Skylake hard locks seeminly when IDLE / CStates ON

[GGlyn]

Myself and others have been struggling with a different hard lock problem with Skylake and we have ran out of options but to assume it is a CPU fault/issue. Does anyone have any comments or feedback?

The long thread is over here, but I will summarise. http://www.tomshardware.co.uk/forum/id-2830772/skylake-build-randomly-freezing-crashing/page-7.html New Skylake Build Randomly Freezing/Crashing - Page 7 - windows 10 - Windows 10

Our systems hard lock randomly but appears on the most part to be when IDLE or next to IDLE. The hardlock can come when playing video, browsing (reading) or mostly when not interacting at all. Some (like myself) noted that they seem to occur more often if you are doing a large amount of disk IO (partition move for example or copying 40+ GB of data disk to disk).

Hard locks happen in either Windows (7,8,10) or Linux.

With CStates on, the crashes are often and usually within 10 minutes of IDLE. With CStates disabled completely most find that the system remains "mainly" stable for days. Generally, no one has reported locks when gaming or under high CPU load (so I do not think this is the Prime95 example but then, since compilers / drivers use the latest instructions (according to CPUID) I cant be sure).

The symptoms happen on many different manufacturers motherboards, RAM types and setups. We have successfully removed or replaced all components and the only change that seems to make a difference is CPU for a few. CStates off in the BIOS is absolutely a guarenteed way to stabilize significantly but not fix.

I changed CPU (after trying everything else) and find it now mostly stable with CStates ON. I must be clear that my system is 100% identical, CPU # 1 crashes all the time and must have CStates off, CPU # 2 does not crash so often even with CStates ON. One may argue this is a problem with my PC, a fault, but with so many people having the exact same issues I cant believe it is a defect. And with Cstates/SpeedStep so specifically (a known new part of Skylakes architecture), I remain convinced that this is something real.

Others were not so lucky, a replacement CPU did not make the situation better. Batches from different factories as well etc etc

My new CPU has hard locked but not when IDLE like the last one (not yet anyway). Mine can lock when under full load, video transcoding, but not more than 1 in 10 sessions. Cstates ON seems equally stable on this CPU as OFF.

Thanks for any suggestions you can give.

[ssote]

Ronin wrote:

wingman99,

This is not the first mainboard that has these problems. On skylake VCCIO is important for C7 state of the CPU, and is the voltage for the memory controller and shared cache also.

This is another generation of CPU's, that document is from 2011. If you really want to read more, here you go : http://www.intel.com/content/www/us/en/processors/core/desktop-6th-gen-core-family-datasheet-vol-1.html 6th Generation Intel® Core™ Processor Family Datasheet, Vol. 1.

Thanks for the link I read through it and have found Vccio is for the memory controller and shared cache. What is your default XMP Vccio Vccsa XMP 3200MHz ram? I I'm having memory problems with my 3200MHz ram.

[AKota1]

Ronin schrieb:

thehackman ,

Set Vcore to Adaptive, set memory on XMP, VCCIO 1.15V .

For VCCSA you need to add step by step voltage till you find the stable point, because I/O controller is forced more when you use a high performance video card + igpu active + memory on non standard timings. You should check in HWiNFO for each step you make because it doesn't correspond with what you set in bios. VCCSA can go safe till 1.28V in bios, in your case, after that you should be carefull (usually you get red numbers when it becomes dangerous, but those red numbers are for voltages you see in real in HWiNFO. There is +0.025 Voltage difference for VCCSA from what you set in bios ...

So if you set in bios white 1.25V in real is 1.275V ... and 1.275 V can be red. I hope you got the point.

But it maybe possible that you obtain stability with VCCSA at 1.25v or even lower.

I know it will be time taking but ...

Thank You very much for your feedback and your suggestions! i am realy grateful for your help.

I changed my settings as You suggested:

- XMP on

- Adaptive Vcore

- VCCIO 1.15V

- VCCSA 1.22V

• i ran Prime95 v2.66 - No Error, about 60-65°C (5min ~)

• Then i tried for fun Prime95 v2.89 - and i get an immediate BSOD with CLOCK_WATCHDOG_TIMEOUT
• The Vcore Changes in Prime95 v2.89 to 1.259V in HWiNFO64

You notice the 1.25V Vcore in Adaptive mode with Turbo Voltage@Auto and Offset@Auto?

Is the CPU defective?

To Verify the Vcore in Auto Setting:

• Vcore@Auto in BIOS - Prime95 v2.66

• Vcore@Auto in BIOS - Prime95 v2.89

Then i Set Adaptive Mode again, but this time offset@Auto and Turbo Voltage to 1.33V

And now i am Stable with Prime95 v2.89 - but with the same Voltage Readings in HWiNFO = 1.259V ????? (i got very confused as i saw that voltage readings)

CPU-Z reads 1.296V Vcore in Adaptivemode with Turbo Voltage set to 1.33V

TL;DR - Vcore in Adaptive Mode too low, Auto Mode is stable with every Prime Version - is the CPU defective?!

[CCris2]

The voltage that you see there is VID per core. You don't have to mess with that voltage as you don't have access to it.

You look at the wrong voltage Look for Vcore.

Those voltages are correct by the way.

Thats why I told you to put voltage on adaptive in bios, because I don't understand what Asus did to voltage on auto ...

On Asus Adaptive voltage is like Auto on Asrock ...

On another mainboard when Vcore was set to auto it was getting fixed voltage and thats wrong.

When CPU is on 8x multi CPU Vcore should be lower and on 42x higher.

It's good that you made it stable

[CCris2]

What is the Vcore that last version of CPU-Z reads ?

[CCris2]

CPUs have 2 states: idle when they run at x8 multiplier and Vcore is about 0.7-0.8V and under load when they run at x42 multiplier, at higher voltage 1.2-1.3V (I don't know the exact standard value).

The CPU-Z version you have is an old one. On the new one you has Vcore not VID voltage ...

[AKota1]

i tested a little bit around:

- VCore @ Auto = Vcore 1.440V load in CPU-Z 1.76 (newest) (while running prime 2.66 or 2.89, no difference)

- VCore @ Adaptive with Turbo@Auto and Offset@Auto = Vcore 1.200V load in CPU-Z (while running prime 2.66; 2.89 = freeze)

- VCore @ Adaptive with Turbo@1.3V and Offset@Auto = Vcore 1.296V to 1.312 (switches around) load in CPU-Z (while running prime 2.66 or 2.89, no difference

- VCore @ Adaptive with Turbo@1.3V and Offset@Auto = Vcore 0.864V @ 800MHz idle in CPU-Z

i did not test the other idle Vcores in each setting.

why is there such a significant difference between the Adaptive Mode Auto setting and the "Full" Auto setting for the CPU Vcore?

[CCris2]

Well, you should ask Asus ... Mainboard sets those voltages ...

[CCris2]

But 1.440V is not good. Did you set manual voltage ? If not that mainboard has a problem ...

The Adaptive voltage seems ok.

[AKota1]

Ronin schrieb:

But 1.440V is not good. Did you set manual voltage ? If not that mainboard has a problem ...

The Adaptive voltage seems ok.

1.440V is on Auto, thats why i am so surprised...

i have a maximus viii ranger here to test with, then i can say for sure it's the mainboard.

do i have to pinpoint the right voltage for the adaptive mode added turbo voltage? Or is 1.3V a good value (will test it with Prime95 V2.89)

[CCris2]

SS with voltage on 800Mhz :

SS with voltage on 4200 Mhz :

[ssote]

Ronin wrote:

If you set memory to XMP it will set automatically the voltage for DRAM to what was programmed. You don't need to set voltage.

I have an Asrock mainboard Z170 extreme 7 +, DDR4 3200. When I activate XMP DDR voltage goes to 1.35, VCCIO goes to 1.2V (Which is to high, on red, and I put it lower manually, because it works the same), and VCCSA to 1.25V.

I have a question when you see what your Default XMP VCCIO is and VCCSA, can you see it in bios or just the utility, also what was your XMP Vccsa value before you changed it. I just put together a new skylake and I'm having memory trouble with XMP enabled on DDR4 3200MHz memory.

[ssote]

thehackman wrote:

Ronin schrieb:

But 1.440V is not good. Did you set manual voltage ? If not that mainboard has a problem ...

The Adaptive voltage seems ok.

1.440V is on Auto, thats why i am so surprised...

i have a maximus viii ranger here to test with, then i can say for sure it's the mainboard.

do i have to pinpoint the right voltage for the adaptive mode added turbo voltage? Or is 1.3V a good value (will test it with Prime95 V2.89)

Can you post a screen shot of CPU-Z with the prime95 running, your clock on 4 cores is only suppose to be 4.0GHz, turbo only goes +2 on one core.http://www.intel.com/content/www/us/en/support/processors/000005523.html Intel® Core™ i7 Desktop Processor Turbo Boost Technology Frequency...

You may have some performance mode enabled, my board has one.

[AKota1]

Ok, i have always 4.2 ghz with load on all cores - ss with Prime 95 v2.89 @8threads:

i dont know where i would have activiated this "performance mode" - maybe its the motherboards setting for the XMP Profile?!

because the XMP Profile activiation sets the CPU Core Ration to "SYNC ALL CORES" - maybe thats why...

ill reset the BIOS to default settings and try again.

[CCris2]

VCCIO and VCCSA you can see in windows in HWiNFO. XMP changes VCCIO and VCCSA only when you activate it in bios.

If you don't set XMP, VCCIO and VCCSA will be set on auto for standard memory on 2133, that is about 0.950V and 1.050V.

[CCris2]

XMP stands for Xtreme Memory Profile. When you activate it bios is optimized to work with that memory. Also the manufacturer need to optimize his bios for that memory in XMP otherwise will not work correctly and you will have to do manual adjustments ...

Yes XMP can modify CPU Core Ratio, because thats the way memory was suppose to work. It can modify other settings also. It all depends how the memory was tested to work.

[AKota1]

Ok thank you, i also read that 1.44V on Auto is normal for some i7 6700K CPU's (asus ROG forum).

right now iam still testing

XMP on

Multicore Enhancement off

core sync ratio "per core" (manually set by me, XMP sets it to "sync all cores" - but i think it does not make any difference, will test it though)

vcore adaptive with turbo voltage@1.3v

vccio 1.15V

vccsa 1.2V

i think these settings are sufficient, but only time will tell

thank you guys for your help!

conclusion: no part has a defect, only my ram is not 100% compatible with my mainboard and i have to adjust the vccio and vccsa voltages because of high performance card + igpu usage? is that plausible for your guys?

thats because the XMP Profile does not adjust the VCCIO and VCCSA voltages, they stay on auto when i enable it.

[CCris2]

Even if they are on auto they should be modified to the corresponding values, and you will not see. You should see this in HWiNFO. The problem is that, on Asus, even if memory is in mainboard memory list, in some cases, the bios doesn't auto adjust the voltages and other settings properly. That happened with sahafiec in another thread. He tested 3 types of memory on his ROG Asus mainboard, and 3xCPUs and in the end he modified voltages .

[AKota1]

ok now i understand, in BIOS the settings stay on auto, but the Motherboard should increase the values according to the XMP setting Profile - if that is not properly implemented, we have to adjust them manually.

thank you for answering all my questions!

cheers,

thehackman

[CCris2]

You got it right !

[LMonk]

I had hard lock with switched off freq scaling and c-states with solid voltage... I think it is time to give up, really give up. I dunno if i really need to try this settings for sa and other... I remember i tried this already.

So what i did, loaded default optimized settings, selected xpm profile, enable adaptive mode and live alone his auto settings, and set that params which you said to 1.15, one of them become red.

UPD. I tested auto adaptive mode and got hard lock, so now i set back there prooved stable settings.

[CCris2]

at 1.15V VCCSA and VCCIO there is no way they are red ...