Strange overclocking results

I managed to successfully stress-test my 8700K to 5.2 GHz for 15 minutes, at a maximum temperature of 81 C (briefly stayed at this temp, but most of the time at mid-70s), a core/cache voltage of 1.39V, a CVX core ratio of 1, and a minimum CPU cache ratio of 46. The thing is, CPU-Z said it was running at 5.2 GHz when the stress test was off, but 5.1 when running stress tests (using Real Bench's test and AIDA64's stress test). Is this unusual? Am I not running enough voltage? I do have CPU core/cache current limit at its maximum setting, and I set all RAM settings in the BIOS according to Corsair's specifications for the RAM.

I have those utilities, although I'm partial to Aida. I remember seeing some variance between what they read and what was set in the bios. I usually just go by whatever 3dmark says after the benchmark. Seems to be accurate for me and that's what it would be judged on as a baseline compared to other users.

As far as the specifics of how each measures and collects the data, and the discrepancy, I can't say.

edit: I guess I should probably mention, I turn off my OC when I play DCS. Generally speaking 7700k and 8700k are potatoes. They run hot. Not much you can do about it and delidding and upgraded TIM only mitigate the issue so much. I notice no performance drop or frames lost. GPU is definitely the bottleneck for DCS, at least in 4K.

I do notice fps drops if my XMP profile for my memory is off. That makes a difference in FPS and stutters. Have fun with the OC. Some people get more into racing their computers than playing games.

It's maybe because you have an AVX offset in your bios. Realbench runs AVX instructions so your frequency gets an offset during stress test. Just set it to zero (actually it should be - 100 Mhz) to keep frequency stable (but it will produce more heat).

When you run DCS you should also see your frequency always bouncing from 5.2 to 5.1 because DCS also runs AVX instructions triggering the offset each time.

But maybe it's other thing.

It says that my multiplier is at 51 when running the test, but 52 when not running, even though every pertinent setting is at manual in the BIOS.

So, do you like 3DMark better than the others, like Cinebench? What about VRMark and the other one that they make? They're $29 each and Cinebench was free. Is it worth spending the money? What about getting the $39 version of AIDA64, versus the free trial version? Am I missing some vital functionality?

It's maybe because you have an AVX offset in your bios. Realbench runs AVX instructions so your frequency gets an offset during stress test. Just set it to zero (actually it should be - 100 Mhz) to keep frequency stable (but it will produce more heat).

 

When you run DCS you should also see your frequency always bouncing from 5.2 to 5.1 because DCS also runs AVX instructions triggering the offset each time.

 

But maybe it's other thing.

So set AVX to zero and keep it there? I'm thinking that heat may not end up being a problem (hopefully). I've got 14 fans running, including 8 top-of-the-line Noctuas running push/pull around the radiators. The cpu was delidded, a copper heat spreader installed and lapped, and liquid metal used underneath too.

Results

Well, getting rid of the offset did it! Thanks!

Nice...

I paid for Aida, it's nice to have and the license is for multiple computers but you definitely don't need it.

I have the 3d mark steam version. I mainly benchmark with Time Spy Extreme and FireStrike Ultra. Those are pretty much the standards right now. I don't own VR yet, and haven't used that bench.

Cinebench is a good CPU test but applies to video encoding and rendering. Not really relevant for gaming. The FireStrike and TimeSpy are good synthetic benchmarks that replicate modern games and CPU physics. TimeSpy is Dx12, Firestrike is Dx11. Or vice versa, I forget.

I haven't been able to break the top 10 in scores, but top 20 in the world is pretty good. Especially for my cute little build. Most of the top scores are from the same people tuning their systems.

The Uniengine benchmarks are a good alternative (heaven, mountain, superposition) they have pay and free versions. To get on the leaderboards I think you have to buy a license.

The leaderboards are a great way to compare prospective components. ASUS and EVGA are usually top scorers.

Have fun, but it's easy to fall down a rabbit hole chasing scores and frames LOL, don't forget to play DCS too.

edit: I use Thermal Grizzly paste and FujiPoly Pads for TIM interface with the cooling loop aluminum. CPU is from Silicon Lottery, so whatever they use, I think they've upgraded their TIM procedure since I bought mine. I think they use liquid metal now.

I have the superposition pay version. Should I also get VR Mark, in addition to 3D with TimeSpy, etc?

Playing DCS in VR now, as a matter of fact. I've gotten over 110 games since getting the Oculus last month (!).

I have the superposition pay version. Should I also get VR Mark, in addition to 3D with TimeSpy, etc?

Playing DCS in VR now, as a matter of fact. I've gotten over 110 games since getting the Oculus last month (!).

Can't say for sure, I think there are several competing VR benchmarks free/pay etc. Still waiting to jump into that.

If you are new or recently back into gaming, make sure you take advantage of Steam sales. They have big ones in the summer and Christmas time. (DCS does too). Build up a wishlist and they'll send you a notification. Games from around 2010 and more recent scale well in the graphic settings. I've accumulated more games then I will ever have time for pennies on the dollar.

Portal and Fallout would be fun in VR. I've yet to try it to be honest, usually when I play DCS I play for hours at a time, not sure I want to wear a visor for that long. I would definitely buy the visor and suit from Ready Player One.

Note all your settings, i.e. make screenshots from Bios with mobile phone etc.. take a pen and a paper, but note each value you changed and set.

Then, as 1.39v is not the lowest, try with VSID to OFF and Voltage to 1.35 manual, LLC to 6 ( Asus as I see, so 6 is your best value for 5+G ).

I am sure, with manual Volts and LLC6 you can lower your applied volts and thus wattage.

Run prime95 and report back what temps you get, SMall FFT's, that's the real deal. Watch your temps !!!

You can try with 1.39v manual ( your current value ) but check in HWinfo what Volts are REALLY applied and working WHILE the machine is FULLY stressed. LLC6 will take care that your volts do NOT drop too low when you suck watts, when you lower the LLC you need more initial Volts to not fall below the minimum required volts/watts. Lower the Volts by x.x1 step by step while you torture it....when it clips away...1 back up.

I started at 1.425v for 5.2G ( LLC2 only but all Phases and CPU Power up to 140% ) and thought this was great....I now run 1.35 in idle and 1.32ish under full load, with much better temps and far more stable.

This applies to a "DELIDDED" CPU, if yours is still original, you should delid it and save some watts. If it does 5.2 without delid it's a good die, hands down.

I have those utilities, although I'm partial to Aida. I remember seeing some variance between what they read and what was set in the bios. I usually just go by whatever 3dmark says after the benchmark. Seems to be accurate for me and that's what it would be judged on as a baseline compared to other users.

 

As far as the specifics of how each measures and collects the data, and the discrepancy, I can't say.

 

edit: I guess I should probably mention, I turn off my OC when I play DCS. Generally speaking 7700k and 8700k are potatoes. They run hot. Not much you can do about it and delidding and upgraded TIM only mitigate the issue so much. I notice no performance drop or frames lost. GPU is definitely the bottleneck for DCS, at least in 4K.

 

I do notice fps drops if my XMP profile for my memory is off. That makes a difference in FPS and stutters. Have fun with the OC. Some people get more into racing their computers than playing games.

Delid'ing/TIM replacement with no other changes dropped my CPU from 58C to 45C. So it has a pretty huge impact.

I forgot to mention; I'm getting a fine jittering in VR, like a small vibration that makes things blurry when you look at them. This began after the overclock.

As a matter of fact, I did have LL set to 6, and pretty much everything else as you stated. Here's a series of pictures during a Prime95 version 26.6 small fft test, about 20 minutes into it. I had to scroll up and snap a new picture to get everything off of HW monitor. Temps never exceeded 80 C, and usually stayed in the low to mid-70s.

Very good results. Will probably run pretty lower when gaming DCS. If you play DCS for 1 hour, your system should stable for 1 hour. You know what I mean.

I use Windows Snipping Tool to capture screenshots. https://support.microsoft.com/en-ca/help/13776/windows-use-snipping-tool-to-capture-screenshots

It's your first overclock with that platform, take your time.

hwinfo is a different program than yours, ...and it is FREE, no " [Trial] " in that one :)

Try this: https://www.hwinfo.com/download.php.

You can add pages so it fills all your screen with 3 or 4 pages, then snip it with snipping tool from 10.

Most important is your VSID. Is that set to OFF in Bios ? That and LLC6 together make sense.

Do you mean SVID? I've got that set to "typical scenario".

I'll look for VSID, though. I'll also post some pictures of my BIOS settings.

Looking at the picture above, that ends in ...27.jpg,, which was taken 20 minutes into a small fft test, what would you set CPU core/cache voltage to? If I'm understanding you correctly, I should set to 1.41, because that's what it's saying the max voltage is?

Okay, I ran small ffts again for a half-hour, and used the software you recommended. The first 5 screenshots are from during the test, the next 5 are immediately after the test, and the last 12 are pictures of the areas of the BIOS where I made changes before. I honestly couldn't find anything called VSID. Could it be SVID you're thinking of? As for that, there's no off or disable setting, just what type of scenarios, and auto for choices.

Oh heck yes, SVID, not VSID.

There are two lines that relate to SVID, the one you mentioned is near the top of that page in Bios and defines the PRESET, which will give more juice the lesser your pick is, aka Worst case Scenario will pump a lot more volts than best case Scenario will. In any case, if SVID is on, you dont control the volts, the CPU and board do instead and you may join the party as cook #3.

Look at the same page in Bios, further down, where it also asks for performance presets...

lemme boot and I edit it.... LoL

CPU SVID SUPPORT is the one you need to disable, pic #3

ok...Asus allows screenshots in Bios with F12 key to a USB drive, very handy..

here are my pics...

Okay, I'll use that, and disable SVID. Based on the pictures of the hardware monitor I posted, what voltage do I need to set? Do my other settings look okaY/

Actually, looking at the pictures I posted, I actually do have CPU SVID support disabled. That being the case, what should I do with my voltage and other settings, based on the results I posted in those 10 pictures of the HW monitor? Also, while I got you here, what should I do about that fine jittering/vibration, or whatever it should be called, that makes things hard to read or see in VR? This is only present in DCS, by the way.

I would try to ower volts by doing this:

Leave the settings were they are and run a prime95 26.6 without any AVX to keep things a bot cooler for testing this. Once you find the lowest volt without AVX, you can use a recent p95 version with AVX and try that, that will for sure need a bit of Volts added again. But we keep it simple and do this:

Open AiSuite->TPU and while p95 runs SmallFFT's on version 26.6, lower the volts by 0.01 every 5-10min until it crashes or drops a core. If it crashes your OS or BSOD it is likely far too low, if it only drops 1 or 2-3 cores it can be fixed by upping the volt again to the previous value.

In HWinfo, you can read the vcore applied, that value will be the highest in idle and the ölowest under full pressure. LLC is what defines how low it will drop but also how far it may overshoot when the load goes back to zero again. Keep that in mind, LLC is similar to preloading, just dont preload too much. This is also very different on MSI Asrock etc.. they all have their own cook book.

SInce your temps are still very nice, I think you have a good die and can lower the volts some 0.05 or even more for 5G. Once you got the lowest stable, add 0.01 again to be safe...and then run p95 with AVX, likely you need 0.03-0.05v more for that. Watch the temps !

Once you know your non-AVX and with-AVX Volts, you can decide what you dial in for 24/7 Desktop use. I can lower my Volts to 1.27 for non-AVX but need 1.30v for bullet proof AVX. For 5.2G I need 1.30v without and 1.35 with AVX on my die, delidded.

YOu might want to set the Performance Mode in Bios to HIGH PERFPORMANCE, it changes a few values for the CPU power handling, duration of burst etc... I recommend you do that too.

For VR stuttering..I have no idea. Make sure all your USB is properly working, that LOD and PD is not too high.

Maybe try MSI IRQ mode on all devices.

Dont know, those are hard to pin down what causes it.

try lowering vccio and vcsaa as well.......

and for ffs....use raja overlock guide for the kabylake and cofeelake....

https://rog.asus.com/articles/guides/the-kaby-lake-overclocking-guide/

Do I need to turn avx on in BIOS when I run the newer Prime95 bwith AVX? How do i go into MSI IRQ mode? What is LOD (PD is a setting in DCS, right?) ?

By how much should I lower VCCIO and VCSAA, or how do I go about finding a more appropriate voltage? I'm drawing a blank on ffs for some resason, too. As far as voltages: am I using the applied voltage in HWInfo to set voltages manually, or am I using AISuite to test out voltages, and then going into BIOS after and setting core voltage to whatever stable voltage I find in AI Suite (just guessing that setting voltage in BIOS is more of a "hard" set, or more durable for lack of a better term.

That Kaby Lake guide you posted; is that applicable to the Coffee lake too?

One note of caution. When you test these things, you have a decent chance of crashing the system. Which *can* lead to corrupted Windows. So make sure you have a good backup (backuped up with verification on).

All it takes is tuning one thing, causing it to bluescreen/die, and not being able to fix it even with win10 boot/recovery disk.