Really NEED Undervolt guide for 14700k+MSI z790 mobo

[AndyTheGreat]

Hi, community. Just bought 14700K for my MSI z790-P Wifi DDR5. It burns like hell. I get 99-101 in Cinebench 24 and 102 in AIDA, around 99 in Premiere Pro and After Effects. Did not even try other software since it's obvious my DeepCool LS720 (promised by Tom's Hardware to cope with 13900k's over 300 watts) cannot withstand the 14700k 278 without thermal throttling.
So I either have to get another AIO - and I have no idea which one because there are aio owners saying their particular aio beats the temps of 14900k getting not more than 87 in tetsts and there are ones saying even the best 420 mm aios cannot cope with their 13700k - or undervolt my CPU.
I followed the instructions from some post right here on the MSI forums (cannot find it) and lowered the long term and short term power to 253 watts and also set the Lite Load from the default Mode 9 to Mode 3 and the temps are now ten degrees less. Is it enough? Is there any other method or is this one correct even?
I'm completely dumb in undervolting - those are just numbers in BIOS for me, so I really need some sort of an MSI mobo guide to correct and safe undervolting - step by step with screensots or a video. Please, help.

 

[FlyingScot]

No idea, bro... I checked every single setting in the Bios. Except for the Intel-related settings (c-states and so on which before the update were "Dis"), cannot think of anything else. A confirmation has come that my MX-6 arrived, so the thermal paste issue will be checked today

CiTay, any thoughts on how we can expedite our way to finding the source of the extra heat?

 

[AndyTheGreat]

Perhaps better cooling solution?

Which one is better? I've studied hundreds of reviews and all the major aios perform mostly the same way, differing between 2-4 degrees, moreover in some tests you see, for instance, Lian Li Trinity Performance being the most effective, in other it lacks behind the Kraken 360, which does worse in a third test... My LS720 was reviewed as the best AIO in 2023 which could harness 320 watts on 13900k. Now somehow it cannot cope with even 250 without throttling. Liquid Freezer 3 you would say? I've looked through the customers' reviews in my city local retails - of those who bought it and all of them say it performs even worse than LT\LS720, giving poor results in cinebench and quite high temps... So what is a better AIO? Mind, that on every your suggestion I can find owners' reviews proving the chosen AIO doesn't perform as in bloggers' reviews. This may be the AIO stagnation problem...

 

[AndyTheGreat]

Only if you leave CEP enabled. IA CEP is the "Current Excursion Protection" mechanism for the IA cores (normal CPU cores), and it wants to prevent any undercurrent or overcurrent from a narrow window that is expected for a CPU. Once it sees a break from the norm, it will work against it by also lowering performance a lot. With an active IA CEP, when using a lower CPU Lite Load mode, the performance can massively decrease, similar to here. Of course the temperatures will be much lower as well, but without the performance, that is no good. So in that case CEP has to be disabled (if it's possible, depending on the board + CPU combination), then the full performance will be kept, at lower power draw.

I'll try testing with the new thermal paste and the CEPs disabled, let's check it up

 

[AndyTheGreat]

Didn‘t you have AVX offset at -3 before? You now have it on Auto. Who knows what auto is doing in relation to your scores and temps. But we do know that Cinebench R23 uses AVX and it generates lots of heat. In regards to your AIO pump, the most basic test for a blockage is to feel the two pipes when under load. One pipe should be warmer than the other. Also, if you have a blockage in the pump itself then you might temporarily dislodge it when repasting - shake the pump and turn it upside down. If temps improve for a little while and then get poor again, you need to order that new AIO you talked about.

Missed that. When I last set it to -3 you said that wasn't too good an idea because avx instructions are used in alotta tasks )) so should I leave it on auto or try with the -3? I think the -0.125 vcore offset got me a bsod with AVX -3, if I remember it correctly. But I can try with that too...
By the way, apart from Tesla energy generation, it's just a vcore offset by CPU negative, so -0.125 is correct, though it might be not as stable as -0,09

 

[Arctucas]

Which one is better? I've studied hundreds of reviews and all the major aios perform mostly the same way, differing between 2-4 degrees, moreover in some tests you see, for instance, Lian Li Trinity Performance being the most effective, in other it lacks behind the Kraken 360, which does worse in a third test... My LS720 was reviewed as the best AIO in 2023 which could harness 320 watts on 13900k. Now somehow it cannot cope with even 250 without throttling. Liquid Freezer 3 you would say? I've looked through the customers' reviews in my city local retails - of those who bought it and all of them say it performs even worse than LT\LS720, giving poor results in cinebench and quite high temps... So what is a better AIO? Mind, that on every your suggestion I can find owners' reviews proving the chosen AIO doesn't perform as in bloggers' reviews. This may be the AIO stagnation problem...

Cooling is the key. Better cooling = better performance.

You probably do not want to hear this, but; custom loop. This is not for everyone.

These 14th gen just run hot. You can settle for lower performance, in which case, you might as well have purchased a 14600 (non-K) or similar. Or, you can delid the processor, build a custom loop with direct die waterblock and liquid metal TIM, and enjoy the full potential of your processor.

 

[Arctucas]

Missed that. When I last set it to -3 you said that wasn't too good an idea because avx instructions are used in alotta tasks )) so should I leave it on auto or try with the -3? I think the -0.125 vcore offset got me a bsod with AVX -3, if I remember it correctly. But I can try with that too...
By the way, apart from Tesla energy generation, it's just a vcore offset by CPU negative, so -0.125 is correct, though it might be not as stable as -0,09

I believe what @FlyingScot was asking is what is your VCore set to in BIOS, before your applied the negative offset?

For example; I set VCore to 1.32 V, and then apply a -.120 V offset. This is Adaptive+Offset.

 

[citay]

CiTay, any thoughts on how we can expedite our way to finding the source of the extra heat?

Heat is the direct result of the power draw. If the cooling capabilities remained relatively constant, then the explanation must be that the power draw was lower before, and is now higher. It's not enough to know that the BIOS settings you changed are identical to before (because the newer BIOS could, unbeknownst to you, set other values differently by itself, including some that are not exposed to the end user). We have to look at the end product, i.e. the resulting power draw in the end. What's worse, some of the settings (like the DC loadline) can also affect how accurate the reported power draw is.

So if we really wanted to be accurate, we would have to compare the measured power draw at the wall, which is something i routinely do in my testing. Then you have some hard data that cannot lie, and the CPU Package Power, while still important, is more of a confirmation of what you can already see on the energy cost meter (hope that's the correct word). This kind of device:

That's from my old system (i5-11500), showing:
Idle: 23 W
Full singlethread load (Cinebench R23 Single): 50 W
Full multithread load (Cinebench R23 Multi): 118 W
Maximum load (Prime95 Small FFTs): 160 W

Sure, it includes the board and its VRM losses, the drives, the GPU and the PSU's internal losses, some of it complicates the idea of wanting to mainy look at the CPU. But still, this is a great way to compare the big picture across different things you try, and i use this little energy meter often, certainly when i mess with the voltage. Which is usually just after having built the system and doing some intitial testing, but i also hook it up sometimes after a BIOS update, to check that there hasn't been any regression for whatever reason.

 

[FlyingScot]

Heat is the direct result of the power draw. If the cooling capabilities remained relatively constant, then the explanation must be that the power draw was lower before, and is now higher.

Exactly my thought…so how about I take a stab at summarizing the issue:
Andy is seeing higher temps (and lower scores) in Cinebench R23.
The higher temps are either due to poorer cooling or higher power consumption.
To rule out poorer cooling, one should look at AIO performance and case fan speeds vs. the original setup. In other words, how confident are you that nothing has changed?
Now to the question of power consumption. Two things to look for:
[1] Has frequency increased, which increases current (amps)? You are no longer running an AVX offset, so the answer on the surface could be yes. However, you would expect higher scores, not lower - unless you are now thermally throttling via a lower TjMax value than before. So, something else has changed.
[2] Has Vcore under load increased? More volts = more power consumption = more heat. Instinctively, it looks like the answer is yes. But without knowing previous Vcore when running R23, it’s hard to know for sure. Lots of BIOS settings (and even microcode changes) can affect Vcore, but I’m dubious that it’s a hidden one. CEP is the most likely culprit if you have reactivated it. Also, relationship of AC LL values to CPU Lite Load could have changed in new BIOS. Do you know what you were using before?
Another thing would be LLC Auto behavior may have changed (I.e. Vdroop management under load), although I would be surprised if MSI did that. If you are manually overclocking then I think we can ignore TVB and eTVB, etc. Although, I don’t know that for sure.

Bottomline: To solve this riddle with any sense of certainty, you need (at the minimum) to be able to compare average/steady core frequency plus average/steady Vcore (e.g. 1.28v vs. 1.34v) when running R23 versus before. If you knew that Vcore had jumped or frequency was higher then we could go looking for which setting or settings was responsible. Do you have a HWinfo64 snapshot from before that you can use to compare to now???

 

[AndyTheGreat]

Ок, this is what's left from the thermal paste (Thermalright TFX romoured to be lake 14 W/mK). So partly the temps was its fault or maybe a coincidence/ And here are the aftewrwards testing in OCCT Power

 

[AndyTheGreat]

I believe what @FlyingScot was asking is what is your VCore set to in BIOS, before your applied the negative offset?

For example; I set VCore to 1.32 V, and then apply a -.120 V offset. This is Adaptive+Offset.

Actually I didn't change anything except for the offset... The rest was Auto. I'mm gonna take all the settings screenshots right now... Oh, yeah, about the custom loop. There are no companies that can give you the service of setting up a custom loop in your PC and later on give the support for it in my city, so it's a dead end. I don't have experience in this either

 

[AndyTheGreat]

Ok, The CEPs are DISed, AVX2 is set to -3.
While I'me testing, please, look through my BIOS settings I have now and tell me what to change (you'll notice no changes in CEPs and AVX there - I took the screenshots before changing these settings)
One thing I have to admit - till this year I had no idea about anything related to undervolting or overclocking, so please, be gracious

 

[FlyingScot]

So if we really wanted to be accurate, we would have to compare the measured power draw at the wall, which is something i routinely do in my testing. Then you have some hard data that cannot lie, and the CPU Package Power, while still important, is more of a confirmation of what you can already see on the energy cost meter (hope that's the correct word). This kind of device:

View attachment 189995

I have been deliberating whether to get one of these. They look very useful for times like this. I have a 1200W PSU in my latest PC and I wanted to get a good laugh when I see how much it pulls from the wall. A little excessive, I know. But it was the same price as the 1000W model, which was the next level up from my 850W.

 

[AndyTheGreat]

5 minutes OCCT Power test with the changed settings I mentioned above

 

[FlyingScot]

One thing I have to admit - till this year I had no idea about anything related to undervolting or overclocking, so please, be gracious

Hey, trying to help you diagnose your problems is forcing me to think about how to troubleshoot. We are learning together, mate!
Once you get the system settled down again, do yourself a favor and screenshot all BIOS screens and HWInfo screens (when running R23) so you have a point of reference. And don't forget to keep the Windows Power Plan the same, too.

 

[FlyingScot]

5 minutes OCCT Power test with the changed settings I mentioned above
View attachment 190008

230Watts at 86C. I don't think that's unreasonable. Maybe you could shave off a little more, but you have to be happy. What's your R23 score look like?

 

[AndyTheGreat]

230Watts at 86C. I don't think that's unreasonable. Maybe you could shave off a little more, but you have to be happy. What's your R23 score look like?

Not as high as it must be. TechPowerUp had 34507 at stock (253\253) with a Liquid Freezer 2 (420 mm). Yes it may be a better AIO, but I'm not testing at stock either, so with undervolting I should get 35k pts at least, but I get this:

 

[AndyTheGreat]

Btw, worth mentioning. You may not believe me, but such a "not many system resources demanding" thing as video editing makes my CPU sweat the same way as the Cinebench test - I got 91 degrees today when working on a video in Premiere Pro.

 

[FlyingScot]

One question I have: Why would 230W peak in OCCT Power = 86C and 230W peak in R23 = 96C?
I can only think of two things: R23 is somehow running AVX and OCCT is not...or the answer lies in looking at the average Wattage. Run the two again and make sure to reset the counters immediately after hitting the run buttons. Then quickly note the average at the end. Also, make sure HWInfo captures the average over the whole runtime, which I think is the default. Ignore my 5 samples below. It should be 0. BTW, a Vcore of 1.4 is going to get hot. That's going to have to be lowered if you want a cooler rig. You appear to be throttling for various reasons, thermals, power, amps, it looks like. All of this is due to the high VID/VCore. Whether your CPU can handle less VCore at that frequency is the question.

 

[AndyTheGreat]

One question I have: Why would 230W peak in OCCT Power = 86C and 230W peak in R23 = 96C?
I can only think of two things: R23 is somehow running AVX and OCCT is not...or the answer lies in looking at the average Wattage. Run the two again and make sure to reset the counters immediately after hitting the run buttons. Then quickly note the average at the end. Also, make sure HWInfo captures the average over the whole runtime, which I think is the default. Ignore my 5 samples below. It should be 0. BTW, a Vcore of 1.4 is going to get hot. That's going to have to be lowered if you want a cooler rig. You appear to be throttling for various reasons, thermals, power, amps, it looks like. All of this is due to the high VID/VCore. Whether your CPU can handle less VCore at that frequency is the question.

View attachment 190011

OCCT Power test was set to auto, so I set it to AVX2 manually. This is a 10 munites result:

 

[FlyingScot]

Andy, you’re going to need to expose the average column. That’s probably where the answer lies. The R23 workload must be more demanding. You’ll need to settle on one or the other app for consistency moving forward, especially if you are having thermal issues when editing videos.

CiTay has been saying for awhile now that these 14700Ks are more like an i9 when it comes to trying to get them under thermal control. I’m starting to get a sense of what he means. Damn… Intel just basically pre-overclocked these chips on the way out the door… If this trend continues, we‘re going to need to consider geothermal solutions!!!