Diving into this hot and controversial topic - undervolting with CEP enabled!
I want to address the elephant in the room first - is disabling CEP potentially dangerous? The short answer is, probably not. I don't really know, and I'm not aware of any evidence that it could be harmful, especially if you have already set sensible settings in your BIOS. This is currently the widespread opinion online, also here in this forum, including with people like citay, who has had lots of experience. Arguing about whether or not CEP is necessary or not is not my goal with this post, I just want to share what I've learned and done.
This is also not intended to be a full guide on how to undervolt, including the basics. Citay's guide is extremely extensive and covers basically everything somebody new to this needs, to get started.
TL; DR - you can check some results and notes here
First a very short backstory, which might provide you with some context.
About a month ago I switched to a desktop PC with a 13700K, from a laptop with a 12900HX, and even before I ordered the components I was already aware of the 13/14 gen issues, so one of my goals from day one was to stick with the basics and follow the official recommendations provided by Intel. Most of them are considered good practice anyway, such as setting ICCMax, proper power limits, enabling C-States and using a power plan in Windows that allows downclocking. IA CEP being enabled is also part of Intel's recommendations, so that's something I made sure is on before I installed Windows, along with applying the rest of the recommended settings, where needed.
My first attempt at undervolting my 13700 was to lower the Lite Load mode as I had read somewhere it does wonders, but I immediately faced a performance hit caused by CEP. Then I read I had to disable CEP in order to properly undervolt using a Lite Load method, but as it was part of Intel's recommendations, I wanted to try a different approach first. With the 12900HX, the only way to undervolt was by using a negative offset as there was no advance BIOS available, so I already had some experience with setting offsets and I just defaulted to this. I tried it with the 13700K and it actually worked great (still does), lowered voltages across the board, temps and power draw noticeably, and there was no performance hit because of CEP.
My Cinebench R23 score with the default motherboard settings is around 29K pts at best, which is enough performance for me, but the problem is the instant thermal throttling at 100C, and hitting the 253W default PL2. Also, voltages spike to 1.46-1.47V during normal usage.
With a -0.125V offset my score went up to 30700 pts, with max power draw 225W and 1.25-1.26V under 225W load. I was happy with this setup so I used it for a few days without issues, then I tried a larger offset to see if it'd be okay. I went with -0.150V which was also perfectly stable, at some point I also set a conservative PL1=125W and PL2=188W and everything was great. Voltages were fine, sometimes spiking to 1.33, but generally under lighter load so no major worries with that. I had tested for stability using y-cruncher, Primer95, OCCT, R23, R24, TimeSpy, and last but not least, through gaming and normal usage, but I watched a Buildzoid video where he mentioned Cinebench R15 is very good at exposing instabilities, so I though I should test with it too. Sure enough, WHEA errors popped up after just 4-5 consecutive runs. I dropped the offset to -0.140V, and it is stable in R15.
Around the same time I started playing The Last of Us Part 1 and for the first time I got a bit concerned by the voltage I was seeing, as I was hitting 1.33-1.34V in-game, and averaging 1.32V, which didn't seem ideal. Just to clarify - it probably isn't a problem, but I wanted to try lower it a bit. So I started experimenting with different ways to lower the max VCore in gaming and also during lighter usage, while keeping CEP enabled. Even though I still have no idea whether it protects my CPU from anything, if I can achieve the results I want with it enabled, I don't see a reason to disable it.
Increasing the voltage offset was obviously not an option, because I had just decreased it from -0.150V to 0.140V. R15 causes me WHEAs when VCore starts hovering just below 1.18V at full load, and -0.150V puts me just in that range. Therefore, I knew what my target voltage under load is - at least 1.18V, but less than 1.19V, so now I needed to find a way to achieve that while maintaining performance, while decreasing the VCore under lighter load and gaming to 1.3V max.
CEP, AC/DC load lines and LLC
If I understand correctly, CEP is triggered by differences between the AC load line (set in mOhms) and the LLC mode (also corresponding to mOhms), where LLC determines how much Vdroop (drop in voltage during heavy CPU load) is being counteracted by the VRM. The AC value lets the CPU know what Vdroop it should expect, so that the CPU can properly calculate the voltage request it should send to the motherboard (at least in theory). If the AC tells the CPU it should expect "x" Vdroop under load, while the LLC allows for "x+5" Vdroop under load, then the CPU effectively gets more undervolted the higher the CPU load is. That's why undervolting by lowering the AC load line is so effective when benchmarking or running heavy loads - it hides from the CPU the fact that Vdroop is expected, so the CPU thinks it's okay with requesting lower voltage as assumes the motherboard will compensate the Vdroop.
If CEP is enabled, this is where it freaks out and starts clock stretching to prevent potential instability, even though the system might otherwise be completely stable and well-performing. This clock stretching effectively reduces the CPU's power and current draw, allowing it to remain stable at a lower voltage, which CEP considers unstable, because it is so much lower than what it expects to receive. So this is why R23 scores can drop by 50% even though you know the Lite Load mode you've selected is stable with your CPU. CEP is not triggered by offsets, because they shift the entire voltage-frequency curve of the CPU, so you can just make it request lower and lower voltages by applying a larger offset, until it is simply unstable. CEP will not kick in as it won't detect a difference between the requested voltage and the supplied one.
However, CEP also seems to have a buffer zone and doesn't kick in unless AC drops to somewhere below ≈67% of the LLC impedance. You can lower the AC load line only, without having a performance hit caused by CEP, just not by much.
The DC load line doesn't directly affect voltage, what it does is to calibrate the power measurement done by the CPU. The DC value in mOhms should match the LLC's impedance in mOhms, so that ideally, when DC and LLC are properly calibrated, VID=voltage supplied to CPU. This ensures proper power measurement, which is especially important if you have a power limit set that's always hit under full load. If DC is set too low, VID will be inaccurately higher, which will lead to inaccurately high power measurement, so you'd effectively power throttle your CPU, on top of the power limits you have set. If DC is set too high, then the VID will be inaccurately lower, which can turn your 200W PL2 into a 205W one, for example. Small differences probably won't be noticeable, but that's the general idea.
So, with all that in mind, what options do we have to undervolt when CEP is enabled, besides just by setting an offset? We have to abide by one general rule - AC should not be set to a value that's below ≈70% of DC=LLC. It sounds simple enough, but it has implications.
If we want to reduce AC to a value similar to a relatively low Lite Load mode, let's say to AC=20=0.2 mOhms (as Lite Load 5 does), DC=LLC cannot be set higher than 20/0,7 = 0.28 mOhms (rounded down). But we have to keep in mind that LLC is applied using presets, so we have a limited number of options for DC, if we want to properly match it to a given LLC mode. Also, going to a lower (as in number, e.g from 8 -> 4) LLC mode (on MSI motherboards, on Asus, e.g., it's the opposite), means that you are requesting from the VRM to compensate more for the Vdroop. To do that, the VRM has to artificially boost the voltage to the CPU when the CPU is under load, but when the load suddenly goes away, this additional voltage applied by the VRM can cause a sudden voltage spike that shoots above the CPU's target VID (called an overshoot), which technically has the potential to be harmful overtime, as it can deliver excess voltage to the CPU. How big the risk is depends a lot on the quality of the motherboard, but it is a risk nonetheless. This exact topic is not something I've researched too much, but the general consensus is that for most people an LLC mode that allows a healthy amount of Vdroop is the better option. I'll appreciate comments on this from people who are using flat LLC or strong modes, what is your experience and setup, and what benefits do you find in this.
Going back to the lowering AC with CEP enabled problem, the above would mean that we have a narrow window to work with for DC=LLC, in my opinion somewhere between 0.4 - 0.7 mOhms. Any lower than that, you'd be asking the VRM for a significant Vdroop compensation. Any higher than that, you can just go with the default DC=110=LLC=Auto, and you don't have to worry about matching DC to LLC, but at the same time you can't lower AC as much as you might want to.
But if you want to worry about matching them... (like me), see below.
With the latest bioses, especially the ones with 0x129 microcode, MSI's motherboards mostly (if not exclusively?) default to the "Intel Default" settings, which have AC=DC=110 (1.1 mOhms) and LLC on Auto. What this should mean is that DC=110=1.1 mOhms is calibrated for LLC=Auto. An important note here is that I've tested LLC=Auto and LLC=8 on my motherboard, and they have the exact same Vdroop behaviour, and other people,with different MSI motherboards, such as the Z790 Tomahawk, have also confirmed the same.
So, this means that with DC=110 (1.1 mOhms) and LLC=Auto=8, VID should match the voltage supplied to the CPU, right?
On mine, and many other MSI motherboards, the only sensor which is available to us for checking the voltage supplied to the CPU is VCore. Unfortunately, it is said to not be completely accurate. According to user SgtMorogan (but not only) on the overclock.net forum, "Vcore will always read somewhat higher than reality due to the impedance between the die and the sensor.". This can be found in this topic, which is widely shared in MSI motherboard-related discussions online. In there, you can find two different tables with supposed impedances, one for Z690 motherboards and one for Z790, with different values in mOhms across the LLC modes. One user with a Z790 Tomahawk board has tested different LLC modes and calculated the supposedly matching DC values. What's interesting is that according to him, LLC=8 pairs with DC=98 (0.98 mOhms), not 110 (1.1 mOhms), as we might assume, given the default settings and the fact that LLC=Auto=8. Additionally, in the same thread, on page 3, user FR4GGL3 has shared the following:
"I asked MSI a few weeks ago. The Questioan was which exact Numbers in mOhms equal to the 1 to 8 Settings of LLC in the Bios.
The answer was:
The “CPU Loadline Calibration Control” settings (Auto, Mode 1 to 8) are fine tune results by RD team’s know-how, so please allow us to keep them secret.
The Auto setting would meet the Intel suggested values.
If user wants less voltage drop (more voltage compensation) when CPU is under high loading, please select Mode 1.
The bigger Mode number the more voltage drop.
So I would say "Auto" is 1.1 mOhms. At least on my Z690 Board. That is also what is listed here on the first few entries"
When I put full load on the CPU using the Intel Default profile with AC=DC=110 and LLC=Auto, VCore always reads higher than VID. I logged data via HWInfo and calculated the average differences across a few short runs of OCCT and R23, by first calculating the difference between VCore and VID for each polling point, and then the average difference, and the result is almost always exactly 0.013V, or 13mV. The runs based on which I've calculated this begin at PL2 and then PL1 kicks in, and I've taken the average of the VCore-VID difference based on all data. But even if I only review the PL2 or PL1 data separately, it is almost always exactly a 0.013V difference, +-2-3mV at most. Setting DC to 98-100 actually causes VID to almost perfectly match VCore. So what does this mean?
Option 1 - assuming that MSI have properly calibrated LLC=Auto to DC=110, being the default, then VCore is indeed inherently inaccurate and always shows higher than it should, about 0.013V higher on average, at least on my motherboard.
Option 2 - if MSI are incorrectly defaulting to DC=110, while LLC=Auto being 0.98-1.0 mOhms, this would more or less explain the lower VID compared to VCore at stock configuration.
I am willing to trust that MSI have not been incorrectly setting DC and LLC by default, as this doesn't even have to do anything with Intel. So, trusting the default settings means that if I want to change LLC to another mode and calibrate DC accordingly, I have to aim for the same 0.013V difference between VCore and VID that I'm seeing with the stock configuration. After some trial and error, I've found out that on my motherboard, LLC=6 paired with DC=68, achieves the same 0.013V average difference as 110/LLC=Auto, under the same conditions.
In order for VID to match Vcore with LLC=6, DC should be set to ≈60, but I've found this impacts performance by a small margin, and I believe it's because it's effectively lowering my PL2 limit.
So, to recap:
- Lowering the AC load line, while keeping LLC=DC=110=1.1 mOhms, is basically what the Lite Load modes do and it's especially effective when high load is put on the CPU. A lot of Vdroop is allowed, but the CPU doesn't know it, so it's not asking for voltage to compensate for it, leading to a significant undervolt during high-load. CEP doesn't like that so it starts slowing down the CPU and reducing the power and current going to it.
- We can undervolt with CEP enabled, it's just more complex and requires a different approach.
- The ground rule is that AC cannot be <70% of DC/LLC; and DC should be calibrated to LLC, so that the VID-Vcore relation is the same as when using the default settings, after measuring it with the most precise sensor you have available.
- Alternatively, you could just go with VID=VCore, as even if this leads to higher inaccurate power reading, you could simply bump up your power limits by a few watts and nobody has to know about it.
- We could technically go as low as we want with AC, as long as we don't break the above rule, but this naturally means that LLC also has to be made stronger (compensate more). Going too low with AC will quickly require an almost flat LLC, which is generally not recommended for most people unless you really know how to set it up and have a good high-end motherboard. It also has other implications too, but I won't go into details.
If we don't want to set a very strong LLC, we have to keep AC at 30-35 the lowest, so that we can set DC=LLC to at least 40. I have not experimented with this range, but went for 1-2 steps above, aiming for LLC=6. It still allows for healthy Vdroop and doesn't have too much compensation. As mentioned above, it seems to match with DC=68, at least as long as I can trust the measurements.
I mentioned that the AC load line undervolt method works the best under high CPU load. This is because even though reducing AC also impacts the VID calculation without load, due to some mysterious way the CPU calculates its VID - using "predicted current", a lowered AC doesn't have the same great undervolting effect when the CPU load is not high enough to induce Vdroop. At least this is how I interpret it. So, what you end up with is higher voltage during light load compared to when you undervolt using an offset, and this can become especially noticeable during gaming. To counteract this, we can combine the two and add a negative offset to a lowered AC load line. This gives us a lower base VID + offset (config 3 below); or slightly lower base VID + surprise Vdroop for the CPU + offset (config 2 below).
I've tested 3 different undervolt configurations, all with CEP enabled, and have compared them with the default Lite Load 5 preset, with CEP disabled. The results illustrate well the benefits of each undervolting method. Here is an Excel file with all the test results, baseline information and some notes.
Config A is with the "Intel Default" lite load profile, with AC=DC=110, LLC on Auto and adaptive+negative offset set to -0.140V. This is my OG setup which I still like due to its simplicity and generally good results. Its only problem is the 1.33-1.34V spikes that can happen during gaming (in specific games).
Config B is a slightly modified version of config A, exploting CEP's buffer zone. Here, AC=80, DC=110 and LLC=Auto. Because AC is reduced from 110 to 80, I've also reduced the offset a bit to -0.125V, and this gives me almost the same VCore under load, but max VCore is lower due to the lower AC, which doesn't cause the CPU to calculate as high VID requests anymore. No impact in performance compared to config 1.
Config C is an experimental one where AC=DC=68=LLC6 (set based on the described above) and again an -0.125V offset. Here we have less VDroop, but also AC is set lower, so the same offset of -0.125V puts me at more or less the same VCore under load as config A and B. However, during light load this gives me even lower max VCore spikes. No impact in performance compared to configs A or B.
Config D is just Lite Load 5 with CEP disabled, so AC=20/DC=110 and LLC=Auto. This gives me higher max VCore spikes than config B and C, but generally performs slightly better at full 188W load. You will see in the file that in Cinebench R23 LL5 achieves on average around 100-150 pts higher result compared to the other setups, but this is not a significant difference. The most potential it has is in an OCCT-like workload, where LL5 could draw noticeably less power, but this seems to be dependent on the specific type of load. I should also note that this is the lowest perfectly stable Lite Load mode for my CPU, as with LL3 CB R24 crashes soon after I start it, and I don't think LL4 will be stable in R15, as the Vcore with it drops to the low 1.170s.
Cinebench R23
This is an interesting one because all four configurations perform similar to each other, but with clear differences based on the power limit.
- At 188W, config D (LL5) has higher average effective clocks compared to the rest, by about 50MHz for the P and E cores, therefore scores a bit higher.
- At 125W, the situation changes and configs A-C perform better, with higher average effective clocks. This sets a trend - the lower the load is, the better the offset configurations perform compared to the Lite Load one.
- The short run R23 scores were very close to each other, with configs A-C being around 30200 pts, and LL5 around 30300 pts.
OCCT Stability test
Here the Lite Load 5 setup is a clear winner at PL2, and it seems that in a heavy load of the type OCCT generates, AC<DC configurations excel due to the large unpredicted VDroop. Because of the low AC value, the CPU doesn't expect much Vdroop, but the OCCT load seems to cause a lot of it, so the bigger the difference between AC and DC/LLC is, the lower the VCore will be.
One thing to note is that the E cores didn't go past 4.1GHz with LL5, while they got up to 4.2GHz using the other three configurations.
Also, I don't understand the mechanism behind it, but the LL5 configuration had a significantly lower power draw at PL2 - 13W less than the runner up, config B.
Config B, where AC<DC=LLC is at second place at PL2, so it seems the AC load line undervolting is definitely the way to go if your use cases generate CPU loads similar to the ones OCCT does.
At PL1, they all effectively perform the same.
Geekbench 6
I tested this because it's a very light load for the most part, but with sharp load spikes here and there, so I thought it'd be a good test of max spikes in Vcore, current and power draw.
Here we also see that the two configurations with DC/LLC=110 + an offset see much lower max power draw spikes compared to the LL5 preset and the DC=LLC=68 + offset modes. LL5 has the highest average VCore, while the VCore spikes are within 10mV range across the four configurations.
Scores were within margin of error, around 2990 pts for single core and 19680 pts for multi core.
The win goes to config B for having the lowest metrics across the board.
Assassin's Creed Odyssey
In this game, Lite Load 5 has by far the highest average Vcore. This resembles the higher average Vcore during Geekbench 6, and is maybe related to the lower average current and/or power draw in these two scenarios. This is also typical during general usage without heavy load. LL5 always maintains the highest average VCore, because there is no offset applied to the V/F curve, and the low AC load line doesn't lead to much of an undervolt during low-load scenarios, when no VDroop is happening.
The win goes to B or C because of the lowest average VCore.
The Last of Us Part 1
In The Last of Us, this time config A, the 110/110 + offset configuration, had the highest average Vcore. Config D/Lite Load 5 still has the second highest average Vcore, and perhaps this game's CPU load is a middle ground where the VDroop is high enough for config D to have lower average Vcore than config A, but not high enough so that the lack of a V/F offset is compensated enough to match config B and D.
The win goes to B or C because of the lowest average VCore.
Conclusions:
Can we undervolt with CEP enabled - definitely! It is certainly more complicated and finicky compared to simply reducing AC and disabling CEP, as there are now multiple parameters to account for - AC, DC, LLC, and offset. But the results can be very good, performance is almost identical compared to Lite Load 5, and the voltage is lower in gaming and light usage.
In Cinebench R23, LL5/config D technically performs the best, no doubt about it, but the performance difference is so negligible it can never be felt. However, LL5 had a significant advantage in the OCCT stability test. Lower VCore, lower power draw, lower temperature, it was a clear winner there. This brings me to a conclusion I never though might be the case - perhaps, there is no best undervolt method (even complexity aside). Some will give you lower voltage in gaming and light usage, others will excel in specific workloads that tax the CPU a certain way. At least this is how I interpret my results, which I admit, are not based on an extensive suite of benchmarks and tests. I could go back and do additional tests with the same configurations, probably first on my list would be a 10-minute R23 run and a 10 minute R24 run with each, but this would take me a lot of time.
Anyway, another thing I think is visible is that basically all four configurations are very capable, and I'm quite happy with the results overall. Cofigurations B and C are the most interesting to me because they combine a reduced AC load line with an offset, and mix the best of both worlds. I think they're great for most people, as they provide good performance and temperatures, and lower the overall max VCore. But the very big difference between AC and DC/LLC that's present with LLC5 seems to be the best choice for optimizing power draw and temperatures, for anybody whose use case is heavy CPU loads such as OCCT, which create heavy Vdroop scenarios. For some reason the same doesn't apply to R23, so if somebody has an idea what's causing this different behaviour, please share.
Hope you enjoyed the read!
I want to address the elephant in the room first - is disabling CEP potentially dangerous? The short answer is, probably not. I don't really know, and I'm not aware of any evidence that it could be harmful, especially if you have already set sensible settings in your BIOS. This is currently the widespread opinion online, also here in this forum, including with people like citay, who has had lots of experience. Arguing about whether or not CEP is necessary or not is not my goal with this post, I just want to share what I've learned and done.
This is also not intended to be a full guide on how to undervolt, including the basics. Citay's guide is extremely extensive and covers basically everything somebody new to this needs, to get started.
TL; DR - you can check some results and notes here
First a very short backstory, which might provide you with some context.
About a month ago I switched to a desktop PC with a 13700K, from a laptop with a 12900HX, and even before I ordered the components I was already aware of the 13/14 gen issues, so one of my goals from day one was to stick with the basics and follow the official recommendations provided by Intel. Most of them are considered good practice anyway, such as setting ICCMax, proper power limits, enabling C-States and using a power plan in Windows that allows downclocking. IA CEP being enabled is also part of Intel's recommendations, so that's something I made sure is on before I installed Windows, along with applying the rest of the recommended settings, where needed.
My first attempt at undervolting my 13700 was to lower the Lite Load mode as I had read somewhere it does wonders, but I immediately faced a performance hit caused by CEP. Then I read I had to disable CEP in order to properly undervolt using a Lite Load method, but as it was part of Intel's recommendations, I wanted to try a different approach first. With the 12900HX, the only way to undervolt was by using a negative offset as there was no advance BIOS available, so I already had some experience with setting offsets and I just defaulted to this. I tried it with the 13700K and it actually worked great (still does), lowered voltages across the board, temps and power draw noticeably, and there was no performance hit because of CEP.
My Cinebench R23 score with the default motherboard settings is around 29K pts at best, which is enough performance for me, but the problem is the instant thermal throttling at 100C, and hitting the 253W default PL2. Also, voltages spike to 1.46-1.47V during normal usage.
With a -0.125V offset my score went up to 30700 pts, with max power draw 225W and 1.25-1.26V under 225W load. I was happy with this setup so I used it for a few days without issues, then I tried a larger offset to see if it'd be okay. I went with -0.150V which was also perfectly stable, at some point I also set a conservative PL1=125W and PL2=188W and everything was great. Voltages were fine, sometimes spiking to 1.33, but generally under lighter load so no major worries with that. I had tested for stability using y-cruncher, Primer95, OCCT, R23, R24, TimeSpy, and last but not least, through gaming and normal usage, but I watched a Buildzoid video where he mentioned Cinebench R15 is very good at exposing instabilities, so I though I should test with it too. Sure enough, WHEA errors popped up after just 4-5 consecutive runs. I dropped the offset to -0.140V, and it is stable in R15.
Around the same time I started playing The Last of Us Part 1 and for the first time I got a bit concerned by the voltage I was seeing, as I was hitting 1.33-1.34V in-game, and averaging 1.32V, which didn't seem ideal. Just to clarify - it probably isn't a problem, but I wanted to try lower it a bit. So I started experimenting with different ways to lower the max VCore in gaming and also during lighter usage, while keeping CEP enabled. Even though I still have no idea whether it protects my CPU from anything, if I can achieve the results I want with it enabled, I don't see a reason to disable it.
Increasing the voltage offset was obviously not an option, because I had just decreased it from -0.150V to 0.140V. R15 causes me WHEAs when VCore starts hovering just below 1.18V at full load, and -0.150V puts me just in that range. Therefore, I knew what my target voltage under load is - at least 1.18V, but less than 1.19V, so now I needed to find a way to achieve that while maintaining performance, while decreasing the VCore under lighter load and gaming to 1.3V max.
CEP, AC/DC load lines and LLC
If I understand correctly, CEP is triggered by differences between the AC load line (set in mOhms) and the LLC mode (also corresponding to mOhms), where LLC determines how much Vdroop (drop in voltage during heavy CPU load) is being counteracted by the VRM. The AC value lets the CPU know what Vdroop it should expect, so that the CPU can properly calculate the voltage request it should send to the motherboard (at least in theory). If the AC tells the CPU it should expect "x" Vdroop under load, while the LLC allows for "x+5" Vdroop under load, then the CPU effectively gets more undervolted the higher the CPU load is. That's why undervolting by lowering the AC load line is so effective when benchmarking or running heavy loads - it hides from the CPU the fact that Vdroop is expected, so the CPU thinks it's okay with requesting lower voltage as assumes the motherboard will compensate the Vdroop.
If CEP is enabled, this is where it freaks out and starts clock stretching to prevent potential instability, even though the system might otherwise be completely stable and well-performing. This clock stretching effectively reduces the CPU's power and current draw, allowing it to remain stable at a lower voltage, which CEP considers unstable, because it is so much lower than what it expects to receive. So this is why R23 scores can drop by 50% even though you know the Lite Load mode you've selected is stable with your CPU. CEP is not triggered by offsets, because they shift the entire voltage-frequency curve of the CPU, so you can just make it request lower and lower voltages by applying a larger offset, until it is simply unstable. CEP will not kick in as it won't detect a difference between the requested voltage and the supplied one.
However, CEP also seems to have a buffer zone and doesn't kick in unless AC drops to somewhere below ≈67% of the LLC impedance. You can lower the AC load line only, without having a performance hit caused by CEP, just not by much.
The DC load line doesn't directly affect voltage, what it does is to calibrate the power measurement done by the CPU. The DC value in mOhms should match the LLC's impedance in mOhms, so that ideally, when DC and LLC are properly calibrated, VID=voltage supplied to CPU. This ensures proper power measurement, which is especially important if you have a power limit set that's always hit under full load. If DC is set too low, VID will be inaccurately higher, which will lead to inaccurately high power measurement, so you'd effectively power throttle your CPU, on top of the power limits you have set. If DC is set too high, then the VID will be inaccurately lower, which can turn your 200W PL2 into a 205W one, for example. Small differences probably won't be noticeable, but that's the general idea.
So, with all that in mind, what options do we have to undervolt when CEP is enabled, besides just by setting an offset? We have to abide by one general rule - AC should not be set to a value that's below ≈70% of DC=LLC. It sounds simple enough, but it has implications.
If we want to reduce AC to a value similar to a relatively low Lite Load mode, let's say to AC=20=0.2 mOhms (as Lite Load 5 does), DC=LLC cannot be set higher than 20/0,7 = 0.28 mOhms (rounded down). But we have to keep in mind that LLC is applied using presets, so we have a limited number of options for DC, if we want to properly match it to a given LLC mode. Also, going to a lower (as in number, e.g from 8 -> 4) LLC mode (on MSI motherboards, on Asus, e.g., it's the opposite), means that you are requesting from the VRM to compensate more for the Vdroop. To do that, the VRM has to artificially boost the voltage to the CPU when the CPU is under load, but when the load suddenly goes away, this additional voltage applied by the VRM can cause a sudden voltage spike that shoots above the CPU's target VID (called an overshoot), which technically has the potential to be harmful overtime, as it can deliver excess voltage to the CPU. How big the risk is depends a lot on the quality of the motherboard, but it is a risk nonetheless. This exact topic is not something I've researched too much, but the general consensus is that for most people an LLC mode that allows a healthy amount of Vdroop is the better option. I'll appreciate comments on this from people who are using flat LLC or strong modes, what is your experience and setup, and what benefits do you find in this.
Going back to the lowering AC with CEP enabled problem, the above would mean that we have a narrow window to work with for DC=LLC, in my opinion somewhere between 0.4 - 0.7 mOhms. Any lower than that, you'd be asking the VRM for a significant Vdroop compensation. Any higher than that, you can just go with the default DC=110=LLC=Auto, and you don't have to worry about matching DC to LLC, but at the same time you can't lower AC as much as you might want to.
But if you want to worry about matching them... (like me), see below.
With the latest bioses, especially the ones with 0x129 microcode, MSI's motherboards mostly (if not exclusively?) default to the "Intel Default" settings, which have AC=DC=110 (1.1 mOhms) and LLC on Auto. What this should mean is that DC=110=1.1 mOhms is calibrated for LLC=Auto. An important note here is that I've tested LLC=Auto and LLC=8 on my motherboard, and they have the exact same Vdroop behaviour, and other people,with different MSI motherboards, such as the Z790 Tomahawk, have also confirmed the same.
So, this means that with DC=110 (1.1 mOhms) and LLC=Auto=8, VID should match the voltage supplied to the CPU, right?
On mine, and many other MSI motherboards, the only sensor which is available to us for checking the voltage supplied to the CPU is VCore. Unfortunately, it is said to not be completely accurate. According to user SgtMorogan (but not only) on the overclock.net forum, "Vcore will always read somewhat higher than reality due to the impedance between the die and the sensor.". This can be found in this topic, which is widely shared in MSI motherboard-related discussions online. In there, you can find two different tables with supposed impedances, one for Z690 motherboards and one for Z790, with different values in mOhms across the LLC modes. One user with a Z790 Tomahawk board has tested different LLC modes and calculated the supposedly matching DC values. What's interesting is that according to him, LLC=8 pairs with DC=98 (0.98 mOhms), not 110 (1.1 mOhms), as we might assume, given the default settings and the fact that LLC=Auto=8. Additionally, in the same thread, on page 3, user FR4GGL3 has shared the following:
"I asked MSI a few weeks ago. The Questioan was which exact Numbers in mOhms equal to the 1 to 8 Settings of LLC in the Bios.
The answer was:
The “CPU Loadline Calibration Control” settings (Auto, Mode 1 to 8) are fine tune results by RD team’s know-how, so please allow us to keep them secret.
The Auto setting would meet the Intel suggested values.
If user wants less voltage drop (more voltage compensation) when CPU is under high loading, please select Mode 1.
The bigger Mode number the more voltage drop.
So I would say "Auto" is 1.1 mOhms. At least on my Z690 Board. That is also what is listed here on the first few entries"
When I put full load on the CPU using the Intel Default profile with AC=DC=110 and LLC=Auto, VCore always reads higher than VID. I logged data via HWInfo and calculated the average differences across a few short runs of OCCT and R23, by first calculating the difference between VCore and VID for each polling point, and then the average difference, and the result is almost always exactly 0.013V, or 13mV. The runs based on which I've calculated this begin at PL2 and then PL1 kicks in, and I've taken the average of the VCore-VID difference based on all data. But even if I only review the PL2 or PL1 data separately, it is almost always exactly a 0.013V difference, +-2-3mV at most. Setting DC to 98-100 actually causes VID to almost perfectly match VCore. So what does this mean?
Option 1 - assuming that MSI have properly calibrated LLC=Auto to DC=110, being the default, then VCore is indeed inherently inaccurate and always shows higher than it should, about 0.013V higher on average, at least on my motherboard.
Option 2 - if MSI are incorrectly defaulting to DC=110, while LLC=Auto being 0.98-1.0 mOhms, this would more or less explain the lower VID compared to VCore at stock configuration.
I am willing to trust that MSI have not been incorrectly setting DC and LLC by default, as this doesn't even have to do anything with Intel. So, trusting the default settings means that if I want to change LLC to another mode and calibrate DC accordingly, I have to aim for the same 0.013V difference between VCore and VID that I'm seeing with the stock configuration. After some trial and error, I've found out that on my motherboard, LLC=6 paired with DC=68, achieves the same 0.013V average difference as 110/LLC=Auto, under the same conditions.
In order for VID to match Vcore with LLC=6, DC should be set to ≈60, but I've found this impacts performance by a small margin, and I believe it's because it's effectively lowering my PL2 limit.
So, to recap:
- Lowering the AC load line, while keeping LLC=DC=110=1.1 mOhms, is basically what the Lite Load modes do and it's especially effective when high load is put on the CPU. A lot of Vdroop is allowed, but the CPU doesn't know it, so it's not asking for voltage to compensate for it, leading to a significant undervolt during high-load. CEP doesn't like that so it starts slowing down the CPU and reducing the power and current going to it.
- We can undervolt with CEP enabled, it's just more complex and requires a different approach.
- The ground rule is that AC cannot be <70% of DC/LLC; and DC should be calibrated to LLC, so that the VID-Vcore relation is the same as when using the default settings, after measuring it with the most precise sensor you have available.
- Alternatively, you could just go with VID=VCore, as even if this leads to higher inaccurate power reading, you could simply bump up your power limits by a few watts and nobody has to know about it.
- We could technically go as low as we want with AC, as long as we don't break the above rule, but this naturally means that LLC also has to be made stronger (compensate more). Going too low with AC will quickly require an almost flat LLC, which is generally not recommended for most people unless you really know how to set it up and have a good high-end motherboard. It also has other implications too, but I won't go into details.
If we don't want to set a very strong LLC, we have to keep AC at 30-35 the lowest, so that we can set DC=LLC to at least 40. I have not experimented with this range, but went for 1-2 steps above, aiming for LLC=6. It still allows for healthy Vdroop and doesn't have too much compensation. As mentioned above, it seems to match with DC=68, at least as long as I can trust the measurements.
I mentioned that the AC load line undervolt method works the best under high CPU load. This is because even though reducing AC also impacts the VID calculation without load, due to some mysterious way the CPU calculates its VID - using "predicted current", a lowered AC doesn't have the same great undervolting effect when the CPU load is not high enough to induce Vdroop. At least this is how I interpret it. So, what you end up with is higher voltage during light load compared to when you undervolt using an offset, and this can become especially noticeable during gaming. To counteract this, we can combine the two and add a negative offset to a lowered AC load line. This gives us a lower base VID + offset (config 3 below); or slightly lower base VID + surprise Vdroop for the CPU + offset (config 2 below).
I've tested 3 different undervolt configurations, all with CEP enabled, and have compared them with the default Lite Load 5 preset, with CEP disabled. The results illustrate well the benefits of each undervolting method. Here is an Excel file with all the test results, baseline information and some notes.
Config A is with the "Intel Default" lite load profile, with AC=DC=110, LLC on Auto and adaptive+negative offset set to -0.140V. This is my OG setup which I still like due to its simplicity and generally good results. Its only problem is the 1.33-1.34V spikes that can happen during gaming (in specific games).
Config B is a slightly modified version of config A, exploting CEP's buffer zone. Here, AC=80, DC=110 and LLC=Auto. Because AC is reduced from 110 to 80, I've also reduced the offset a bit to -0.125V, and this gives me almost the same VCore under load, but max VCore is lower due to the lower AC, which doesn't cause the CPU to calculate as high VID requests anymore. No impact in performance compared to config 1.
Config C is an experimental one where AC=DC=68=LLC6 (set based on the described above) and again an -0.125V offset. Here we have less VDroop, but also AC is set lower, so the same offset of -0.125V puts me at more or less the same VCore under load as config A and B. However, during light load this gives me even lower max VCore spikes. No impact in performance compared to configs A or B.
Config D is just Lite Load 5 with CEP disabled, so AC=20/DC=110 and LLC=Auto. This gives me higher max VCore spikes than config B and C, but generally performs slightly better at full 188W load. You will see in the file that in Cinebench R23 LL5 achieves on average around 100-150 pts higher result compared to the other setups, but this is not a significant difference. The most potential it has is in an OCCT-like workload, where LL5 could draw noticeably less power, but this seems to be dependent on the specific type of load. I should also note that this is the lowest perfectly stable Lite Load mode for my CPU, as with LL3 CB R24 crashes soon after I start it, and I don't think LL4 will be stable in R15, as the Vcore with it drops to the low 1.170s.
Cinebench R23
This is an interesting one because all four configurations perform similar to each other, but with clear differences based on the power limit.
- At 188W, config D (LL5) has higher average effective clocks compared to the rest, by about 50MHz for the P and E cores, therefore scores a bit higher.
- At 125W, the situation changes and configs A-C perform better, with higher average effective clocks. This sets a trend - the lower the load is, the better the offset configurations perform compared to the Lite Load one.
- The short run R23 scores were very close to each other, with configs A-C being around 30200 pts, and LL5 around 30300 pts.
OCCT Stability test
Here the Lite Load 5 setup is a clear winner at PL2, and it seems that in a heavy load of the type OCCT generates, AC<DC configurations excel due to the large unpredicted VDroop. Because of the low AC value, the CPU doesn't expect much Vdroop, but the OCCT load seems to cause a lot of it, so the bigger the difference between AC and DC/LLC is, the lower the VCore will be.
One thing to note is that the E cores didn't go past 4.1GHz with LL5, while they got up to 4.2GHz using the other three configurations.
Also, I don't understand the mechanism behind it, but the LL5 configuration had a significantly lower power draw at PL2 - 13W less than the runner up, config B.
Config B, where AC<DC=LLC is at second place at PL2, so it seems the AC load line undervolting is definitely the way to go if your use cases generate CPU loads similar to the ones OCCT does.
At PL1, they all effectively perform the same.
Geekbench 6
I tested this because it's a very light load for the most part, but with sharp load spikes here and there, so I thought it'd be a good test of max spikes in Vcore, current and power draw.
Here we also see that the two configurations with DC/LLC=110 + an offset see much lower max power draw spikes compared to the LL5 preset and the DC=LLC=68 + offset modes. LL5 has the highest average VCore, while the VCore spikes are within 10mV range across the four configurations.
Scores were within margin of error, around 2990 pts for single core and 19680 pts for multi core.
The win goes to config B for having the lowest metrics across the board.
Assassin's Creed Odyssey
In this game, Lite Load 5 has by far the highest average Vcore. This resembles the higher average Vcore during Geekbench 6, and is maybe related to the lower average current and/or power draw in these two scenarios. This is also typical during general usage without heavy load. LL5 always maintains the highest average VCore, because there is no offset applied to the V/F curve, and the low AC load line doesn't lead to much of an undervolt during low-load scenarios, when no VDroop is happening.
The win goes to B or C because of the lowest average VCore.
The Last of Us Part 1
In The Last of Us, this time config A, the 110/110 + offset configuration, had the highest average Vcore. Config D/Lite Load 5 still has the second highest average Vcore, and perhaps this game's CPU load is a middle ground where the VDroop is high enough for config D to have lower average Vcore than config A, but not high enough so that the lack of a V/F offset is compensated enough to match config B and D.
The win goes to B or C because of the lowest average VCore.
Conclusions:
Can we undervolt with CEP enabled - definitely! It is certainly more complicated and finicky compared to simply reducing AC and disabling CEP, as there are now multiple parameters to account for - AC, DC, LLC, and offset. But the results can be very good, performance is almost identical compared to Lite Load 5, and the voltage is lower in gaming and light usage.
In Cinebench R23, LL5/config D technically performs the best, no doubt about it, but the performance difference is so negligible it can never be felt. However, LL5 had a significant advantage in the OCCT stability test. Lower VCore, lower power draw, lower temperature, it was a clear winner there. This brings me to a conclusion I never though might be the case - perhaps, there is no best undervolt method (even complexity aside). Some will give you lower voltage in gaming and light usage, others will excel in specific workloads that tax the CPU a certain way. At least this is how I interpret my results, which I admit, are not based on an extensive suite of benchmarks and tests. I could go back and do additional tests with the same configurations, probably first on my list would be a 10-minute R23 run and a 10 minute R24 run with each, but this would take me a lot of time.
Anyway, another thing I think is visible is that basically all four configurations are very capable, and I'm quite happy with the results overall. Cofigurations B and C are the most interesting to me because they combine a reduced AC load line with an offset, and mix the best of both worlds. I think they're great for most people, as they provide good performance and temperatures, and lower the overall max VCore. But the very big difference between AC and DC/LLC that's present with LLC5 seems to be the best choice for optimizing power draw and temperatures, for anybody whose use case is heavy CPU loads such as OCCT, which create heavy Vdroop scenarios. For some reason the same doesn't apply to R23, so if somebody has an idea what's causing this different behaviour, please share.
Hope you enjoyed the read!
Last edited:
