the AVX2 instructions that yield better performance per clock. The latest version of linpack, however, does use the new AVX2 instructions, and it generates more heat than anything I've seen.

But the motivation is simple: Prime95 is a really quick way to see if something is "possibly" stable, and worth exploring. We all have our favorite quick-stress test for this.

Lately I've been using the Intel Extreme Tuner's built-in stress test for long-term. It doesn't use AVX code, so it seems to be a good stress test for someone that just wants to play games at max GHz without any crashes.

 

[Kursah]

Well at this point it seems that I can get 4.4 stable at 1.36v, and 4.3 stable at 1.25v. Still testing though. I am finding that Wargames: AirLand Battle uses 8 threads and may push my CPU more than BF3 game-stress-wise. But after AIDA and Prime that's where I went next and so far so good.

I will say the 4.4 w/1.36v, temps can get a up into the low 80's while gaming. Which compared to some chips might not be too shabby? I dunno...seems decent on air to me for a haswell chip at that volts on air. Dunno if it's worth the extra 100mhz tho for all that extra temp and voltage.

 

[TheHunter]

Here is my hopefully final 4.6Ghz@ 1.220v bios setting, it looks ok in almost everything atm

 

[Womper]

Here is my hopefully final 4.6Ghz@ 1.220v bios setting, it looks ok in almost everything atm

Thanks for sharing! My input voltage is 1.8 when I use Auto rules, interesting that your MB uses such a lower value.

Also, I played around with Adaptive some more and found two things:
1. Setting CPU core voltage offset to "Auto" when using adaptive will be the same as 0.0v. So no need to have a +0.001 offset. Double check the Total Adaptive Mode CPU Core Voltage box- it should equal the Additional Turbo Mode voltage.
2. Adaptive mode will only boost voltage if you're running faster than 3.9GHz (yes, you knew that already). Even if you sync all cores to 39x (which is non-stock behavior). Furthermore, if you're trying to add voltage to the default cache speed of 39x, you can't use Adaptive mode for it- it needs to be at least 40x to use Adaptive.

Even if I stay at 39x cache speed, I still have to bump up cache voltage when going after core overclocks. So use manual or offset with 39x.

 

[TheHunter]

About cpu vrm @ 1.70v, in uefi it says its better to leave off when OC'ing so its not raising it further. Otherwise yes its the same here ~ from 1.71 to 1.79v

Same recommendation by system vrm spread spectrum and by cpu power management - CPU integrated VR fault management, but i left both at auto.
By cpu I've put extra normal offset +0.002v just in case



I didnt play with adaptive cache offset yet, but lets say i leave it at 42x, will it be stable at lower voltage?
Or will I have to set min 39x so its stable at default voltage?

 

[Womper]

About cpu vrm @ 1.70v, in uefi it says its better to leave off when OC'ing so its not raising it further. Otherwise yes its the same here ~ from 1.71 to 1.79v

Same recommendation by system vrm spread spectrum and by cpu power management - CPU integrated VR fault management, but i left both at auto.
By cpu I've put extra normal offset +0.002v just in case



I didnt play with adaptive cache offset yet, but lets say i leave it at 42x, will it be stable at lower voltage?
Or will I have to set min 39x so its stable at default voltage?

I'm not sure what you mean by leaving cpu vrm off. I can either leave input voltage on Auto (which will be 1.8v) or specify an absolute voltage value of my own. Your pic has it on Auto.

Adaptive Cache Offset acts identically to the normal Offset mode. It raises all voltage bins by the specified amount. This is helpful if you need to increase the voltage for the 39x and less speed-steps for some reason. I can't really think of a reason though- if you set to 42x for example, then you're covered by the Additional Turbo Voltage value during load. And if you're idling, then you probably won't have stability issues anyways since everybody speed-stepped down to 800Mhz.

Adaptive mode as a whole won't increase stability, it just eliminates the undesirable side-effect of Offset voltage- which is when you speed-step down to 800MHz for idle, you're still going to be adding that Offset to your voltage, which needlessly consumes power and stuff.

 

[cadaveca]

I can either leave input voltage on Auto (which will be 1.8v)

1.8V is not "stock". it varies from chip to chip as well as being dynamic based on load, which further complicates OC'ing. That's why some boards offer the override feature, which is what he is talking about, most likely. disabled follows board maker rules, enabled follow Intel rules.

I kept one chip 1.7V, one chip 1.75V. I noticed a couple boards set 1.8V default, but not all do. I think Intel spec is 1.65 V to 1.85V at stock, 1.65 V for C-states, up to 1.85V at high load, too, so I'm not sure why my two chips read this differently.

 

[Kursah]

1.8V is not "stock". it varies from chip to chip as well as being dynamic based on load, which further complicates OC'ing. That's why some boards offer the override feature, which is what he is talking about, most likely. disabled follows board maker rules, enabled follow Intel rules.

I kept one chip 1.7V, one chip 1.75V. I noticed a couple boards set 1.8V default, but not all do. I think Intel spec is 1.65 V to 1.85V at stock, 1.65 V for C-states, up to 1.85V at high load, too, so I'm not sure why my two chips read this differently.

The disable feature for CPU Input Voltage on my Asus Z87-Pro is described pretty much as not allowing the CPU VRM to control what setting for the MB's Voltage to the CPU...but does not tell you if keeping it on AUTO changes anything. Mine seems to be at 1.712 atm.

I had mine manually set to 1.70 and it still read 1.712. I had read a couple of Haswell OC guides saying you may need up to 1.90V or more to get a stable high OC and that you want .4-.5v more than what you're feeding your CPU. And that while raising this voltage may help with OC stability it may also hinder it as more voltage into the CPU = More Heat which can just as quickly kill stability.

I wish there was a program that monitored that voltage variable via Windows...it would be interesting to see just how much it really changes. Might have to re-install Ai Suite...though I'd rather not. It's a neat program..but I prefer BIOS myself.

I'm also curious and eager to see what the next BIOS updates will bring.

Oh ya and another note, Crazyeyesreaper posted a couple pages ago in this thread about cooler orientation on these chips. Well I have tested my Noctua NH-U14S both with the standard horizontal to the rear exhaust and the vertical to the top fan exhaust in my current case...the Lian Li Lancool PC-K62 (love this case! wish they still made 'em!). What to keep in note here is in standard mount my cooler is aimed directly at the rear 120mm fan, which can move a decent amount of air for the stock fan. All fans are stock in my case atm. The top has two 140MM fans that while quiet do move a good amount of air. My previous build I had the Xig Gaia aimed veritcally thinking that was best..just like I did in my HAF 932 before that.

It's a hot part of the day...ambient temps are around 32-33C in the room I'm in atm..no A/C. I ran vertical exhaust since last night...and was hitting 99-100C pretty quick into Intel's ETU's CPU test and if I ran it too long, say 20 minutes I would get a BSOD 124 and restart (this is with my recent 4.4GHz @ 1.36v). Well I reset the orientation to the standard out the back. I just ran a 5 minute test, and the peak temp was 90C with the hottest core hitting 92C. With the vertical exhaust I was hitting 99-100C on cores and CPU readings. I think that's a pretty extreme difference when I was a believer like Crazy that a vertical exhaust would be best...maybe not the case. Maybe that's with this case? Maybe that rear 120mm is more effective than the two 140mm's? Maybe the cooler is designed more for the standard mount style though it's easily interchangeable between orientations...fastest cooler swap EVER. gonna go try it on the 20 minute test after playing some Wargame: AirLand Battle. Sweet game, very awesome stability test for your PC too.

More to come!

Had a great and first 1v1 MP match, and it was a riot! Max temps while in-game during this hot evening was between 75c for the coolest core and 80c for the hottest. Where I was hitting lower to mid 80's with a vertical exhaust. Again, could be a mix of case and stock fans.

 

[Womper]

The disable feature for CPU Input Voltage on my Asus Z87-Pro is described pretty much as not allowing the CPU VRM to control what setting for the MB's Voltage to the CPU...but does not tell you if keeping it on AUTO changes anything. Mine seems to be at 1.712 atm.

Oh, gotcha. SVID Control is the option we're talking about. Auto probably = disabled as soon as you increase the multiplier, but either way, 1.8v seems reasonable for my needs. I meant to note that 1.8 is my stock voltage, since I figured every chip is different. Seems like a pretty large range though.

I had mine manually set to 1.70 and it still read 1.712. I had read a couple of Haswell OC guides saying you may need up to 1.90V or more to get a stable high OC and that you want .4-.5v more than what you're feeding your CPU. And that while raising this voltage may help with OC stability it may also hinder it as more voltage into the CPU = More Heat which can just as quickly kill stability.

Does the input voltage really increase heat by itself? If so, what kind of numbers are we talking about? Or does that too vary per-chip?

 

[PolRoger]

I wish there was a program that monitored that voltage variable via Windows...it would be interesting to see just how much it really changes. Might have to re-install Ai Suite...though I'd rather not. It's a neat program..but I prefer BIOS myself

HWiNFO will show VCCIN voltage provided that your motherboard has a sensor for it. You can customize the program to show as little or as much as you like.

The LLC control on the motherboard (for Haswell class) is now applied to the CPU Input voltage. ASUS auto rules seem to default to Level 8 (100%) which actually applies a little overvoltage to VCCIN under load. I've been running with mine set to Level 7 which seems to have the least variance between idle and load. Lower level settings provide more droop to VCCIN under load.

Oh, gotcha. SVID Control is the option we're talking about. Auto probably = disabled as soon as you increase the multiplier, but either way, 1.8v seems reasonable for my needs. I meant to note that 1.8 is my stock voltage, since I figured every chip is different. Seems like a pretty large range though.

Does the input voltage really increase heat by itself? If so, what kind of numbers are we talking about? Or does that too vary per-chip?

It took me a while to figure why his was lower as well...(SVID Control setting).

I've left my setting mostly to default on auto but I've also experimented with manually lowering the CPU Input voltage. I think the optimal settings for this will vary by chip as well as to what actual overclock settings you are trying to run... I just disabled SVID Control in BIOS and mine also dropped to ~1.712v under load but I then I started getting random BSOD to my "stable" 47x43x 2133C9 overclock... I've now manually bumped CPU Input voltage up to ~1.75v and the BSOD's have seemed to have gone away...

In the past I remember experimenting with a lower voltage 44x overclock while running CPU Input down around ~1.65v. I suspect that running higher CPU Input voltages above ~1.8v will also impact load temps when running higher multi overclocks? I suppose that this could be offset if by running a higher CPU input voltage it would also allow you to lower vcore and still achieve stability?

 

[TheHunter]

Ah sorry yes I meant SVID, in Aida64 its described as CPU VRM

I noticed if its at auto it changes a bit max i saw was 1.85v, by disabled its always at fixed 1.70v

And small update, 1.220v wasnt 100% stable (bsod 0x124 in BF3), im now at adaptive 1.223v - this passed BF3 64player TDM test and the rest..

Level 7 which seems to have the least variance between idle and load. Lower level settings provide more droop to VCCIN under load.



It took me a while to figure why his was lower as well...(SVID Control setting).

I've left my setting mostly to default on auto but I've also experimented with manually lowering the CPU Input voltage. I think the optimal settings for this will vary by chip as well as to what actual overclock settings you are trying to run... I just disabled SVID Control in BIOS and mine also dropped to ~1.712v under load but I then I started getting random BSOD to my "stable" 47x43x 2133C9 overclock... I've now manually bumped CPU Input voltage up to ~1.75v and the BSOD's have seemed to have gone away...

In the past I remember experimenting with a lower voltage 44x overclock while running CPU Input down around ~1.65v. I suspect that running higher CPU Input voltages above ~1.8v will also impact load temps when running higher multi overclocks? I suppose that this could be offset if by running a higher CPU input voltage it would also allow you to lower vcore and still achieve stability?

I've read somewhere 75% LLC is the way to go with adaptive offset. Im at 6 atm, but like you noticed 7 is probably the best choice.


About SVID I think it has to be ~0.5v gap between input and cpu voltage.. So I at ~1.225v can still get away with 1.70v..
1.75v would be ideal, but if i manually put 1.75v its in the red zone and this kinda scares me lol, should i set to 1.75v anyway?


Also yes higher values will heat up, just like higher cache volts.

 

[PolRoger]

About SVID I think it has to be ~0.5v gap between input and cpu voltage.. So I at ~1.225v can still get away with 1.70v..
1.75v would be ideal, but if i manually put 1.75v its in the red zone and this kinda scares me lol, should i set to 1.75v anyway?

I'm not really sure I follow you about the 1.75v to 1.225v being in the red zone?

I've seen it mentioned in this guide:
http://www.overclock.net/t/1401976/the-gigabyte-z87-haswell-overclocking-oc-guide

"Too high of a VIN can result in added heat while setting VIN too low can cause instability and VCore droop if the VIN is 0.1-0.2v away from the VCore. Intel recommends keeping VIN Override at least 0.4v above the VCore."

I'm not sure if this is some sort of firm rule or if it can vary by individual chip and overclock settings?

 

[Kursah]

Ah sorry yes I meant SVID, in Aida64 its described as CPU VRM

For some reason I'm not seeing all those voltages in my Aida? Maybe it's related with me removing AiSuite? I remember them showing up...but maybe windows has crashed enough times something else is bugged...

 

[TheHunter]

^
Try latest beta, im using beta 3.255 there
http://www.aida64.com/downloads/aid...64&utm_medium=update&utm_campaign=betaproduct

I'm not really sure I follow you about the 1.75v to 1.225v being in the red zone?

I've seen it mentioned in this guide:
http://www.overclock.net/t/1401976/the-gigabyte-z87-haswell-overclocking-oc-guide

"Too high of a VIN can result in added heat while setting VIN too low can cause instability and VCore droop if the VIN is 0.1-0.2v away from the VCore. Intel recommends keeping VIN Override at least 0.4v above the VCore."

I'm not sure if this is some sort of firm rule or if it can vary by individual chip and overclock settings?

lol I said it wrong, in UEFI it shows red color anything over 1.50v, although now with v1205 its black again.


Ah thanks so its a 0.4v gap.

 

[Kursah]

That fixed it! I can now see voltages.

 

[cadaveca]

Ah thanks so its a 0.4v gap.

I never did find this info in Intel documents. I did look for it, and did find it in ASUS guides, but not in Intel info.

Some of the other info that was related in such guides has been changed, too, so I do not think this is a "hard" rule..it's something that LN2 users want to look at when running high vcore, which none of us will do. Kinda frustrating for me to see 24/7 guides written that are based off of extreme-oc guides, where much of the info for that just simply doesn't relate.

I have my own guide in progress still; it's mostly done, just making benchmark screenshots the past few days for that and a few other reviews I am working on. There actually isn't much different in my guide than others, honestly, other than different voltage values, and the whole keeping cpu cache @ 39 thing, rather than trying to keep it in sync, which only seems to benefit benchmarks, and not all of them, either.

 

[petedread]

I need at least 1.860v to have cpu and cache at 4.3. CPU 1.225v. On a gigabyte board.
(Edit) My cache is static set at anything other than 35x. Set it at 39x, static, it doesn't boost to my OC or drop when idle.

 

[Womper]

I've left my setting mostly to default on auto but I've also experimented with manually lowering the CPU Input voltage. I think the optimal settings for this will vary by chip as well as to what actual overclock settings you are trying to run... I just disabled SVID Control in BIOS and mine also dropped to ~1.712v under load but I then I started getting random BSOD to my "stable" 47x43x 2133C9 overclock... I've now manually bumped CPU Input voltage up to ~1.75v and the BSOD's have seemed to have gone away...

Hmm I think I'll try disabling SVID. Maybe my CPU sometimes lowers the input voltage too much or something, causing those long term BSOD.

What voltage mode are you using on your 47x43x 2133c9 overclock? Is that a static 1.26-ish voltage, offset, or adaptive? We might have really similar chips- I'm running 47x45x 2133C9 with 1.275v adaptive mode on the core and the cache. In terms of manual/static voltage, this needs 1.325v to support AVX tests. I've taken it down to a static 1.28v, and it can handle AVX for several minutes before crashing.

Edit: Last night I decided to explore 4.9GHz a bit more. 1.39v static voltage, cache at 39x with a +0.1v cache offset, 1600C9. IET stress test blue screened after 4 minutes...darn. 5GHz boots with the same, but freezes fairly quickly after I get to my Windows desktop. So...how much voltage can the core take again? Is 1.42v a bad idea? What do these chips die at?

 

[Kursah]

I never did find this info in Intel documents. I did look for it, and did find it in ASUS guides, but not in Intel info.

Some of the other info that was related in such guides has been changed, too, so I do not think this is a "hard" rule..it's something that LN2 users want to look at when running high vcore, which none of us will do. Kinda frustrating for me to see 24/7 guides written that are based off of extreme-oc guides, where much of the info for that just simply doesn't relate.

I have my own guide in progress still; it's mostly done, just making benchmark screenshots the past few days for that and a few other reviews I am working on. There actually isn't much different in my guide than others, honestly, other than different voltage values, and the whole keeping cpu cache @ 39 thing, rather than trying to keep it in sync, which only seems to benefit benchmarks, and not all of them, either.

Dave! I can't wait to read your guide and apply your findings and knowledge to my build!

I have tried messing with cache above 39 and it's proven nothing buck lackluster or not noticable as you've stated. I really haven't ran many benchmarks beyond a 3DM11 or AIDA64 Cache and Memory Benchmark where even going from 39-41 I think was minimal on change. I will say I did read elsewhere in a guide I stumbled upon by google the other night...and a few thread posts that most recommend to run a 39-41x cache no matter what OC. But I find staying at 39 seems to keep overall stability in check.

Hey TheHunter thanks again for the link..I love this program but I don't keep up with the betas. I find that using the Intel Extreme Tuning Utilty CPU Stress Test that my CPU...I've grown fond of this app. Between that and Wargames AirLand Battle I feel stability is at least more enjoyable to find! Unless I BSOD in the middle of a good skirmish! Doh!

But I noticed that during idle, my CPU VRM reading is right at 1.712v, same as reported by BIOS. Then during load via Intel ETU, it goes up to 1.728v and stays there. Then a few seconds after the test is done and everything is cooling down I see it drop to 1.696v for a couple seconds before levelling out at 1.712v. I am currently on AUTO for the value, but the setting for CPU control of this value is set to disabled. I wonder if it floats around like that with a manual setting as a vDroop situation or if it's voltage adjustments? Seems small enough that I almost doubt it's the chipset purposely kicking up to 1.728v but I could be way off base here.

EDIT: So that's another thing I'm sure I've read about but don't remember and haven't tried... how different can the min and max variables for the cache multi be? Like say can you do a min of 8 to match the CPU and a max of say 39 and have it scale? Does it really matter? Does that hinder stability? Does it help?

 

[Womper]

EDIT: So that's another thing I'm sure I've read about but don't remember and haven't tried... how different can the min and max variables for the cache multi be? Like say can you do a min of 8 to match the CPU and a max of say 39 and have it scale? Does it really matter? Does that hinder stability? Does it help?

The difference between cache min and max doesn't matter in my experience. I'm at 8x min, 45x max and it works fine. No difference in stability from when I left Min at Auto (which chose 39x), at least so far. Gotta let that cache take a break too!

 

[Kursah]

Maybe you've already posted this, but can you tell if it's in fact switching clocks like the cores do? What software is reading that speed for ya if you can read it? I would assume that is the case...I just set my min to 8x as well, and 39x as max. You have sweet chip man!

 

[PolRoger]

Hmm I think I'll try disabling SVID. Maybe my CPU sometimes lowers the input voltage too much or something, causing those long term BSOD.

What voltage mode are you using on your 47x43x 2133c9 overclock? Is that a static 1.26-ish voltage, offset, or adaptive? We might have really similar chips- I'm running 47x45x 2133C9 with 1.275v adaptive mode on the core and the cache. In terms of manual/static voltage, this needs 1.325v to support AVX tests. I've taken it down to a static 1.28v, and it can handle AVX for several minutes before crashing.

Edit: Last night I decided to explore 4.9GHz a bit more. 1.39v static voltage, cache at 39x with a +0.1v cache offset, 1600C9. IET stress test blue screened after 4 minutes...darn. 5GHz boots with the same, but freezes fairly quickly after I get to my Windows desktop. So...how much voltage can the core take again? Is 1.42v a bad idea? What do these chips die at?

I'm now testing 47x44x 2133C9... It seems that I'm able to run cache at 44x with just a slight increase to cache voltage over 43x. Both CPU and Cache are set to fixed voltage... energy savings settings are actually a moot point for me because my chip pretty much always runs at 100% load.

BIOS Settings:

47x/44x 100bclk:

CPU Core: 1.259v
CPU Cache/Ring: 1.237v
CPU S/A: +.100
CPU Digital I/O: +.100 (Auto)
CPU Analog I/O: +.000 (Auto)
DRAM: 1.525v
CPU Input: 1.792v (Auto)

Just this week I revisited 48x again and when under load at 48x my chip always results in BSOD's. The instability was severe enough to corrupt/crash my Rosetta data set on my BOINC software. :shadedshu

 

[TheHunter]

Ok thanks for min cache info, I've set it to
min 8x
max 41x
total adaptive offset 1.120v

And it jup works the same way like cpu, no issues.

 

[Womper]

Maybe you've already posted this, but can you tell if it's in fact switching clocks like the cores do? What software is reading that speed for ya if you can read it? I would assume that is the case...I just set my min to 8x as well, and 39x as max. You have sweet chip man!

You can actually see this with CPU-Z on one of the tabs in the NB Frequency box. Or you can use Aida64 and others I'm sure.

I'm now testing 47x44x 2133C9... It seems that I'm able to run cache at 44x with just a slight increase to cache voltage over 43x. Both CPU and Cache are set to fixed voltage... energy savings settings are actually a moot point for me because my chip pretty much always runs at 100% load.

BIOS Settings:

47x/44x 100bclk:

CPU Core: 1.259v
CPU Cache/Ring: 1.237v
CPU S/A: +.100
CPU Digital I/O: +.100 (Auto)
CPU Analog I/O: +.000 (Auto)
DRAM: 1.525v
CPU Input: 1.792v (Auto)

Just this week I revisited 48x again and when under load at 48x my chip always results in BSOD's. The instability was severe enough to corrupt/crash my Rosetta data set on my BOINC software.

Does your 1.259v core work for AVX loads?

And yeah, I've noticed with my chip's cache that 46x or greater is just a wall. 45x is fine with 1.25v, but 46x needs 1.325 or possibly more (it may have been a culprit for lack of long-term stability). 47x isn't happy with even 1.375v.

As for the core, 4.8GHz is possible, but it needs such a high voltage for non-AVX that Adaptive and Offset risk boosting voltage into the mid 1.4v's if a stray AVX instruction enters the pipe. Something like 1.35v-1.375v should be stable for games (heat is rarely a problem with games).

I'm super glad I bought this swiftech H220, I think it's fantastic. Definitely lets me stress test previously impossible voltages. I completed 3 hours of that IET stress test at 1.275v adaptive, and the max temp on the hottest core was 80C. Ambient temp was ~24c. Idle temps are barely above ambient, so this thing is sweet.

 

[PolRoger]

Does your 1.259v core work for AVX loads?

And yeah, I've noticed with my chip's cache that 46x or greater is just a wall. 45x is fine with 1.25v, but 46x needs 1.325 or possibly more (it may have been a culprit for lack of long-term stability). 47x isn't happy with even 1.375v.

As for the core, 4.8GHz is possible, but it needs such a high voltage for non-AVX that Adaptive and Offset risk boosting voltage into the mid 1.4v's if a stray AVX instruction enters the pipe. Something like 1.35v-1.375v should be stable for games (heat is rarely a problem with games).

I'm super glad I bought this swiftech H220, I think it's fantastic. Definitely lets me stress test previously impossible voltages. I completed 3 hours of that IET stress test at 1.275v adaptive, and the max temp on the hottest core was 80C. Ambient temp was ~24c. Idle temps are barely above ambient, so this thing is sweet.

I haven't really tested it with AVX... I sometimes run brief Prime AVX runs as a quick check to see if an overclock will BSOD or freeze etc. I usually test for long term stability by "crunching" for 24/48(+) hrs. I would think that it would probably not be long run AVX stress stable.

With my chip... 44x cache seems to be the high range and 45x more like approaching a wall. I've never even attempted to do 46x...

I really think something is up with this particular sample when trying to run core multi @48x on ambient cooling. It just doesn't seem to be able to do it. :shadedshu But the chip runs pretty well using 44x thru 47x... So I can't get too upset about it.