Raptor Lake Settings Survey and Spreadsheet "Database"

[FlyingScot]

Welcome to the Raptor Lake Settings Survey and Spreadsheet "Database"

While no two systems (or CPUs) are the same, it is my hope that you will find this centralized information useful when comparing your own specific setup to those who have the same CPU - and maybe similar hardware. In addition to this objective, I have also provided an opportunity for Raptor Lake owners to periodically submit additional entries to the survey database when anything changes. If degradation becomes more of a widespread issue than originally anticipated then these periodic "field reports" should allow us to see these trends as they unfold.

Instructions for Survey Submissions
Once you're satisfied with your system's stability and performance, please follow the instructions below to prepare for your first submission to the survey database.

Please download the attached Survey Guide.pdf. This guide has been designed to give each survey respondent a chance to gather all the relevant information ahead of time. This is an important step. The last two pages of the guide describes how you can make immediate use of the survey results, either before or after you complete your online entry.

In addition to the many questions in the survey, you will find a section (Section 2) that asks for performance related data. I have selected Cinebench R23 for this purpose. However, Cinebench R23 is only one small snapshot of performance and I realize will not necessarily be indicative of gaming performance, etc. In fact, as our recent testing has demonstrated, some "undervolting approaches" are much better suited to all-core CPU workloads (like Cinebench R23) than those favoring lighter, non all-core workloads, such as gaming.

[1] Use this link to access the Online Survey
NOTE: Survey responses cannot be edited once they are submitted. However, I can still assist you in making changes (see NOTEs below). Survey responses can also be deleted (by me) if you wish.
Please don’t forget to enter your Forum name (if you’re a member) to help others link you to your settings. But for privacy reasons, please don’t use your real name or email address, etc.

[2] Use this link to access the Online Spreadsheet "database".
NOTE: Pink cells indicate data that I have corrected on behalf of the respondents.
NOTE: To read extra long “User Comments”, double-click on them to pop up a scrollable box.

If you are very new to the topic of Raptor Lake tuning then my guide might help you get started.
Guide: "Almost" everything you need to know about Raptor Lake Voltage/Power/Temperature Tuning

If you’re already quite comfortable with undervolting via the Intel Loadline settings (LLC, AC_LL, DC_LL) then you should find Vassil_V’s case study quite interesting.
Different undervolting methods with IA CEP enabled, and how they compare to Lite Load
You will also find his latest “favorite" settings in the Survey database.

 

[FlyingScot]

I haven’t looked closely at your data yet, but what I would try first (if I were in your shoes) would be to try increasing the LLC from 3 to 4, and then back off my negative offset by 25mV. This is what I would try if I thought the problem was transient loads and the rapidly changing core frequencies that appeared to be causing the instability. Then I would keep playing with both of those values until I found stabiltiy.

I’m actually working on my own instability (WHEA errors) that showed up yesterday and today. If it’s what I think it is, it’s a perfect example of how crazy things can be when you have transient loads.

 

[jamiesaun85]

I haven’t looked closely at your data yet, but what I would try first (if I were in your shoes) would be to try increasing the LLC from 3 to 4, and then back off my negative offset by 25mV. This is what I would try if I thought the problem was transient loads and the rapidly changing core frequencies that appeared to be causing the instability. Then I would keep playing with both of those values until I found stabiltiy.

I’m actually working on my own instability (WHEA errors) that showed up yesterday and today. If it’s what I think it is, it’s a perfect example of how crazy things can be when you have transient loads.

Thanks for the tips. I'm actually not having any....with the exception of I got a page file full error, which i have never seen before and have no idea what that was all about. How could the page file possibly come into play? I have 64 gigs. But I happened to open task manager while running my usual apps and noticed that, which led to my theory, so I shared my thoughts on it and why i believe its a better test than synthetic stress testing.

And I did get one out of video memory error that's been a tell tale sign of degradation. I would be shocked if it actually was degradation though, I was doing so many task simultaneously i was just begging for an issue. That's kinda the point I guess, so I'm not worried about that one at all. Perhaps I should be, but it still hits all speeds with a max of what like 1.26 or so? something like that. Scores are very consistent along with voltages, current and temp. The only thing that has me a little worried is the amperage is about equal to the power? that's just crazy, never seen that before. You'll never run into that outside of CPU circuity, that's for sure.

Now, what I have expedienced is odd behavior inside of the apps themselves. Things take a little longer to launch, or menus to open. Its rare, and minor, but its there. Maybe a handful of times at most. But enough to warrant further testing. Because I dig testing, and I guess im not fat enough yet, though im getting there.

So you're saying to drop llc, and raise the voltage, correct? I was actually about to raise the voltage a bit anyway. I could also raise the acll a bit too. But you think a higher baseline would help with transients. Do i have that right?

 

[FlyingScot]

So you're saying to drop llc, and raise the voltage, correct? I was actually about to raise the voltage a bit anyway. I could also raise the acll a bit too. But you think a higher baseline would help with transients. Do i have that right?

Well, it does sound like you could be on the ragged edge of stability. Things getting sluggish can be a good indicator of voltage stability issues.

What I was suggesting is something I would try in an effort to potentially save time in my investigation. I’d simply hope it would change behavior in a way that might tell me what’s going on inside the CPU.

I personally think there are five types of instability:
[1] Voltages drop too low at the bottom of a “downswing” - kind of like when an engine stalls at a stoplight due to revs dropping too low because the ECM couldn’t respond fast enough to something outside of its expected behavior.
[2] Voltage spikes too high at the top of the “downswing” - i.e. transient spike immediately after workload drops off/changes. Although, this one might be more relevant to all-core workloads when amps are at their highest.
[3] Insufficient voltage for the frequency that one or more CPU cores (or Ring) are operating at.
[4] Noisy signals when operating at peak frequency. Again, for one or more CPU cores (or Ring). A CPU will have a frequency limit (due to silicon quality) no matter how much voltage you feed it.
[5] Voltage drops too low due to Vdroop.

I might be forgetting something, but that’s how I think of instability. So, by changing to a less aggressive LLC, I would hopefully be addressing scenario [2]. And by reducing the Adaptive Offset (which applies to all frequencies) I would hopefully be addressing [1] and [3]. [5] is not relevant - but you would know better than I based upon the apps you’re using. That just leaves us [4], which would require lowering the peak frequency for one or more cores, especially if they’re overclocked. If my “Hail Mary” adjustment improves things then I would try to narrow down which of the two settings contributed the most. However, if I see no improvement then the next step is to reset things and try one tick lower on the peak frequency. And if neither change worked then I might try all three changes at the same time.

One other thing to consider would be the use of the OCTVB feature, which could potentially be buggy, or just your settings are not stable enough. If you can’t solve your issues any other way then I would take a closer look at that setup.

 

[jamiesaun85]

What I was suggesting is something I would try in an effort to potentially save time in my investigation. I’d simply hope it would change behavior in a way that might tell me what’s going on inside the CPU.

Yes sir, I understand that. You'll have to forgive my diction and parlance. It gets me trouble quite a bit. I often come across as terse when in fact its completely unintentional. My questions are always genuine. As we interact more and more you'll see exactly what I mean. Another thing is my "statements" are actually more like questions usually. And Ill sprinkle questions throughout my writings that were meant to actually be questions, not pronouncements. I never state what i consider to be obvious as i opperate almost exclusiley on the benefit of the doubt, if that makes sense. So for example if I say or ask something that comes across as snarky, its not. Its a serious inquiry and I give whomever the benefit of the doubt that they already understand that. And yes, it does get me into trouble quite a bit. But I believe its important in todays world. Too many dumb kids with pink hair want me to police my language and say everything with an asterick, so that we dont hurt their feelings. I wont do it. They can just use their brain and learn to inteperate nuance and context.

Not that I was in trouble, Im just getting all that out of the way ahead of time. Because it will show up, if it hasnt already. Now tjat that is out of the way:

I was just making sure I was understanding you correctly, And I agree with you. I was somewhat heading that way myself. "On the ragged edge of stability," no. Definitely not. But something isn't quite right and I'll take any advice i get. Great analogy with the stalling engine, by the way.

So you have brought up two things in particular that caught my attention. One I have actually been wondering from the very begigning.

#2) spikes. Will a spike cause instability and/or a crash? I know that too low voltage will, obviously. But what about too high? If so that's odd to me. Id like to know more about why that is.
#4) noisy signals. Ya lost me on this one. Do you just mean its opperating outside of it capabiities or being pushed further that its capable of going?

Lastly, do you have any Idea what the page file issue was all about? and am I safe to ignore that one off "out of video memory" error?

 

[FlyingScot]

I certainly appreciate your candor, but I do like to be mentally challenged. So, please don’t worry about how you think you might come across in your writings. I, for one, have never read anything negative into them. In fact, I very much enjoy how you articulate your observations and lines of inquiry, deductive reasoning, etc. One does have to think like a detective when trying to solve computer problems. Quite often it turns out to be a crime of passion…your very own!! [i.e. trying to overclock too high or undervolt too low]

2) spikes. Will a spike cause instability and/or a crash?

That’s a suspicion on my part, and one that I think BuildZoid has considered when watching things on his oscilloscope. But I think it’s likely very uncommon even if it does occur. So, I do tend to think in terms of the other side of the spike, which is the VRM’s late reaction that causes a sudden voltage drop, presumably while the core frequencies are still high. It’s just easier to describe it the way I did in #2.

#4) noisy signals. Ya lost me on this one. Do you just mean its opperating outside of it capabiities or being pushed further that its capable of going?

I used the noisy signal analogy because I recently heard it described for DDR5 and I thought it was quite descriptive. While DDR4, especially Samsung B-Die chips, generally liked more and more voltage (as long as you kept them cool) DDR5 has some kind of electrical wall where more voltage makes stability worse. It likely causes “noise” in the signal carrying the ones and zeros between memory modules and memory controller.

When referring to noise with a CPU, I think in terms of a leakage of voltage outside of the transistor gates, which leads to data errors (i.e. WHEA errors and crashes). I would think this is a condition related to the silicon quality. Yep. In other words, pushing it outside of its comfort zone. Once the leakage starts, no amount of extra voltage helps. It fact it often makes things worse. More voltage = more pressure. We know that extra voltage can cause this leakage, either immediately or over time (i.e. degradation), but frequency must also be able to do this - or something similar - because most CPUs definitely have a frequency wall.

"On the ragged edge of stability," no. Definitely not. But something isn't quite right and I'll take any advice i get.

Maybe I used the wrong words when I wrote “ragged edge”. What I really should have said is maybe a “hidden edge condition“ kind of instability. For those who stability test memory timings, etc, they routinely see stress tests that can run 12hrs without issue and then in the thirteenth hour spit out an error. Achieving true 100% stability might actually be exceedingly difficult.

Case in point is my situation. My PC has been stable for 1-1/2 years and 70 hours of winter gaming, at least half of which is in the last couple of months. But I think I just found a way to provoke a WHEA error while in one particular game. I have set WHEA errors to open the CALC.exe (Windows Calculator) [because Microsoft, in their wisdom, got rid of the very useful Windows 7 Event Viewer notifications options] so I suddenly had my game window minimize to show me the calculator. The experience was quite jarring, which was my original intention, and I was quite confused until I realized what had just happened. I hadn’t seen a WHEA error in over 18 months. You gotta love computers. I’ll update you on my findings if my detective theory turns out to be the cause. It might be relevant.

Lastly, do you have any Idea what the page file issue was all about? and am I safe to ignore that one off "out of video memory" error?

The page file issue could be caused by an app with leaky memory. But I really don’t have a solid answer. I would perhaps just try adjusting the page file parameters. I’m old school, so I like to set a fixed size. The issue with the out of video memory was what prompted Intel and MSI to jack up the default voltage (via very high CPU Lite Load values) for CPUs that were not fully stable - both degraded ones and also some i9’s straight out of the box. So, if you take a page out of their book, it would suggest that a slight increase in voltage might be the answer. However, some UE5 games are just a bear when it comes to shader compiles. Increasing voltage dramatically just to satisty one rogue application/game or stress test might not be worth it. That’s always your call. I am a little concerned about the sluggishness inside app, though. That could be an indication of improper voltage during transients, maybe causing CEP to step in (if enabled) or even CEP-Lite as BuildZoid calls it. But I’d also consider taking a look at your C-States. And then there’s that Windows latency issue that some people experience on the new BIG.Little architectures of Alder Lake+ [I will try to find my video on that one].

 

[jamiesaun85]

excellent. I believe I will start with the voltage a bit and go from there. I'm tempted to limit c states to 8 also, but probably shouldn't do 2 adjustments at the same time.

That might be the strangest issue I've heard yet with the calculator launching itself.

 

[FlyingScot]

excellent. I believe I will start with the voltage a bit and go from there. I'm tempted to limit c states to 8 also, but probably shouldn't do 2 adjustments at the same time.

That might be the strangest issue I've heard yet with the calculator launching itself.

I'm looking forward to the results of your testing as we're still trying to learn from each other's experiences. And we usually end up having to write our own manuals and guides. I do that A LOT! Speaking of guides, I've attached my SOP on how to link the Calculator to a WHEA event. I find it very useful to be made aware of WHEA errors as soon as they happen. It also means that I don't have to keep HWInfo64 running all the time, which also tracks WHEA errors in real-time.

 

[Edginess0767]

Hi FlyingScot, thanks for the opportunity to share all the profiles for other user's BIOS settings.
I just added my result there on the online spreadsheet as Kevin Z, at row 37.
Do you have any questions regarding my BIOS setup?

 

[FlyingScot]

Hi FlyingScot, thanks for the opportunity to share all the profiles for other user's BIOS settings.
I just added my result there on the online spreadsheet as Kevin Z, at row 37.
Do you have any questions regarding my BIOS setup?

Hi there, Kevin!
Well, I never thought I would see a Gigabyte entry in my database. You must have worked hard to do the translation between the two BIOSes. Nice work!
I see that you chose a rather high P-Core Ratio of 60 all-core, but then used a -8 AVX offset. What was your thinking on that one? I would think that your -8 AVX offset would be hurting your R23 score, but it's still a nice score. Plus, it's obviously stable from all the stability testing you indicated that you had done. I assume that you implement all the latest microcode fixes, right?

 

[Edginess0767]

Hi there, Kevin!
Well, I never thought I would see a Gigabyte entry in my database. You must have worked hard to do the translation between the two BIOSes. Nice work!
I see that you chose a rather high P-Core Ratio of 60 all-core, but then used a -8 AVX offset. What was your thinking on that one? I would think that your -8 AVX offset would be hurting your R23 score, but it's still a nice score. Plus, it's obviously stable from all the stability testing you indicated that you had done. I assume that you implement all the latest microcode fixes, right?

Yeah, MSI's BIOS settings are rather different with lite load options with LLC levels.

For my max p-core, I just set them to all 60x because I do not want 2 cores goes to 60x and other at 56x.
Cinebench score is not affected by the avx offset which is not triggered. I mainly set it to be stable at prime 95 small fft. By BIOS default it is -5 for 56x for all cores so you are looking around 51x base except 55x for two 60x cores. Avx offset is only triggered with heavy avx workloads, unlike 11th gen where it got triggered when running any avx workload.

My BIOS has the 0x12B microcode. I could undervolt it to AC LL of 15 (around 1.22v in cinebench r23) but I do not want to just get on the edge of stability and then instability in summer. Btw, if Ac ll undervolt combine with adaptive offset, I will get voltage fluctuations in prime95 workloads.

 

[jamiesaun85]

Maybe I used the wrong words when I wrote “ragged edge”. What I really should have said is maybe a “hidden edge condition“ kind of instability. For those who stability test memory timings, etc, they routinely see stress tests that can run 12hrs without issue and then in the thirteenth hour spit out an error. Achieving true 100% stability might actually be exceedingly difficult.

No sir, I don't believe you did use the wrong term after all, not that I disagreed to begin with. Well, maybe a little, but I knew you were right to be skeptical, I,m the same way myself. Plus, I could tell using the settings that something wasnt quit right. Turns out this quote was quite prophetic of you. It was indeed on the ragged edge. To be sure, my last set up was {mostly} stable, safe, and sustainable I believe. And man was it fast. Having said that, It absolutely would crash if hammered hard enough, for long enough. Got me a new program called iMatch thats another media organizer and serial abuser of FFmpeg. To put it mildly. Man does it hammer the cpu, the memory and the disk. Like 100% utalization of a 990 pro, ive not seen that before.

Anyway, when scanning a very large database, it woiuld soft crash at least half the time. Also started getting up to about 91c and even blue screened a couple times. Oh, and WHEA erors up the ying yang, whatever those are. How bad they are, im not really sure. But we can agree that they're not a great sign, for sure.

Ive since diled some things back and have no problems with any of the above anymore. What do you think about utilizing the avx setrtings? Im not even sure if FFmpeg is AVX or even what AVX is. I just know its there as another offset for workloads that are particullarly punishing. Maybe using what I was, with an avx ofset of -1 or -2 would be the ticket?



Edit: Also, I did end up bumping the voltage up a touch after the warnings of odd behavior like sluggisg opening of apps and whatnot, and you were correct there too, i think it was only 5mv if i remeber right. Might have been 10mv, not very much at all, but it was enough. Fixed the mouse and apps and everything. Its just a little too wide open for punishing workloads.

 

[Edginess0767]

No sir, I don't believe you did use the wrong term after all, not that I disagreed to begin with. Well, maybe a little, but I knew you were right to be skeptical, I,m the same way myself. Plus, I could tell using the settings that something wasnt quit right. Turns out this quote was quite prophetic of you. It was indeed on the ragged edge. To be sure, my last set up was {mostly} stable, safe, and sustainable I believe. And man was it fast. Having said that, It absolutely would crash if hammered hard enough, for long enough. Got me a new program called iMatch thats another media organizer and serial abuser of FFmpeg. To put it mildly. Man does it hammer the cpu, the memory and the disk. Like 100% utalization of a 990 pro, ive not seen that before.

Anyway, when scanning a very large database, it woiuld soft crash at least half the time. Also started getting up to about 91c and even blue screened a couple times. Oh, and WHEA erors up the ying yang, whatever those are. How bad they are, im not really sure. But we can agree that they're not a great sign, for sure.

Ive since diled some things back and have no problems with any of the above anymore. What do you think about utilizing the avx setrtings? Im not even sure if FFmpeg is AVX or even what AVX is. I just know its there as another offset for workloads that are particullarly punishing. Maybe using what I was, with an avx ofset of -1 or -2 would be the ticket?



Edit: Also, I did end up bumping the voltage up a touch after the warnings of odd behavior like sluggisg opening of apps and whatnot, and you were correct there too, i think it was only 5mv if i remeber right. Might have been 10mv, not very much at all, but it was enough. Fixed the mouse and apps and everything. Its just a little too wide open for punishing workloads.

Did you test with prime95 small fft for stability at all? Or other test software such as aida64 or occt?

I think ffmpeg uses avx to encode depend on you CPU's supported instruction set.

What is avx: https://linustechtips.com/topic/805187-what-exactly-is-avx-and-do-i-need-it/

Explainer: What are MMX, SSE, and AVX?

Computer technology is no strange to acronyms: CPU, RAM, SSD, to name but a few. Ever so often, new ones appear on the scene in the never-ending...

www.techspot.com

 

[FlyingScot]

No sir, I don't believe you did use the wrong term after all, not that I disagreed to begin with. Well, maybe a little, but I knew you were right to be skeptical, I,m the same way myself. Plus, I could tell using the settings that something wasnt quit right. Turns out this quote was quite prophetic of you. It was indeed on the ragged edge. To be sure, my last set up was {mostly} stable, safe, and sustainable I believe. And man was it fast. Having said that, It absolutely would crash if hammered hard enough, for long enough. Got me a new program called iMatch thats another media organizer and serial abuser of FFmpeg. To put it mildly. Man does it hammer the cpu, the memory and the disk. Like 100% utalization of a 990 pro, ive not seen that before.

Anyway, when scanning a very large database, it woiuld soft crash at least half the time. Also started getting up to about 91c and even blue screened a couple times. Oh, and WHEA erors up the ying yang, whatever those are. How bad they are, im not really sure. But we can agree that they're not a great sign, for sure.

Ive since diled some things back and have no problems with any of the above anymore. What do you think about utilizing the avx setrtings? Im not even sure if FFmpeg is AVX or even what AVX is. I just know its there as another offset for workloads that are particullarly punishing. Maybe using what I was, with an avx ofset of -1 or -2 would be the ticket?

Oh boy, I see that you've been having quite a time wrestling your system into submission. That was quite some instability you ran in to. Of course, it really isn't too surprising being that you are deliberately dancing on the edge given your increasing knowledge of how things work. There's nothing quite like failure when it comes to learning new skills! Obviously, what has made things worse is that Mad Max of a program called iMatch. I wonder if it's the workload that deserves all those resources or whether it's just another badly coded app? Either way, you're perfectly within your right to just keep dialing back your performance desires until you get things stable again. A lot of CPUs do give you the impression that you have won the silicon lottery until you start pushing them in new and novel ways; then they have a tendency to fall on their butt.

Edit: Also, I did end up bumping the voltage up a touch after the warnings of odd behavior like sluggisg opening of apps and whatnot, and you were correct there too, i think it was only 5mv if i remeber right. Might have been 10mv, not very much at all, but it was enough. Fixed the mouse and apps and everything. Its just a little too wide open for punishing workloads.

That's good feedback. It's always nice to get "field reports" that match with your understanding of cause-and-effect. Thanks for that info! I'm looking forward to reading your next report, especially when you "really" finalize your settings. That might be a good time for another Raptor Lake Database entry.

 

[FlyingScot]

Yeah, MSI's BIOS settings are rather different with lite load options with LLC levels.

For my max p-core, I just set them to all 60x because I do not want 2 cores goes to 60x and other at 56x.
Cinebench score is not affected by the avx offset which is not triggered. I mainly set it to be stable at prime 95 small fft. By BIOS default it is -5 for 56x for all cores so you are looking around 51x base except 55x for two 60x cores. Avx offset is only triggered with heavy avx workloads, unlike 11th gen where it got triggered when running any avx workload.

My BIOS has the 0x12B microcode. I could undervolt it to AC LL of 15 (around 1.22v in cinebench r23) but I do not want to just get on the edge of stability and then instability in summer. Btw, if Ac ll undervolt combine with adaptive offset, I will get voltage fluctuations in prime95 workloads.

Interesting. I would have thought that R23 would make heavy use of AVX instructions. By the way, my understanding (from what i would consider a reliable source: Skatter) was that the AVX changes came into effect with Rocket Lake. I have a Rocket Lake so I could always test that understanding when I have some time. But I'm certainly open to being corrected. Here's what I wrote in my main undervolting guide. Would you care to critique it for me based upon your understanding and/or findings?

** Tip: In the Comet Lake (and earlier) days, AVX offsets could actually lead to instability due to it constantly cutting in an out and causing transient spikes. However, Intel have recently changed the logic to where light AVX workloads will no longer trigger the offset. This change in approach means that only the heavier AVX workloads will downclock the CPU if you use negative offsets. In addition, beginning with Rocket Lake, AVX offsets are now applied to each individual maximum core ratio rather than just applying to the factory All-core Turbo 2.0 ratio. While both these enhancements have the potential to increase the utility of AVX offsets, I still recommend trying to avoid using them if at all possible. It will ultimately make stability testing a little easier.

 

[Edginess0767]

Interesting. I would have thought that R23 would make heavy use of AVX instructions. By the way, my understanding (from what i would consider a reliable source: Skatter) was that the AVX changes came into effect with Rocket Lake. I have a Rocket Lake so I could always test that understanding when I have some time. But I'm certainly open to being corrected. Here's what I wrote in my main undervolting guide. Would you care to critique it for me based upon your understanding and/or findings?

** Tip: In the Comet Lake (and earlier) days, AVX offsets could actually lead to instability due to it constantly cutting in an out and causing transient spikes. However, Intel have recently changed the logic to where light AVX workloads will no longer trigger the offset. This change in approach means that only the heavier AVX workloads will downclock the CPU if you use negative offsets. In addition, beginning with Rocket Lake, AVX offsets are now applied to each individual maximum core ratio rather than just applying to the factory All-core Turbo 2.0 ratio. While both these enhancements have the potential to increase the utility of AVX offsets, I still recommend trying to avoid using them if at all possible. It will ultimately make stability testing a little easier.

I think I might remember it wrong, so it started with Rocket Lake.

Both R23 and 2024 version of cinebench apparently does not heavy load of AVX. That means they will let the cpu boost to max all core turbo.

https://www.reddit.com/r/overclocking/comments/tz5dr2

From this reddit post it seems like Cinebench r23 does not use 256 bit AVX (Or AVX2), this is probably same for Cinebench 2024.

For the other part of avx offset. Yes, the offset applied to per core max turbo, therefore -8 offset with 60x core turbo will result to 52x.

The problem for me is that, if I do avx offset of 0, P core could boost to more than 52x, which could in turn may need more vcore, and also due to the increased p core ratio, your cpu's temperature will get worse.

If no avx offset, the undervolt settings for 52x avx workload may not be stable anymore, due to more vcore needed and also your temperature is higher because the hottest p core basically determines your cpu's package temperature and whether cpu will throttle.

I could probably undervolt more if I set my avx offset to -10, or 50x in AVX2 workloads.

 

[FlyingScot]

I think I might remember it wrong, so it started with Rocket Lake.

Both R23 and 2024 version of cinebench apparently does not heavy load of AVX. That means they will let the cpu boost to max all core turbo.

https://www.reddit.com/r/overclocking/comments/tz5dr2

From this reddit post it seems like Cinebench r23 does not use 256 bit AVX (Or AVX2), this is probably same for Cinebench 2024.

For the other part of avx offset. Yes, the offset applied to per core max turbo, therefore -8 offset with 60x core turbo will result to 52x.

The problem for me is that, if I do avx offset of 0, P core could boost to more than 52x, which could in turn may need more vcore, and also due to the increased p core ratio, your cpu's temperature will get worse.

If no avx offset, the undervolt settings for 52x avx workload may not be stable anymore, due to more vcore needed and also your temperature is higher because the hottest p core basically determines your cpu's package temperature and whether cpu will throttle.

I could probably undervolt more if I set my avx offset to -10, or 50x in AVX2 workloads.

Okay. Cool. I think we're on the same page. I tend to not use AVX offsets (probably old habits) but if you find the newer AVX behavior less intrusive then it sounds like it can be useful. I'll have to take a closer look at the Cinebench apps. What you uncovered was definitely something I wasn't aware of. Perhaps that's another reason why jamiesaun85 ran into stability issues. Perhaps his new apps are really heavy AVX users. I always thought that if you could pass the Cinebench R23 test then most all-core workloads should be okay - or at least close to being okay - and outside of Prime95 of course, which causes way more vdroop. Typing quickly here...forgive the sentence structure...

 

[Edginess0767]

Okay. Cool. I think we're on the same page. I tend to not use AVX offsets (probably old habits) but if you find the newer AVX behavior less intrusive then it sounds like it can be useful. I'll have to take a closer look at the Cinebench apps. What you uncovered was definitely something I wasn't aware of. Perhaps that's another reason why jamiesaun85 ran into stability issues. Perhaps his new apps are really heavy AVX users. I always thought that if you could pass the Cinebench R23 test then most all-core workloads should be okay - or at least close to being okay - and outside of Prime95 of course, which causes way more vdroop. Typing quickly here...forgive the sentence structure...

I do not trust my cpu to be stable while undervolting unless I test prime95 small fft with avx on. Cinebench r15 runs consecutively without crash is good for fast test. But AVX workload needs more voltage at the same frequency to be stable. I usually test prime 95 small fft around minimum of 3 hours.

Another thing to consider is the difference of v/f curve between p core and e core. For example if you have a good p core SP but not e core SP, you e core cannot undervolt as low as p core while staying at the stock frequency.

Also, different level of LLC affect vdroop, for example, for medium LLC vs low LLC on my motherboard, for the similar vcore under cinebench, the difference of vdroop under prime95 small fft is around 20mv.

I would recommend that use a less droopy LLC that matches DC LL and change AC LL 5 at a time to see if system is stable first using cinebench r15 with 10 runs, and then occt, y-cruncher and prime 95 small fft to test them, prime95 needs at least 2-3 hours.

 

[jamiesaun85]

Did you test with prime95 small fft for stability at all? Or other test software such as aida64 or occt?

I think ffmpeg uses avx to encode depend on you CPU's supported instruction set.

What is avx: https://linustechtips.com/topic/805187-what-exactly-is-avx-and-do-i-need-it/

Explainer: What are MMX, SSE, and AVX?

Computer technology is no strange to acronyms: CPU, RAM, SSD, to name but a few. Ever so often, new ones appear on the scene in the never-ending...

www.techspot.com

Thank you for this. New information is always welcomed, I look forward to reading it.

 

[jamiesaun85]

I would recommend that use a less droopy LLC that matches DC LL and change AC LL 5 at a time to see if system is stable first using cinebench r15 with 10 runs, and then occt, y-cruncher and prime 95 small fft to test them, prime95 needs at least 2-3 hours.

That's exactly what I do, though no explanation why. Just makes sense in my brain I guess. And Everything else y'all said is above my head.

And no, I haven't ever once ran prime 95, and Ive never used a stress specific software of any kind. I just don't like doing that, It seems like it would be particularly taxing for something that could be irrelevant. But, if its checking AVX and that is indeed what I'm fighting with now, then I guess its not so irrelevant after all.

But no, generally i check stability just using the same programs i always use.

 

[Edginess0767]

That's exactly what I do, though no explanation why. Just makes sense in my brain I guess. And Everything else y'all said is above my head.

And no, I haven't ever once ran prime 95, and Ive never used a stress specific software of any kind. I just don't like doing that, It seems like it would be particularly taxing for something that could be irrelevant. But, if its checking AVX and that is indeed what I'm fighting with now, then I guess its not so irrelevant after all.

But no, generally i check stability just using the same programs i always use.

How is your PC now? Is it stable after further testing?