Super low 8700k core voltage / temps question

[tommurto]

I just built a new system and have been doing some tweaking and testing to check if my airflow is in check, and if I may possibly have some overclocking headroom.

First thing I did was undervolt the CPU to see how that affects temperatures (I am working towards silence at the moment). To my surprise, using the offset in bios, my Core Voltage is fluctuating as low as 0,624 according to CPU-Z (usually under stress test 1,216, maximum of 1,312). That seems really really low, but running stress tests such as AIDA64 and Prime95 have been absolutely stable. But can such a low reading actually be correct? CPU-Z and Aida64 both show the same figures. I have no MCE enabled, only XMP.

AIDA64 FPU test gave me a max temp of 84, which I am relatively satisfied with considering my cooler isn't high-end. Will invest later when I overclock. I'm not comfortable installing a super-heavy heatsink, as I move my system to lans regularly.

System specs:
i7 8700k
Thermaltake True Spirit 140 Direct
Asrock Z370 Extreme4
16gb (2x8gb) G.Skill Ripjaws V 3200mhz CL16 DDR4
GTX 970 (2080 Ti is preordered)
NZXT H500 (2 intake fans + 2 exhaust fans)

 

[Darkbreeze]

Aida and Prime are not really acceptable as primary stability tests. Download and run Realbench when you make any configuration change to the CPU. If it runs 8 hours at that setting, using realistic AVX workloads, then it is stable. If it stops due to errors, then it is not stable. Period.

If you are making a number of changes and wish to only address basic stability between changes, you can run 1 hour tests of Realbench, but before considering any setting as stable, you need to perform an 8 hour run of Realbench. This has become pretty much the gold standard in stability testing. Workloads are realistic and and the methods are acceptable. Choose the amount of memory that reflects your installed capacity of memory and don't try using the system while it is running the test. It won't be usable.

Then, and only then, can you really come back and say the CPU is stable at a given setting. These other tests you've mentioned are very useful in their own right, but mostly as supplemental tests or for thermal testing rather than stability. Big difference.

Your CPU stability should be validated BEFORE you make any changes to the memory configuration from the default JEDEC SPD configuration of 2133mhz. Your XMP profile is an OC, and as such, requires testing of it's own but not until AFTER you are ENTIRELY finished with whatever changes you intend to make to the CPU configuration. It is a process and needs to be done in a specific order.

 

[tommurto]

Aida and Prime are not really acceptable as primary stability tests. Download and run Realbench when you make any configuration change to the CPU. If it runs 8 hours at that setting, using realistic AVX workloads, then it is stable. If it stops due to errors, then it is not stable. Period.

If you are making a number of changes and wish to only address basic stability between changes, you can run 1 hour tests of Realbench, but before considering any setting as stable, you need to perform an 8 hour run of Realbench. This has become pretty much the gold standard in stability testing. Workloads are realistic and and the methods are acceptable. Choose the amount of memory that reflects your installed capacity of memory and don't try using the system while it is running the test. It won't be usable.

Then, and only then, can you really come back and say the CPU is stable at a given setting. These other tests you've mentioned are very useful in their own right, but mostly as supplemental tests or for thermal testing rather than stability. Big difference.

Your CPU stability should be validated BEFORE you make any changes to the memory configuration from the default JEDEC SPD configuration of 2133mhz. Your XMP profile is an OC, and as such, requires testing of it's own but not until AFTER you are ENTIRELY finished with whatever changes you intend to make to the CPU configuration. It is a process and needs to be done in a specific order.

Alright, so in terms of thermals Realbench will be equivalent to Prime95 and Aida64, but stability-wise is a better test? In other words I should turn off XMP, do an 8 hour bench with the lower vcore, and if that is stable I will try with XMP and do it again? I presume the required vcore for XMP will be higher with XMP than without, since as you said, it is an overclock.

 

[Darkbreeze]

Actually, it will not be equivalent to Prime 95 version 26.6 (And ONLY version 26.6, which does NOT use AVX instructions) on Small FFT for thermal testing, and that's ok, because Realbench isn't intended for thermal testing. It's intended for stability testing and benchmarking only. Also, two separate things. The benchmarking portion is not for stability, and the stress test is not for benchmarking. We are only concerned with the stress testing portion of Realbench, which will not in fact apply a steady state thermal load like Prime95 version 26.6 Small FFT.

For thermal testing, that is what you want to run, for 15 minutes. Run this EVERY time you make a change to the CPU multiplier or voltage, to verify thermal compliance again.

Then, run Realbench every time to verify stability. Anything else is just throwing stuff at the wall and hoping something sticks.

XMP does not have a meaningful direct relationship with vcore. Obviously there is a relationship due to the memory controller component, but those are no longer IN the cpu, so it's less of a factor and no vcore adjustments are necessary once you have validated your CPU overclock OR any time you change memory settings.

What you DO need to be aware of and potentially make changes to are the DRAM voltage settings, the VCCIO voltage settings and the System agent voltage settings. VCCIO and SA both affect memory performance and stability, mainly stability. If you are overclocking and have a high RAM frequency, I recommend setting those two between 1.1 and 1.2v. There is a lot of information to found about this here:

https://www.tweaktown.com/guides/8481/coffee-lake-overclocking-guide/index.html

But no changes to vcore are necessary due to anything you might do to the memory, EXCEPT that a lot of times you cannot get memory to run at it's rated speed if that is a rather high frequency, say, 3000mhz or higher, without often having to apply at least a moderate overclock to the CPU first, which of course requires making sure the CPU overclock is fully stable before moving on the memory configuration so you will know that the problems you are having ARE due to memory settings and not due to problems with the CPU configuration.

 

[mdd1963]

I'd suspect the only time core voltage was .65 V was when it was idling for a short while at some ~800 MHz or so...

 

[Darkbreeze]

Exactly right. Any other time if the voltage was that low, you'd crash, have errors or some other form of instability such as the system resetting.

 

[tommurto]

CPU-Z is showing 1,232 V at load, and the lower number 0,624 are at idle, yet even then CPU-Z still shows 3.7 - 4.3 Ghz.

 

[Darkbreeze]

CPU-Z is maybe the worst utility you could use for monitoring, next to windows resource monitor.

Download HWinfo and run "sensors only", or CoreTemp. Both are miles above CPU-Z.

HWmonitor, Open hardware monitor, Realtemp, Speccy, Speedfan, Windows utilities, CPU-Z and most of the bundled motherboard utilities are often not the best choice as they are not always accurate. Some are actually grossly inaccurate, especially with certain chipsets or specific sensors that for whatever reason they tend to not like or work well with. I've found HWinfo or CoreTemp to be the MOST accurate with the broadest range of chipsets and sensors. They are also almost religiously kept up to date.

CoreTemp is great for just CPU thermals including core temps or distance to TJmax on AMD platforms.

HWinfo is great for pretty much EVERYTHING, including CPU thermals, core loads, core temps, package temps, GPU sensors, HDD and SSD sensors, motherboard chipset and VRM sensor, all of it. Always select the "Sensors only" option when running HWinfo.

In cases where it is relevant and you are seeking help, then in order to help you, it's often necessary to SEE what's going on, in the event one of us can pick something out that seems out of place, or other indicators that just can't be communicated via a text only post. In these cases, posting an image of the HWinfo sensors or something else can be extremely helpful. That may not be the case in YOUR thread, but if it is then the information at the following link will show you how to do that:

*How to post images in Tom's hardware forums



Run HWinfo and look at system voltages and other sensor readings.

Monitoring temperatures, core speeds, voltages, clock ratios and other reported sensor data can often help to pick out an issue right off the bat. HWinfo is a good way to get that data and in my experience tends to be more accurate than some of the other utilities available. CPU-Z, GPU-Z and Core Temp all have their uses but HWinfo tends to have it all laid out in a more convenient fashion so you can usually see what one sensor is reporting while looking at another instead of having to flip through various tabs that have specific groupings.

After installation, run the utility and when asked, choose "sensors only". The other window options have some use but in most cases everything you need will be located in the sensors window. If you're taking screenshots to post for troubleshooting, it will most likely require taking three screenshots and scrolling down the sensors window between screenshots in order to capture them all.

It is most helpful if you can take a series of HWinfo screenshots at idle, after a cold boot to the desktop. Open HWinfo and wait for all of the Windows startup processes to complete. Usually about four or five minutes should be plenty. Take screenshots of all the HWinfo sensors.

Next, run something demanding like Prime95 version 26.6 or Heaven benchmark. Take another set of screenshots while either of those is running so we can see what the hardware is doing while under a load.

*Download HWinfo


For temperature monitoring only, I feel Core Temp is the most accurate and also offers a quick visual reference for core speed, load and CPU voltage:

*Download Core Temp

 

[tommurto]

I got both HWiNFO64 and Core Temp, but they only seem to show me my VID (fluctuating between 1,3199 and 1,4019 but apparently that is only the stock core voltage and doesn't take into account the offset I've setup in bios? Those core voltages would be way out of my comfort zone, especially running stock.

I am surprised to CPU-Z being garbage, since it is mentioned in so many places.

 

[Darkbreeze]

There are many things that are common, that still suck. Old Corsair CX power supplies, common, but suck.

Raidmax cases, commmon because they are cheap, and suck.

Using weak or low quality power supplies because people don't think the PSU is important, when it is actually the MOST important component in every system ever assembled. Without good, clean, stable power, every other component is at the mercy of the PSU. So they choose other parts as expensive as they can while providing for whatever PSU can be purchased with the small amount of the budget leftover at the end, and it will suck.

LOL. See a trend here? Yes, CPU-Z is ok for some basic information, no doubt there. But the information you get there is limited and CPU-Z is not kept up to date as religiously as some other utilities. It also doesn't have the option to change settings for how often it updates it's information.

Anyhow, it's really not useful for complex monitoring.

CoreTemp is not intended for voltage monitoring. It's only useful for very accurate core temps.

HWinfo should have a value for every important value your motherboard has sensors for. You may have chosen the wrong options or not configured your sensor screens correctly. You have to scroll down, the page is extensive and generally requires three screenshots to obtain all of the sensor data.

I guarantee your voltage is listed there somewhere. Like this. Vcore will be shown further down the page if you scroll down on practically any configuration I've ever worked on:

It's possible you might have disabled the option to see those sensors if you were greeted by a message saying that there could be trouble with a particular sensor and disabled it. If so, go into the options at the bottom of the sensors only window and choose the option to restore the original order on the layout tab in the configure sensors options, accessible by clicking the gear at the bottom of the window.

Also, make sure on the main options page in there that the box next to voltage is checked. Further, when starting HWinfo be sure that ONLY the option for "sensors only" has a check mark next to it.

 

[tommurto]

Alright cheers, found it. Ran the Realbench stress test and everything was stable for a 1 hour run, will run it overnight to make sure though.

HWiNFO64 (sensors only) is showing Vcore minimum at 0.624 and a maximum of 1.280. This is surprising, as HWiNFO also shows my minimum clocks at 4297,9 Mhz and top clock on all cores at 4,701,1 Mhz (despite not even having MCE enabled). Shouldn't stock i7 8700k have ony core #0 at 4,7 turbo and the others at 4,3?

 

[Darkbreeze]

Not necessarily. That single core turbo is ONLY if it is a single threaded process. The max ALL CORE turbo is only 4.3Ghz.

https://www.gamersnexus.net/guides/3077-explaining-coffee-lake-turbo-8700k-8600k

Also, IMO, you should go into the Windows control panel power options and enable the Performance power profile. Then click on Change plan settings. Click on change advanced power settings. Double click to open the Processor power management to open it and set the max to 100% and the min to 10%.

Then go into the bios and verify that Intel speed step is enabled. You will have to experiment to see if Intel speed shift is working right or not. It does not work right on a great many configurations. I personally leave Intel speed shift disabled, or I disable it, and enable Intel speed step as I know that feature works correctly on everything I've seen so far. These two things, the power profile and Intel speed step will allow the cores to drop voltage and frequency to much lower levels when they are not under a load and do not need to be at full frequency or voltage.

This will not only use less power but it will create less heat in the core, which allows cores to have some relax and cool down time, and that helps to keep the overall package temperature lower when not under a load, so that when it is under brief loads it might not climb as high. It will also help to extend the longevity of your CPU. Imagine, driving your car at 3500RPM, all the time, no matter what the conditions are, is going to cause the engine to fail much sooner than if you are coasting or idling when those RPMs are not necessary.

 

[tommurto]

Alright I made the change to the power profile. I had both speed shift and speed step enabled, but I don't quite understand what speed shift does. If I was able to run Realbench at the current setup, it is working as intended?

 

[Darkbreeze]

If your core speeds are normally fluctuating between about 800mhz and 4.3Ghz/4.7Ghz (Depending on load) and locking in at 4.3Ghz under all core full load, then it's working normally. If speeds are not dropping below the standard default base clock of 3.7Ghz, then I would disable Speed shift and leave Speed step enabled, and then check it again using HWinfo.

 

[tommurto]

Oh, thanks for that. After disabling speed shift, I am getting a minimum core clock of 799,6 Mhz. Before when it was still enabled, the minimum was at 4,3 despite being idle. I'm not exactly running budget parts in my sytem (Asrock Z370 Extreme4 should be a solid mobo), but is there a downside to speedshift not working if speed step manages 800mhz - 4700mhz fluctuation?

 

[vMax]

I wanted to mention that your vcore at load is not that low and it can be lower under load Prime/AIDA etc. especially as you are not overclocking. '0,624 according to CPU-Z (usually under stress test 1,216, maximum of 1,312).'

Under stress testing at stock you should be able to keep it at 1.23 to 1.26v. As an example on a offset/adaptive vcore for my 8700K overclocked to 4.8GHz, I can get away with 1.278v with a 'High' LLC and at 4.9GHz, my max under load (Prime95/AIDA) is 1.308v with a 'Turbo' LLC setting..

So I think you can try for a lower number and then test...

 

[tommurto]

I wanted to mention that your vcore at load is not that low and it can be lower under load Prime/AIDA etc. especially as you are not overclocking. '0,624 according to CPU-Z (usually under stress test 1,216, maximum of 1,312).'

Under stress testing at stock you should be able to keep it at 1.23 to 1.26v. As an example on a offset/adaptive vcore for my 8700K overclocked to 4.8GHz, I can get away with 1.278v with a 'High' LLC and at 4.9GHz, my max under load (Prime95/AIDA) is 1.308v with a 'Turbo' LLC setting..

So I think you can try for a lower number and then test...

Actually that's a good point, the minimun vcore was what was surprising me with how low it is.

I am now at -0.30 (mV) offset in bios, but I will see how far I can go.

I don't really need to consider vdroop and LLC until I start overclocking and upping the vcore, do I?

 

[vMax]

I wanted to mention that your vcore at load is not that low and it can be lower under load Prime/AIDA etc. especially as you are not overclocking. '0,624 according to CPU-Z (usually under stress test 1,216, maximum of 1,312).'

Under stress testing at stock you should be able to keep it at 1.23 to 1.26v. As an example on a offset/adaptive vcore for my 8700K overclocked to 4.8GHz, I can get away with 1.278v with a 'High' LLC and at 4.9GHz, my max under load (Prime95/AIDA) is 1.308v with a 'Turbo' LLC setting..

So I think you can try for a lower number and then test...

Actually that's a good point, the minimun vcore was what was surprising me with how low it is.

I am now at -0.30 (mV) offset in bios, but I will see how far I can go.

I don't really need to consider vdroop and LLC until I start overclocking and upping the vcore, do I?

With regards to LLC, in general no but lets say you get to a very low vcore and hit stability problems, then trying LLC up one notch might help....To be honest, I have never tried going the undervolt route...but it must be as much fun as overclocking....hope it works..

 

[tommurto]

Yeah basically seeing how low I can go in temps and noise, I am working out the limits low-end. Might work on overclocking later on.

 

[Darkbreeze]

Oh, thanks for that. After disabling speed shift, I am getting a minimum core clock of 799,6 Mhz. Before when it was still enabled, the minimum was at 4,3 despite being idle. I'm not exactly running budget parts in my sytem (Asrock Z370 Extreme4 should be a solid mobo), but is there a downside to speedshift not working if speed step manages 800mhz - 4700mhz fluctuation?

Not really. There is a theoretical faster response to load demands and return to idle clocks with Speed shift, but so far for me at least I've not seen Speed shift work correctly on any system yet. Others say it works on their system, but there are threads upon threads of it not working right across a variety of forums. I wouldn't worry about it. I don't. Maybe one day they'll get it right.

 

[Darkbreeze]

I wouldn't undervolt, at all. The stock voltage is a result of thousands of hours of determining where the lowest stable voltage is for the stock frequency settings already. Anything under that is just as risky as overclocking without an increase in voltage.

Much as when you must test each incremental change when overclocking, if you are going to undervolt, and there really is no provable benefit of doing this, I would highly recommend running Realbench for 8 hours before calling it stable. Choose the stress test option, select the amount of installed RAM and click start. Run for 8 hours. Anything less than that and it's pretty uncertain.

Also, Prime95 and Aida are not good metrics for use as stability tests of the CPU configuration. Prime95 version 26.6 is good for testing thermal compliance using the Small FFT option. Aida, is not much use at all really as it does not test using steady state workloads like Small FFT and it's stress workloads are not nearly as reliable as Realbench. Prime95 Blend or Custom tests are good for testing memory stability however.

 

[tommurto]

I managed an 8 hour stress test with RealBench, Vcore 0,592 - 1,296 temps maxing out at 81 C.

Why undervolting? Because stock VID would pump 1,435 into the core at stock clocks. Less power = less heat = less noise is my aim at the moment, since until I get my 2080 Ti installed, my GTX 970 will be bottlenecking me anyway. If there will be tangible improvements to game performance by overclocking, at least I have some reference on where to start on the lower end and work my way up. Working 8h Realbench cycles and Prime95/Aida64 to test thermals.

True Spirit 140 seems to be doing a tolerable job, custom loop is too laborious for me, NH-D15 is too heavy for my comfort, and AIOs seem to be rather underwhelming in relation to noise (except the x62 seems to be pretty good for noise/thermals ratio).

 

[vMax]

I managed an 8 hour stress test with RealBench, Vcore 0,592 - 1,296 temps maxing out at 81 C.

Why undervolting? Because stock VID would pump 1,435 into the core at stock clocks. Less power = less heat = less noise is my aim at the moment, since until I get my 2080 Ti installed, my GTX 970 will be bottlenecking me anyway. If there will be tangible improvements to game performance by overclocking, at least I have some reference on where to start on the lower end and work my way up. Working 8h Realbench cycles and Prime95/Aida64 to test thermals.

True Spirit 140 seems to be doing a tolerable job, custom loop is too laborious for me, NH-D15 is too heavy for my comfort, and AIOs seem to be rather underwhelming in relation to noise (except the x62 seems to be pretty good for noise/thermals ratio).

I have to admit running the Corsair H150i Pro and boy is it quiet with the Mag Lev fans which the smaller H115i pro also has...very, very quiet and good cooling on top.

 

[Darkbreeze]

Wait, wait wait. No, no, no, no, noooooooooooo!

First of all, there is no possible way your CPU could require that high of voltage at the stock configuration, even under load. Stock voltage for the 8700k shouldn't exceed maybe 1.245v, MAX, RUNNING AVX versions of Prime95, with the default configuration. If your vcore is automatically being pumped up to 1.435v, something is seriously wrong.

ALSO, if you are getting 81°C temps, at the STOCK default configuration, with an UNDER volted vcore, and a True spirit 140 CPU cooler (Which is a pretty good air cooled option), then some is even further wrong.

Your max temps with that cooler and a stock configuration shouldn't exceed MAYBE 65-70°C, at full load on all cores. And that is probably pushing things a bit. I realize this is a higher core CPU, but the default boost profile and vcore settings, along with the fact that you're not running a stock cooler type configuration should mitigate most of that since it's not overclocked.