New PC with i7-13700K and Z790 Gaming Plus WiFi - undervolt experience

V

Vassil_V

Member
Joined
Jul 14, 2024
Messages
89
Two days ago I finished a build with an i7-13700K (with AK620 cooler) and a MSI Z790 Gaming Plus motherboard. During my research while deciding what components to go with for the PC, I naturally came upon the dreaded issues with instability and degradation of 13th and 14th gen i7 and i9 CPUs, so I had already done some reading and preparation before building the PC. Before installing Windows I went ahead and changed some settings in the BIOS to the Intel recommended ones, as follows:
PL1 = PL2 = 253W
ICCMax set to 307A
IA CEP = enabled
eTVB, TVB, TVB voltage optimization = enabled
C-States = enabled (I've left C1E disabled)
I also set the TJMax to 95C instead of 100C, for peace of mind.

Note that the latest stable BIOS for my MB doesn't actually have the Intel Default Settings preset, but PL1/PL2, ICCMax are set automatically as above by selecting "Boxed Cooler" mode. There are two beta BIOS versions with Intel Default Settings, but I'd rather avoid betas. If somebody has experience with those, I'd appreciate the feedback.

I don't really understand the AC and DC loadline setting - if I understand correctly, MSI have many (23) different presets for that, called Lite Load Control. Intel recommend that AC=DC. Here, I've selected "Intel Default" mode, which shows up as AC=DC=110 in "advanced mode", which I believe means 1.1 oHms. According to official Intel 13th and 14th gen datasheet, this is the maximum value but I'm not sure what to think of that. Nevertheless, I left it at "Intel Default".

After installing Windows, drivers and all updates, I installed HWInfo and Cinebench 23. CB score was around 29000 and thermal throttle was instantaneous, so I started looking at voltages. What really surprised me was that the maximum VID (a bit later I learned to look at the VCore instead) was hitting 1.5V under light load. During all-core CB load it was about 1.35-1.37
After some experimenting with Lite Load modes ranging from Mode 13 (default for "Normal" profile) to Mode 9, I saw the voltage decreasing and temps improving, but the CB score gets significantly decreased - I saw as low as 20K points with Lite Load = 9. I believe this is because of IA CEP. As a result, I went back to the "Intel Default" preset where AC=DC=110 (as I also would prefer to stick with Intel's recommended settings), and decided to try just setting a negative CPU voltage offset.

Note - Loadline Calibration Control is set to Auto, haven't tested any of the 7 different modes.

I switched the CPU voltage to "Adaptive + Offset", and after some testing, it seems that -0.125V is stable - no crashes in any benchmarks or tests (including Prime95), nor in any of the games I've tested so far. -0.130V was also stable for some tests, but I decided to dial it back to -0.125V to be on the safer side.

With -0.125V offset, the CB 23 score for a single run is around 30700, maitaining 5.3 for P cores and 4.1/2 for E cores, at 225W. The voltage during this load is around 1.25-1.26V which seems great to me. The 10-minute run sees a bit of thermal throttling and the score drops a little, with the clocks decreasing by 100-200Mhz for both the P and E cores.
Idle CPU package temp (currently while writing this) is 32C, ambient temp = 24C.
During gaming, the average ranges between 50C and 70C and overall I'm very impressed by the temps during gaming.

I am very curious about voltage under light load and idle - when the CPU is downclocked (P cores at 3.3 and E cores at 2.6), the voltage can drop to as low as 0.7V, the absolute lowest I've recorded is 0.66V. On the otherside, when the clocks get boosted (but under light load) the Vcore can spike up to around 1.35-1.37V.
Do those voltages sound okay and what are you seeing on your 13th/14th gen i7/i9s?

My goal is stability and longevity (as much as I can influence this), I'm not after overclocking. I've also turned off enchanced turbo boost and turbo boost 3.0, and manually set the ratio limits to my two "star" cores to 53, so there is no boosting to 5.4GHz.

Overall, I am quite happy with my results but I'm under the impression that most people approach undervolting by adjusting Lite Load and Load Line Calibration, so I feel like I'm missing something.

Thanks to everyone who has read everything, and I'll welcome any feedback, suggestions, comments and personal experiences!
 
Last edited:
Sort by date Sort by votes
I've been running with AC/DC=1, LLC=7, CEP off until now, but now I'm trying different settings so that I can have CEP=on. If I turned CEP on with my AC/DC=1 I'd get heavy clock stretching
 
Upvote 0 Downvote
yes, I'm using the advanced offset config for individual V/f points. With AC/DC=1 LLC=7 I used these offsets: -0.02v @ x56, -0.02v @ x55, +0.03v @ x53, +0.11v @ x51, +0.15v @ x43, +0.12v @ x24, +0.11v @ x14, +0.08 @ x8
Now I'm trying with AC/DC=18/33, LLC=4, CEP on so far the offsets are: -0.14 @ x56, -0.14 @ x55, -0.09 @ x53, +0.01 @ x51, +0.08 @ x43, +0.06 @ x34, x24, x14, and x8
I don't like the "auto" settings because I don't know what they end up being in reality and they might even change with other settings or with bios updates
 
Upvote 0 Downvote
hungwai.lim said:
Found this discussion which set AC/DC to 0.01, which results same as me, lower temp. and better performance.
https://community.intel.com/t5/Proc...ty-Issues-Lower-Temps-amp-Better/td-p/1608867

I will test speedshift and EIST off when I am free, I will also test even lower CPU Lite Load to Mode 1 (AD_LL/DC_LL = 0.010/0.010 mOhm) to see if there are any different.
Click to expand...
Interesting. Thanks for sharing that link. I‘m rather fond of my old setup with LL=1,LLC=4. I have been wondering for awhile whether it would work on Raptor Lake. If you can get LLC close enough to LL then you can finish off the rest with manual offsets. And I don’t see why this wouldn’t allow you to make it work with CEP on or off.
 
Upvote 0 Downvote
Vassil_V said:
Actually, not only there are regular and extreme versions, many people on overclock.net claim that the regular version 15.0.37 is the one to beat, even over the Extreme mod, so that's what I've been using. But I just tried 6-7 runs of the Extreme one and it's also stable.

In any case, in just 4-5 consecutive runs, 15.0.37 found instability in a setup that had been 100% stable for more than three weeks, and that had successfully passed many runs of R23, 2024, OCCT, y-cruncher, some Prime95, UE5 shader compilation tests, actual gaming, and just general usage.
Click to expand...

I just tried something, got the regular CB R15 15.0.37 from here, then overwrote the plugins folder with that from the Extreme Edition (which normally uses 15.0.38).
Now there's 15.0.37 with four times the render resolution.

Before:
R15_1.jpg


After:
R15_2.jpg


According to some reports, this is more difficult again than the unmodded 15.0.37. Maybe you want to try this too.
One is supposed to run it a handful of times in quick succession.
 
Last edited:
Upvote 0 Downvote
is there an easy way of increasing the resolution even more, ideally by just editing some text file?
Of course if would be even more ideal if we could make it run in loops like R23
 
Upvote 0 Downvote
charonme said:
is there an easy way of increasing the resolution even more, ideally by just editing some text file?
Click to expand...

It's probably not too difficult to find out if you compare the files in the plugins folder. But i'm currently working on some other testing.

charonme said:
Of course if would be even more ideal if we could make it run in loops like R23
Click to expand...

Not sure how ideal that actually is, because buildzoid had previously said something to the extent of, especially when a run is starting and stopping in R15, those two moments can provoke transients. So it might be beneficial to have a slight pause with no load in between.
 
Upvote 0 Downvote
charonme said:
there is a pause with very little load in between the loops in R23 that triggers those transients, that's where I usually measure the highest voltage spikes on a scope (sometimes even 120mV higher than max vcore reading in hwinfo) you can also see the dips in power draw between the loops on this graph
Click to expand...
That's why it is important to use maximum LLC to avoid under and overshoots. Buildzoid is experimenting with CEP enabled but isn't not CEP disabling that causes the spikes it is the lower LLC. Enabling the CEP doesn't cure the problem, it fights the symptoms.
 
Upvote 0 Downvote
persiz154202d8 said:
That's why it is important to use maximum LLC to avoid under and overshoots. Buildzoid is experimenting with CEP enabled but isn't not CEP disabling that causes the spikes it is the lower LLC. Enabling the CEP doesn't cure the problem, it fights the symptoms.
Click to expand...
Not if you want to keep your average voltages lower, and the long spikes that are even worse than transient spikes. The transients being an Intel thing for generations. With RPL, I believe it’s better to drive down that AC_LL as far as it will go. Unfortunately, you can’t do that at LLC=Auto(8). It’s best not to take BuildZoid out of context. He says and tries a lot of things and his theories develop over time, and over many hours of videos.
 
Last edited:
Upvote 0 Downvote
It is not important to keep the average voltages low. It is important to prevent high voltage spikes while there is high current. I appreciate Buildzоid's work, but I disagree with some of his conclusions. At the end of the day, everyone is free to do their own tests and decide what is best for them.
 
Upvote 0 Downvote
citay said:
I just tried something, got the regular CB R15 15.0.37 from here, then overwrote the plugins folder with that from the Extreme Edition (which normally uses 15.0.38).
Now there's 15.0.37 with four times the render resolution.

Before:
View attachment 192130

After:
View attachment 192131

According to some reports, this is more difficult again than the unmodded 15.0.37. Maybe you want to try this too.
One is supposed to run it a handful of times in quick succession.
Click to expand...
Thanks for that, I've added it to my suite of testing tools! It is stable for me with configs B and C from my other post, and also with I'm runnning at the moment - AC=DC=50 / LLC=5 / -100mV offset / PL2=190W / P-cores at 5.5GHz all-core (not reached under full load).
 
Upvote 0 Downvote
I once had to run CB R15 approx 20 times in a row to discover an instability that wasn't revealed with occt, p95, y-cruncher nor linpack. But ffmpeg video reencode did crash even though it didn't even fully load the cpu
 
Upvote 0 Downvote
charonme said:
I once had to run CB R15 approx 20 times in a row to discover an instability that wasn't revealed with occt, p95, y-cruncher nor linpack. But ffmpeg video reencode did crash even though it didn't even fully load the cpu
Click to expand...
Sometimes something just crashes. I used my pc for 6 months without issues and discovered then that r15 gave hw errors. Didnt cause any instability in anything ive done before that.

Now, everything is stable, no r15 errors after 10 runs, prime95, occt, r23, 2 hours handbrake, all fine but cinebench 24 crashes sometimes. In the end, its a whole boatload of hardware with an even bigger boatload of software running all at the same time. It will never be 100% stable
 
Upvote 0 Downvote
Yeah, that's why when I test for instability I stick to 5-10 runs of R15, R23 and TimeSpy, one 10-minute run of CB 2024, and 30 minutes of OCCT. If this goes well, I just start using my computer normally and if there's something that needs to be addressed, it will present itself at some point. I firmly believe that if one is stubborn enough, you can always stumble upon a crash or two under some specific circumstances, and if you attribute it to hardware every time, you'll never be satisfied with your setup. My iPhone also crashes apps from time to time, even though I trust its loadlines and voltages are okay.
 
Upvote 0 Downvote
Back
Top