Guide: "Almost" everything you need to know about Raptor Lake Voltage/Power/Temperature Tuning

FlyingScot

Well-known member
Joined
Apr 30, 2024
Messages
1,378
This guide assumes that you already have the basic knowledge necessary to enter the MSI BIOS and navigate the various settings screens. It also assumes that you have already upgraded (or plan to upgrade) to one of the more recent BIOS versions (August 2024 or more recent) that includes microcode 0x125, 0x129 or 0x12B. If you need more background information on these microcode releases, I have summarized them in this post.

I’ve tried to arrange the information in this guide in a way that takes the reader down this rabbit hole we call “undervolting” at a pace that matches their curiosity. At any point along the way, you are free to jump off to the much simpler beginners guide created by CiTay. If you do decide to use CiTay’s guide then I can recommend this example of a before and after case study.

However, if you decide to keep reading then think of the information I present in this guide as both a prelude to Citay’s guide and a “next-steps” knowledge base for those who want (or are curious) to take undervolting to the next level. At a minimum, it might serve as useful reference material once the reader has gained some prior undervolting experience.

[LEVEL ONE KNOWLEDGE]

I've been working on a standardized approach to Raptor Lake voltage tuning. But before I get to that, perhaps we should ask the question "Why do we undervolt?" In the old days, all we talked about was "overclocking." But these days, undervolting is often much more desirable. After all, our PCs are already noisy and hot enough as it is, don't you think?

Reasons to undervolt might include one or more of the following:
[1] Fix a PC with excessive fan noise, especially while running stability tests, benchmarks, heavily multi-threaded applications and playing games.
[2] Increase the longevity of PC components by reducing the kind of degradation that occurs when components get overly hot.
[3] Squeeze every last bit of CPU performance from your PC within the limitations of your existing cooling system.
[4] Less common, but maybe to stabilize an unstable system.

How does silicon quality, referred to as the "silicon lottery," affect your undervolting and overclocking results:

[1] While CPU stability will generally increase with higher voltage levels, a CPU's silicon quality will ultimately have a Core Frequency (Core Ratio) "wall." This means that you will not be able to increase your CPU's core frequency no matter how much voltage you feed it. In some cases, the quality of a single P-Core or E-Core can determine the maximum frequency a CPU can operate at and still remain stable.

[2] When undervolting is the objective, the same frequency-to-voltage relationship will apply. Your CPU's silicon quality will ultimately determine the minimum voltage that each core frequency (Core Ratio) can operate at and still remain stable.

[3] CPU temperature has a direct effect on the minimum voltage your CPU will require in order to remain stable at a given frequency. A CPU's temperature will typically increase in direct relationship to the power consumption (Watts). It is also influenced by the quality of the silicon (i.e. voltage leakage), the ambient temperature of your room, the CPU cooler performance and the case cooling performance.

A quick note: Don't forget that your case cooling performance (in addition to your CPU cooler) can have a direct effect on both CPU temperatures and overall system stability. In particular, cases with solid front panels plus front mounted fans are particularly problematic when it comes to component temperatures. These cases can easily create a "negative pressure" environment, basically where more air is exiting the case than entering the case. Symptoms of a negative pressure condition include unusually high motherboard chipset (e.g. Z690, Z790) temperatures (>65C) and RAM temperatures (>50C).

If you are relatively unaccustomed to this level of BIOS tuning, please don’t worry if the first time you try to follow these steps (as sequenced) you have to keep stepping back. It might take a little prior knowledge and personal experience before you can more easily make these decisions in the order presented. Hindsight is everything. However, if you have any suggestions on how I might improve this guide, which is still a work-in-progress, then please don’t hesitate to leave a comment in this thread. Cheers!

So without further ado, here are the steps that I think make the most sense based upon watching two dozen people struggle with this effort. I've tried to lay out a step-by-step process that (a) provides some important background information for beginners, and (b) side-steps any potential conflicts caused by doing things in the wrong order.

Step 1a: Decide on which CPU Cooler Tuning [preset] to use to configure important BIOS defaults.
Step 1b: Decide on whether you want to try to tune with Current Excursion Protection (CEP) Enabled [can add more complexity] or Disabled [usually results in less complexity].
Step 2a: Decide on which type of workloads are most important to you when it comes to your temperature and voltage tuning. Do you want the lowest possible voltage and temperatures for Prime95 (i.e. highly multi-threaded/all-core workloads) or Battlefield 5 (i.e. single threaded/less than all-core workloads)?
Step 2b: Decide on whether to retain the stock factory boosting behavior or whether you want to implement manual turbo offsets or manual max-core ratios. You can use this step to either manually “overclock” (to maximize performance) or “underclock“ (to reduce power usage/temperatures) depending upon your priorities.
Step 3: Select your CPU’s Power Limits (PL1, PL2, IccMax) to match your cooling solution and target temperatures.
Step 4a: [Optional/Advanced Tuners] Decide on your Loadline Calibration Control (LLC) setting.
Step 4b: Begin tuning CPU Lite Load [Novice Tuners] or AC_LL [Advanced Tuners].
Step 5: [Optional/Advanced Tuners] Add a Manual Voltage Offset.
Step 6: [Advanced Tuners] Tune DC_LL to fix VID and CPU Package Power inaccuracies.
Step 7: Save your final settings (after stability testing) to a BIOS OC Profile (1-6) and then repeat steps 1 - 6 making different choices. Compare your different configurations and pick a winner that works best for your normal PC usage.
Step 8: [Optional] Add a manual CPU VR Voltage Limit to your final choice of settings.

Background Information
Step 1a
: Selecting the best CPU Cooler Tuning preset requires a certain level of knowledge and experience, and in some ways is an emotional decision. For example, do you really trust Intel and MSI to pick the best defaults for your unique CPU and usage case? If you are new to tuning then the answer is probably ”Yes!” However, as you gain more tuning expertise, you will likely want to take more direct and preemptive control over the power, voltage and core frequency settings. Selecting a different CPU Cooler Tuning mode can sometimes simplify things. Luckily, however, regardless of the preset you choose (e.g. Intel Defaults, MSI Performance, MSI Unlimited, etc.), you can usually achieve the same end results by using a slightly different combination of defaults and user-overrides - although Intel may make this more difficult with future CPU generations.

Tip: The CPU Cooler Tuning preset you choose can, and will, alter the defaults for the following settings (still a work-in-progress list):
Long Duration Power limit(W) >> [PL1]
Short Duration Power limit(W) >> [PL2]
CPU Current Limit(A) >> [IccMax]
IA CEP Support
IA CEP Support For 14th Gen
TVB Ratio Clipping >> [All i9 CPUs]
TVB Ratio Clipping Enhanced >> [14th gen i9 CPUs ONLY]
CPU Under Voltage Protection >> [Intel whitepaper]
LoadLine Calibration Control (LLC) >> [Auto** value can change behind the scenes]
CPU Lite Load Control
CPU Lite Load [setting] >> [affects the hidden values AC_LL** and DC_LL**]

** "MSI Unlimited" preset appears to select an Auto LLC value of maybe 5, 6, or 7 and then sets AC_LL and DC_LL to “hopefully” match this LLC level, e.g. AC=50/DC=80.
** "MSI Performance" preset ....???
** "Intel Defaults" preset has consistently selected Auto LLC value of 8 and set AC_LL and DC_LL to “hopefully” match this LLC level, normally AC=110/DC=110 (but sometimes 170/170).
Tip: Don’t worry if the above three ** points make no sense to you. It should make more sense after you finish reading my Level Two Knowledge post in this thread.

Tip: Less experienced Raptor Lake tuners should consider starting out with the Intel Defaults preset because it mirrors most of Intel's recommended guidance, including reasonable power limits for your CPU. However, note that the Intel Defaults preset will likely need some serious manual tuning to bring down potentially high voltage and power usage conditions. That's because it has been configured for maximize stability rather than efficiency. More experienced tuners should consider using one of the MSI presets.

Tip: The latest BIOS releases from MSI will likely enable C1E Support by default. This setting was previously set to disabled. In a related point about C-States, be aware that, in some cases, the Package C State Limit setting, especially in the highest sleep states, can cause instability when the CPU is at idle, presumably because voltage drops too low after your undervolting efforts. If you run into idle instability (but all other activities are stable) then either try changing the above C-State settings to be less impactful or simply back off a little on your undervolt. Keep this point in mind for future reference.

Step 1b: The benefits of Current Excursion Protection (“IA CEP Support”) is an ongoing source of debate. The decision to enable or disable CEP is partly an emotional decision in the absence of more convincing data one way or the other. In general, tuning with CEP is often more difficult than tuning without it. It is therefore an important decision that should be made at the earliest stage in the process. As you continue learning more about the role of CEP (keep reading this thread and the recommended links) you may come to better understand both its importance and how to tune with it enabled.

Tip: Be aware that when selecting CPU Cooler Tuning preset “Intel Defaults“, CEP will normally be enabled by default. But when selecting one of the two MSI presets, CEP might be disabled by default.

Step 2a: Do you want to tune your CPU for highly multi-threaded workloads (such as Video processing, Unzipping files, etc.) or non-all-core workloads - like most game engines. This decision can end up driving you crazy if you are trying to maximize benchmark scores and have good gaming temps. For example, Cinebench R23 loves lots of voltage and sky-high power limits (PL1, PL2 and IccMax) - assuming your cooling solution can keep up. On the other hand, games, web browsing, Microsoft Office tasks, etc. are generally low-power activities and therefore allow for much lower voltage (and power limits) to maximum efficiency and run with the coolest temperatures. In the end, you will undoubtedly have to prioritize one over the other. If bragging rights or highly multi-threaded applications are important to you then consider a separate BIOS OC Profile for just those occasions....then have another OC Profile for all other activities, including gaming.

Step 2b: This is an important decision. Do you want Intel to control all the core boosting behavior in order to maximize light load performance (especially single and dual core workloads) but at the risk of more elevated CPU voltages? Or do you want to implement your own individual CPU core frequencies (including the old fashioned All-core frequency approach), which can help to reduce voltages at the expense of slightly reduced light-load performance, e.g. some gaming workloads? Keep in mind that the highest Voltage request (VID) from all the individual IA Cores (both P-Cores and E-Cores), plus the Ring VID, will always decide the voltage the CPU requests from the motherboard’s VRM power module. Due to the Voltage/Frequency (V/F) curves (one for each IA Core) voltage requests (i.e. VIDs) are directly linked to core frequency. If you ask a particular CPU IA Core to hit a higher frequency, it will often require more voltage to remain stable. The same is true if you increase the Ring Ratio. The amount of additional voltage will ultimately depend upon the quality of the silicon in that part of the CPU Die and the temperature you run it at. Conversely, when you lower the peak frequency of a CPU Core (by using a manual core ratio limit or negative turbo offset) you will usually see a lower peak VID request from that particular IA Core.

See the BIOS configuration example below where one of the manual core ratio options is used to lower the maximum P-Core Ratio to 56 and the maximum E-Core Ratio to 41. Notice that in the factory default configuration, P-Core 4 and 5 are considered the “preferred” cores (i.e. better silicon/less voltage needed) and therefore allowed to boost higher than the other cores.

Capture.JPG


Tip: When implementing manual core ratios, you are forced to use the ”MSI Unlimited” CPU Cooler Tuning preset. This will likely not bother you if you were already using the other MSI preset called “MSI Performance”. However, if you were using the “Intel Defaults” preset, and this is the preset mode you prefer, then you can try to switch back to the “Intel Defaults” preset after making the core ratio change (or any other change that kicks you out of this preset). Some users report that this trick works. But, either way, make sure to revisit Step 1a/1b before moving on to Step 3.

Tip: Of course, you do have a third choice. You could allow the factory boosting behavior to remain in place (see example below) but control the behavior with Turbo offsets, AVX offsets**, Ring offsets, or by disabling Intel Turbo 3.0 that controls single core and dual core boosting for the preferred cores, which is often responsible for the highest peak CPU voltages. Also, while it’s not generally recommended to disable Thermal Velocity Boost (TVB) [i9’s only] and Enhanced Thermal Velocity Boost (eTVB) [14th gen i9’s only], you can in fact alter the factory TVB behavior with manual settings, referred to as OCTVB. In each one of these cases, you can control CPU voltages by manipulating the maximum frequency a core will automatically boost to via the factory algorithms. Again, higher frequency of any one IA Core usually means higher voltage for the entire CPU. With Raptor Lake’s history for degradation, lower voltage is always safer than higher voltage. [This guide may cover Turbo offsets and OCTVB configurations in the future.]

Factory Boot 14900K.JPG

The above table shows the standard factory frequency boosting technologies for the eight (8) P-Cores inside the 14900K SKU. The bottom axis shows the number of active P-Cores and the side axis shows the core frequency.

Key:
Base is the base frequency if you were to disable Intel Turbo 2.0.
TB2 (Turbo Boost 2.0) is limited by the maximum core frequency (per Intel spec. sheet), the main power limits: PL1, PL2, and TAU (how long PL2 can remain active), plus the CPU Package temperature (TJMax), which usually defaults to 100C.
ABT (Adaptive Boost Technology) was first introduced with Rocket Lake. It is not user-configurable for Raptor Lake.
TB3 (Turbo Boost 3.0) was introduced with Alder Lake and is exclusive to i7 and i9 CPUs. It identifies two “preferred“ cores to boost higher under certain conditions. It’s primary focus is to benefit single-threaded applications/events. While not always the case, TB3 can result in the highest voltages seen for your CPU. You can disable this feature in the BIOS.
TVB (Thermal Velocity Boost) was introduced with Comet Lake and is exclusive to i9 CPUs (regardless of whether it shows up in the BIOS). It can be user-configured via a separate OCTVB BIOS submenu. The default behavior is to opportunistically boost core frequencies whenever CPU temperatures are maintained below ~70C.
eTVB (Enhanced Thermal Velocity Boost) works in conjunction with TVB to boost Core Ratio by +1.

** 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.

Step 3: We all start from the perspective that we want our CPUs to run unencumbered. However, CPU temperatures have never been such an issue as they are with Raptor Lake. It’s therefore important to be rather conservative with your CPU’s power limits in order to stay within the capabilities of your particular cooling solution. The goal should be to keep temperatures below 95C for true stress tests (e.g. Prime95, OCCT, y-cruncher, etc.), below 85C for simulation and synthetic benchmarks (e.g. Cinebench R23, Geekbnech, etc.), and below 75C for everything else. Ideally, your gaming temps should average in the 60’s, only spiking into the 70’s during heavy shader compiling that makes heavy usage of all the CPU’s cores and threads. Keeping those high temperatures down can help to reduce the normal rate of silicon degradation and therefore extend the life of your CPU and surrounding components.

Power Limits.JPG

Long Duration Power limit(W) >> [PL1] >> See note for PL2.
Long Duration Maintained(s) >> [TAU] >> The time (in seconds) that PL2 power limits will remain in control until PL1 takes over. It's common practice to leave this setting at the default.
Short Duration Power limit(W) >> [PL2] >> While measured in what Intel describes as TDP, you can simply think of this setting as limiting the peak Watts that can be consumed by the CPU.
CPU Current Limit(A) >> [IccMax] >> This setting attempts to limit the "peak" current that can flow through the CPU.

Tip: For help with selecting safe and reasonable limits for your CPU, see the information provided in the following post, including Intel’s latest Raptor Lake guidance. Note that when you select the CPU Cooler preset ”Intel Defaults“, the BIOS should already default to sensible power limits based upon Intel’s guidance.

Tip: These days, the easiest and most common approach to establishing your PL1 and PL2 power limits is to use Cinebench R23. You can often use Cinebench R23 to identify the maximum capabilities of your cooling solution and then use that understanding to set your PL2 value. You would then set your PL1 value equal to or less than PL2 depending upon your risk tolerance. Lower temperatures for long running workloads mean less chance for CPU silicon degradation. To dial in the PL1 value, you should experiment with your normal long-running workload tasks. For example, you could set a lower PL1 value to ensure that your gaming temperatures remain in the 60’s. Many of today’s games only require 150W or less, so this is a good place to start when it comes to setting a lower PL1 than PL2. Incidentally, when it comes to games, long shader compiles can really hammer your CPU. A lower PL1 value versus the PL2 value can help to keep a lid on temperatures during these activities. The same is true for other long running “spiky” workloads, such as video rendering, uncompressing files, etc. For short bursts in workload demand, PL2 will be in effect. The rest of the time, PL1 is in effect. In regards to setting a Current limit (IccMax), see Intel’s recommendations in their chart. For a practical guide on how to establish power limits within the MSI BIOS, refer to step 1 in CiTay's Guide.

Tip: I deliberately placed the tuning of the Power Limits early in the sequence of events for one important reason. If you dial in your undervolting settings by following the Steps 4 - 6 and then change the power limits afterwards, you can run into stability issues. So just be aware that if you increase power limits at a later date, you may have to revisit your CPU Lite Load and/or negative offset voltages to re-stabilize the system, i.e. back them off a little. Keep this in mind.

Step 4a: People often have strong feelings about which Loadline Calibration Control (LLC) setting to use. For less experienced tuners, I recommend just starting out with the Auto setting. The biggest differences come from your choice to use a droopy LLC versus a flatter LLC. It will often depend on how you use your PC. If you are doing a lot of highly multi-threaded activities, such as video rendering, then a more droopy loadline (LLC 6,7,8) is often best. If you predominately play games (typically non-all-core workloads) then a less droopy loadline (LLC 3,4,5,(6=overlap)) combined with a voltage offset via adaptive+offset or V/F points might be best. This is because, with gaming type workloads, PL1, PL2, IccMax and AC_LL are far less active at reining in higher voltages. However, if you can lower the AC_LL value even when using a droopy Loadline and use a negative offset then that's the best of both worlds.

Step 4b: Okay, this is now the least difficult step thanks to step 2 of CiTay's Guide. If you have chosen to keep CEP enabled then you will also want to read Vassil_V's case study guide and his follow-up comment. As you follow these various guides that focus on reducing CPU voltage, you may also find that your CPU can now boost higher in frequency (and therefore performance) while still staying under the power limits that you established in Step 3. If, after lowering your voltage, you find your CPU is running too cool (i.e. your entire PC system has more thermal headroom) then you always have the option to return to Step 3 and slightly increase the power limits you previously chose.

Tip: Before you attempt to read the above guides, you may benefit from reading my next post in this thread [LEVEL 2 knowledge] for background information on the various settings of Intel's voltage control system, which is referred to as the Intel Loadline mechanism. However, it's quite likely that you'll need to revisit this information periodically. In other words, don't try to understand all of it in one reading.

Step 5: A manual voltage offset can have the biggest impact on voltages for non all-core workloads and normal activities, like gaming. The most popular voltage offset is the Adaptive+Offset option. Note that while manual offsets are relatively easy to implement, they usually require extensive, and potentially time-consuming, stability testing. You really won't know for several months whether the CPU voltage and frequency related settings you chose in conjunction with the manual voltage offset are truly stable. Still, the rewards can often outweigh the extra effort of testing and tweaking.

Adaptive+Offset.JPG

In the above example, the user has selected to implement a negative(-) 110mV offset to the automatic "factory voltage" by using the Adaptive + Offset mode. Be aware that this voltage reduction will apply to all frequencies and all workloads, including your idle voltage. Due to this behavior, what can sometimes happen is that you find your system is perfectly stable under heavy load, but unstable under lighter loads or even at idle. If this situation occurs then you have no choice but to reduce the amount of negative offset even if it means higher voltages under full load. The amount of voltage offset your CPU can handle typically ranges from -10mV to - 150mV, and will ultimately depend upon many variables, including:
>Silicon Quality and luck
>Degradation
>The settings you choose for the Intel Loadline mechanism
>The frequencies and workloads your CPU experiences
>Cooling system capabilities

Tip: If your BIOS does not have the (-) option for CPU Core Voltage Offset Mode then use "- (By CPU)".

Tip: In addition to the usual BSOD, system freeze, application crash, or unexplainable mouse hesitation/stutter (all common signs of CPU voltage issues - usually because it dropped too low) also keep an eye out for WHEA errors in the Windows Event Log (also trackable in HWInfo64). These errors can occur at any time, likely indicating that your CPU voltage settings are not stable under all conditions. In most cases, slightly increasing CPU voltage should fix the problem. But you can also try lowering the power limits if the instability occurs during all-core workloads. For more information, refer to Level 5 Knowledge.

Custom View NameSourceEvent#Location
WHEA ErrorsWHEA-Logger19Windows Log \ System

Step 6: This task should be a piece of cake, but instead is made unnecessarily difficult because MSI will neither do it for us (like ASUS), nor will they provide the information we need (i.e. the LLC impedance for each setting 1-8, and Auto) to make it easier to do it for ourselves. Luckily, this is not a mandatory step. Your system will continue to function whether you complete this step or not. A separate guide on how to tune your DC_LL is planned for the future, but in the absence of this guide I would simply advise setting the DC_LL a little higher than the AC_LL when adjusting the AC_LL value.

Step 7: Trying different undervolting and power tuning approaches requires patience and a considerable time commitment. It will also require another emotional decision when it comes to picking the winning configuration.

To recap: We know that undervolting via lowering the Lite Load setting (in Normal mode) or AC_LL (in Advanced mode) greatly benefits voltage, power usage and temperatures for all-core workloads. We also know that undervolting via a negative voltage offset (e.g. Adaptive+Offset) is effective at lowering voltage, power usage and temperatures for all workload types. However, this undervolting method is not usually as effective as the former method when it comes to taming temperatures, etc. for all-core workloads. In addition, proving a large Adaptive Offset as stable can be more challenging than just lowering the AC_LL value because, unlike the Adaptive Offset, the effect that AC_LL has on voltages is inversely proportional to the current (amps), with idle being the least affected. In the real world, this means that Cinebench and Prime95 multi-core tests are greatly affected, while games and MS Office are far less affected. Adaptive Offset is the opposite. It has a statistically greater impact on low current applications than high current applications.

Given our understanding of the behavior of the two most popular methods of undervolting is the reason why it can be a good idea to implement both methods at the same time. I find a good approach is to first find your lowest undervolt via lowering AC_LL then see how much Adaptive Offset you can add on top. Any approach, including this one, requires making decisions about whether you want to keep CEP enabled, and which LLC level you want to start out with - Auto is fine. You can then experiment with different approaches. Sometimes a flatter LLC (e.g. 3 - 6) with a lower AC_LL is more optimal than a droopier LLC (e.g. 7 - 8) and a higher AC_LL. As a general rule, you can usually stabilize a lower AC_LL (with or without CEP enabled) when selecting a flatter LLC setting than a droopier one.

Tip: I always advocate for having a saved OC Profile for when bragging rights are important (e.g. Cinebench R23 high-scores), but having a game-tuned profile for daily use.

Tip: Access our Raptor Lake Survey Settings “database” to see what others have chosen for their settings. And for additional research material, keep reading this thread. Plus, don’t forget to submit your settings to the database when you eventually find that happy equilibrium between temperatures and performance.

Step 8: A CPU VR Voltage Limit is a new MSI BIOS setting that was implemented when the 0x12B microcode was released. If you do not enter a manual voltage limit, the default behind the scenes is set to 1.55V. If you are concerned about high voltages, and you notice that your CPU usually stays below a certain peak voltage (both VID and Vcore), then you are now able to implement a manual voltage limit below the 1.55V default. I think a good general rule of thumb is to add 50mV to the highest peak voltage you have seen during extended monitoring with HWInfo64 (e.g. 1.40V => 1.45V) or use 1.35V. Pick whichever is the highest for your system based upon your monitoring of different activities. If you think you have lost performance in some activities, or can measure it via benchmarks, then add another 25mV. You can see how we went about establishing a VR limit for one forum member in this conversation found here.

In my opinion, Raptor Lake voltages (as observed by HWInfo64 "Vcore") can be categorized as follows:
Voltage peaks above 1.50V = potentially bad for the life of the CPU
Voltage peaks between 1.45V - 1.49V = possibly okay, but still in the historically uncomfortable zone
Voltage peaks between 1.40V - 1.44V = likely okay, but not ideal
Voltage peaks between 1.35V - 1.39V = should be safe
Voltage peaks below 1.35V = very safe (historically) and easier on your cooling solution

Note: Note that the CPU VR Voltage limit will not stop voltage jumps due to momentary transient spikes when the CPU rapidly transitions from high workloads (especially all-core workloads) to very low workloads or idle. When measured on an oscilloscope, these voltage spikes are typically in the neighborhood of +50mV on top of peak voltages seen in HWInfo64. However, transient spikes have been a thing since the early days of CPU development and should not, in theory, meaningfully shorten the life of your CPU - especially if you undervolt your CPU and/or implement a CPU VR Voltage Limit well below the default 1.55V.
 
Last edited:
[LEVEL TWO KNOWLEDGE]

The neatest explanation I have come across regarding Intel’s Loadline mechanism (which includes the three BIOS settings AC_LL, DC_LL and CPU Loadline Calibration Control (“LLC“ for short)) is courtesy of BuildZoid’s latest set of videos (as of August 2024) on Raptor Lake tuning, and one video that focuses on the role of the LLC setting.

Firstly, a little background for us non electrical engineers out there. Have you ever plugged in an iron, or space heater, or vacuum, etc. and seen the lights in your home dim for a second? Well, that’s Vdroop - a momentary drop in voltage due to a sudden increase in current. A CPU draws current based upon how many CPU cores are working, and how hard they are working. It also increases with core frequency. More cores running at higher frequency = more current = more potential for more Vdroop. Too much Vdroop and a CPU will crash. CPUs will basically crash whenever voltage is too low for the task being worked on, whether that’s a high-frequency single core workload, a full-core workload, or while at idle.

Other than selecting a CPU with fewer cores (referred to as “IA Cores” in the Intel universe), or disabling cores or hyper-threading, the most effective way to reduce power usage (and temperature) of a CPU is to reduce the voltage. The reason why voltage reduction is far more effective than, say, reducing frequency alone is because for all intents and purposes amperage will drop in direct relationship to a drop in voltage. We all know that Power (measured in Watts) is the result of Volts x Amps, so the benefit of a drop in Voltage of 10% is magnified by a corresponding drop in amperage, too. For example, 4 x 4 = 16, 3 x 3 = 9, 2 x 2 = 4. On the other hand, dropping the max frequency from say 5.7Ghz to 5.5Ghz (assuming voltage is relatively unaffected) will have an effect on peak power usage, but percentage wise is often far less impactful than undervolting.

When it comes to taming your hot running CPU, voltage is the KING! The real trick, however, is trying to find the lowest voltage your CPU can happily operate at without making computational errors that lead to WHEA errors (corrected errors), stutters, freezes and crashes.

Intel CPUs use a formula that includes several variables to calculate a target minimum voltage (Vmin) that it thinks will be required to remain stable under the predicted workload conditions that the Windows OS has notified the CPU is coming. The user has three ways to manipulate (or even totally override) this target voltage: [1] by altering the relationship between the BIOS settings called Loadline Calibration Control (LLC) and AC_LL, [2] by using manual voltage offsets, or [3] by using a fixed voltage. The first two options can be used in conjunction with each other, while a fixed voltage overrides all other voltage mechanisms. We shall now take a closer look at the the various elements that make up the first option, referred to as the Intel Loadline mechanism.

———————

The Intel Loadline mechanism consists of three BIOS settings/values: the Loadline Calibration Control (LLC) setting, the AC_LL value, and the DC_LL value.

The CPU Lite Load setting makes it easy to alter the AC_LL and DC_LL values, and is in turn controlled by the CPU Lite Load Control setting, which has three modes: Normal, Intel Default (new), and Advanced.

CPU Lite Load - Normal Mode.JPG


If Normal mode is selected then the motherboard BIOS will choose default values based upon your type of CPU. These default values can also vary depending on the motherboard model/generation. Input options in this mode include a CPU Lite Load setting of "Auto" and manual presets 1 - 20+. In many cases (and especially with the recent BIOS updates) the values chosen for AC_LL and DC_LL are “aggressive” meaning they err on the side of boosting the voltage even more than what Intel specified for the CPU when it was tested at the factory, referred to as the unique V/F table or VID table. The VID table (one per Core and one for the Ring) is permanently stored in the CPU to establish the voltage target for each frequency bump. For example, a CPU Core running at 4.8GHz will need more voltage to be stable than if it was running at 4.5GHz.

By the way, when Intel SpeedStep Technology is enabled in the BIOS (in most cases this will be the Auto setting) the CPU will automatically alter its frequency up or down depending on workload. At idle, the Intel SpeedStep will also determine the idle voltage for the idle frequency (usually 800 - 1000MHz).

The additional voltage supplied to the CPU by the CPU Lite Load mechanism (across the entire range of frequencies) is intended to guarantee that the CPU will be stable out of the box when placed into the motherboard. Note that more voltage usually means more stability, but that only works up to a point. Too much voltage can damage the CPU, either over extended periods of time (months/years) or, in extreme cases, instantaneously. Too much voltage in excess of what is needed to keep the CPU stable also generates unnecessary heat, and can have the effect of slowing down the performance when the power consumption gets too high. But let’s get back to the discussion of the different CPU Lite Load modes. In the Normal mode, a single CPU Lite Load setting will be made available to the user (mode 1 - 20+) that will have preset values for AC_LL and DC_LL. In many, but not all cases, the two values will be the same.

CPU Lite Load - Advanced Mode.JPG


If Advanced mode is selected then the individual AC_LL and DC_LL values will be exposed to the user so they can be changed independently by the user. WARNING: Caution is required to not accidentally enter 600 instead of 60, as damage to the CPU is possible. Proceed with caution.

Tip: If you are switching back and forth between CPU Lite Load mode “Normal” and “Advanced” then be aware of this potential issue, especially if you are manually entering AC_LL/DC_LL BIOS values in the Advanced mode.

CPU Lite Load - Intel Default.JPG


If Intel Default mode is selected (a new mode that arrived mid-2024 via a BIOS update to address Raptor Lake CPU instabilities) then the AC_LL and DC_LL will very likely be set to 110/110 for most desktop CPUs, but sometimes these values are set as high as 170/170, which is a great way to “cook” your CPU. This mode may also have an impact on what the Auto means for another BIOS setting called Loadline Calibration Control (“LLC“ for short). In regards to Raptor Lake CPUs, this "Auto" setting for the LLC will likely be equivalent to a user selected mode 8, from the range 1 - 8.

Bonus Info.: Intel and motherboard vendors appear to be standardizing on configuring the out-of-the-box BIOS settings to anticipate the motherboard’s voltage delivery system (VRM) as having a resistance (impedance) of 1.100 mOhms. The latest Intel Default mode assumes this is the case when it sets the “Impedance“ for AC_LL = 110, DC_LL = 110 and the LLC = 110 (MSI Mode 8).

———————

The Loadline Calibration Control (“LLC“ for short)

LLC.JPG


The Loadline Calibration Control (“LLC“ for short) is a multiplier that alters the amount the motherboard's Voltage Regulator Module (VRM) compensates for Vdroop between the VRM components and the CPU Die when you have a lot of electrical current flowing. Motherboards with differing VRM circuitry will have different LLC multipliers. The more CPU cores that are active, the more Current (Amps) there is. LLC Mode 1 = 100+% compensation and LLC Mode 8 = Little to no compensation. The LLC Auto setting will approximate one of these manual modes. The value assigned to the Auto default can depend upon which CPU Cooler Tuning preset you have selected. There should be a graph in your BIOS that gives you an approximation of what each value does in terms of the voltage drop, with a flat line representing 100% compensation. Note: The CPU does not know what the LLC setting is set to and therefore cannot anticipate for changes that you make to the LLC setting.

Tip: Lower compensation = Lower voltage arriving at the CPU = Lower Vcore/VR VOUT reading.

Below is an example of how one user tried to find (read "guess") the impedance values associated with each rung on the LLC ladder. This information is important as it makes tuning the AC_LL and DC_LL values much easier, especially if you plan to keep IA CEP enabled and happy! More on that later...

LLC.JPG

Table A - LLC Impedance relationship for MSI Z690 Edge motherboard

Note: We tend to say 0.15 mOhms or 15 interchangeably. So when we say AC_LL = 15, what we really mean is 0.15 mOhms. AC_LL = 1 means 0.01 mOhms.

Tip: To find the default LLC value when set to Auto, switch to mode 5 and see if the steady-state Vcore is higher or lower when running Prime95 Small FFTs. If Vcore is higher then you know that Auto must be equivalent to mode 6, 7 or 8. If Vcore is lower then you know that Auto must be equivalent to mode 4, 3, 2 or 1. Repeat the process with a different LLC mode until your steady-state Vcore closely matches the Auto setting. Note: If you experience crashes or other instability when using mode 5 (especially if it crashes while under load) then it's a good indicator that the default LLC Auto value is less than mode 5 (i.e. 4, 3, 2, 1). If it crashes when you stop the test, or wiggle the mouse, then that might be an indicator of the opposite. I plan to go into more detail on this behaviour as part of Level Five Knowledge write up.

———————

The IA Domain Loadline

The next two BIOS settings that are related to CPU voltage are commonly referred to as AC_LL and DC_LL, and formally referred to as the IA Domain Loadline. To see what values are currently assigned to your system, see How to find your "hidden" AC / DC Loadline values.

———————

The AC_LL value represents an impedance value that tells the CPU how much Vdroop to expect (due to the loadline resistance between the VRM and the CPU Die) and therefore how much to adjust the VID voltage request upwards by to hit a specific target voltage under load. The higher the anticipated Current (Amps) will be under load the more effect the AC_LL setting has on the VID request. The AC_LL additive will be at its greatest when running heavy all-core workloads like Prime95 Small FFTs and Cinebench R23.

Tip: Lower AC_LL = Lower VID (request) = Lower Vcore (under load).

Tip
: In theory, the AC_LL (impedance) setting should always reflect the actual LLC impedance (as per your current LLC setting); but it doesn’t have to be perfect. In fact, using a lower AC_LL value (impedance) than the LLC impedance is a very popular way to undervolt your CPU to keep it cooler. See Example A.

Note: If CEP is enabled, current thinking is that AC_LL (impedance) cannot be set lower than 67% of the LLC impedance value, e.g. AC_LL min = 73 when LLC = 110 (Auto/Mode 8). If there is a mismatch then you may see your Cinebench R23 scores plummet. If that happens, you simply increase the AC_LL value. Be aware that if you make changes to the LLC setting (1 - 8) then you will have to make sure that you have not violated the above rule. If you have, then you will need to adjust the AC_LL to more closely match your new LLC setting.

Note: If for some reason you elect to use a fixed voltage for the CPU then the AC_LL will have no effect. Also, CEP must be disabled to prevent the kind of performance issues mentioned above.

——————

The DC_LL value tells the CPU how to "guesstimate" the actual power usage so it doesn't need to keep polling the motherboard's Voltage Regulator Module (VRM) for this information. The super-fast switching frequency of Intel's VID requests (as fast as 100 microseconds) might even make this impossible or too expensive to achieve, so the DC_LL is a constant that is used to adjust/calibrate the math. DC_LL is used to adjust the VID reported by applications like HWInfo64 so it can be more accurately compared to the actual CPU voltage while under load (as reflected by either VR VOUT* or the less accurate Vcore). It has no effect on real voltages, but it can have an impact on power consumption limits. The good news is that DC_LL does not have to be perfectly accurate when compared to the LLC impedance value (as per your current LLC setting). It just has to be close. Note: Only change the default DC_LL if you are also changing the LLC value - or you are fine tuning the the DC_LL after everything else is dialed in and stability verified.

Tip: Higher DC_LL = Lower “reported” VID. Think of it as a negative offset.

If the DC_LL value is too low (compared to the LLC impedance) then it's likely that your CPU Package Power (TDP) reading in HWInfo64, as well as the VID value, will read on the high side. In this situation, you may prematurely hit your PL1/PL2 limiters and therefore power throttle your CPU a little early. Conversely, if the DC_LL is too high (compared to the LLC impedance) then it's likely that your CPU Package Power (TDP) reading in HWInfo64, as well as the VID value, will read on the low side. It this situation, actual power usage may slightly overshoot your PL1/PL2 limiters. If you get the DC_LL value relatively close to the actual LLC impedance then power usage inaccuracies are usually less than 15 Watts, and hardly enough to cause any instability issues.

Tip: As a general recommendation, it’s a good practice to try to keep the “reported” VID and VCore/VR VOUT* within a +/-5% tolerance to avoid any major discrepancies with power consumption (as it relates to PL1/PL2 limits). However, this objective can be quite challenging to “dial in” and therefore deserves its own topic (see Level Five Knowledge). In the meantime, I would simply advise using the pros/cons table below to set the DC_LL value. It should be close enough not to cause you any problems.

——————

Example A: If we set the AC_LL value lower than what we believe the LLC impedance value is currently set to (using mode 1 - 8, and Auto) then the voltage requested by the CPU will be less than it needs to compensate for 100% of the actual drop in voltage (i.e. Vdroop) that will likely occur when the workload spikes the current. This is one way to "undervolt". For example, if we set AC_LL to 80 and we set LLC to 8 (which is usually intended to match an AC_LL of 110) then the result is that the voltage delivered to the CPU by the VRM will be lower than a traditionally matched setup of AC_LL=110 and, LLC=8 (110). This is because we have told the motherboard VRM to allow more voltage drop than we told the CPU to expect. The motherboard references the LLC setting to adjust its behavior while the CPU uses the AC_LL setting to adjust its behavior. If your CPU remains stable then congratulations you have just tricked the CPU into requesting less voltage, i.e. you have successfully undervolted your CPU via the Loadline settings. It will now run cooler as a result, and may even boost to higher core clock frequencies.

The table below attempts to summarize the pros and cons of popular Z670/Z790 Loadline settings chosen by users. Please note that while the values presented in the table will quite likely work for your particular motherboard model, slight variances in the LLC impedance relationship (versus Table A) are to be expected. This means that you may have to use trial-and-error to match the AC_LL and DC_LL values to the LLC when trying to keep CEP happy and when attempting to recalibrate the VID reported by HWInfo64 after altering the AC_LL from its default value. I hope to cover this topic in more detail in the Level Five Knowledge.

The table below is still very much a work-in-progress thing and my Pros/Cons are still being validated. As a result, take these table comments with a grain of salt as they could be gross generalizations.
LLCAC_LLDC_LLIA CEPCommentsProsCons
Auto / 8110110YesNew CPU Cooler Tuning "Intel Default" configuration for Z690/Z790

Loadline settings in balance with each other (per Table A)
Limited "Vdroop" with all-core workloads

Maximum stability

CEP = Happy!

Highest headroom for manual undervolting via Adaptive+Offset mode
High voltages

Maximum power usage (HWInfo64 "CPU Package Power")

Maximum heat
Auto / 880110YesA good undervolt for beginners when CEP is enabledMore "Vdroop" with all-core workloads will reduce both power usage and temperatures

Safe levels of stability (very likely)

CEP = Happy!
Temperatures may still be too high

Potentially less headroom for manual undervolting (via Adaptive+Offset mode) versus LLC6
Auto / 8<80110NoAs above but requires CEP to be disabledEven more "Vdroop" with all-core workloads further reduces both power usage and temperaturesThe least amount of headroom for manual undervolting via Adaptive+Offset mode

Stability must be verified

CEP cannot be enabled
65050YesEverything in balance with LLC (per Table A)Potentially more headroom for manual undervolting (via Adaptive+Offset mode) than a LLC8 AC_LL undervolt

CEP = Happy!
Stability must be verified

Higher transient spikes than LLC8
63050NoA significant user undervolt versus LLC8As aboveStability must be verified

Higher transient spikes
than LLC8

CEP will not be happy!
(AC_LL value less than 67% of LLC impedance)

Less headroom for manual undervolting (via Adaptive+Offset mode) than LLC6 above
51520YesEverything in balance with LLC (per Table A) + a slight undervoltPotential for best Cinebench R23 scores when power limits increased

May allow for maximum stable Single/Dual Core frequencies for i9 CPUs.

CEP = Happy!

Still leaves some headroom for manual undervolting via Adaptive+Offset mode
Not suited to all-core workloads (i.e. lower core frequencies likely due to higher power usage).

Stability must be verified

Higher transient spikes
than LLC6

Less headroom for manual undervolting (via Adaptive+Offset mode) versus LLC6 and LLC8
51020NoA slight undervolt to LLC5As aboveMay not be suited to all-core workloads (i.e. lower core frequencies likely due to higher power usage).

Stability must be verified

Higher transient spikes
than LLC6

CEP will not be happy!
(AC_LL value less than 67% of LLC impedance)

Less headroom for manual undervolting (via Adaptive+Offset mode) than LLC5 above

The below categorizations are from another source (although edited by me), and are worthy of consideration:

1.10mOhm/1.10mOhm (Advanced 110), LCC Mode 8
This should be called "Intel Failsafe" rather than "max." because it's meant for bad CPUs that are unstable at anything lower. In the real world, this setting is fine for low clocking CPUs like the 12400K because the max VID will only reach ~1.10v but it's not fine for an 8 p-core unlocked CPU because the max VID will hit ~1.6v.

0.80mOhm/0.80mOhm (80), LCC Mode 7
This should be "Intel default" because it should be stable for most CPUs and if your unlocked CPU can't run at this setting then you should RMA it. The VID will still be high so you'll want to try manual undervolting via a negative Adaptive+Offset in the MSI BIOS.

0.50mOhm/0.50mOhm (50), LLC Mode 6
May be the sweet spot if you use your PC for both work and play as it will lower VIDs when gaming (compared to LLC7, 8, Auto) and will have less potential voltage overshoot than LLC5 and below when running all-core applications like video rendering, i,e. “the jack of all trades but master of none. However, the VIDs may still be too high so you'll want to try undervolting via a negative Adaptive+Offset in the MSI BIOS.

0.20mOhm/0.20mOhm (20), LLC Mode 5
Might have higher voltage overshoot but the VIDs should be much lower so you might only need a minimal undervolt or none at all. Mode 5 may be very effective when manually overclocking your P-Cores and E-Cores as it will help to prevent excessive vdroop during high current (amps) workloads. Mode 5 may also play well with the higher voltage requirements of peak single/dual core frequencies produced by Intel’s various turbo algorithms, especially the i9 CPUs.

Tip: If you would like to see more real-world pros and cons of different Loadline configurations, I can think of no better place than the test results in Vassil_V’s case study, Different undervolting methods with IA CEP enabled, and how they compare to Lite Load as well as his recent follow-up comments.
You can also find all of his latest “favorite" BIOS settings in our Raptor Lake Survey Settings “database”, as well as see what others have chosen for their loadline settings.

<<Insert future link to alex-predator undervolting test results table>>>
——————

* All about VR VOUT
Some of the more premium MSI motherboards (such as the Carbon, Ace, GodLike, etc.) expose an additional CPU voltage sensor to the HWInfo64 application called VR VOUT. The reason VR VOUT is useful is that, under load, it is the voltage as reported by the CPU itself (or close enough) - meaning ”as it enters the CPU die“ - making it far more accurate than Vcore, which comes from the motherboard PCB somewhere near the CPU socket. [EDIT: I recently read that it comes from the VRM before any vdroop occurs. But I kinda doubt that for MSI motherboards. Could be wrong, though.] Luckily, in the MSI world, it’s mainly just a nice-to-have feature because MSI super I/O monitoring has an overwhelming propensity to overestimate Vcore, which is not the case with some other motherboard manufacturers (depending on model) where Vcore might be significantly underreported. That’s when VR VOUT becomes not just a nice-to-have, but indispensable - especially when we’re trying to keep “real” CPU voltage under a certain target like we are with Raptor Lake. Tip: VR VOUT is usually very inaccurate when the system is in a low power/idle state.

Tip
: I would advise anyone with VR VOUT to only report it in addition to Vcore, and never in place of it. Otherwise, people are going to be comparing apples and oranges.
 

Attachments

  • LLC.JPG
    LLC.JPG
    12.7 KB · Views: 59
Last edited:
[LEVEL THREE KNOWLEDGE]

In this brief write up, I attempt to answer three questions that people ask:
[1] Why do we need an Intel Loadline?
[2] What the heck does Current Excursion Protection (“IA CEP Support”) actually do?
[3] What are my thoughts regarding an optimum BIOS configuration?


Lucky for you guys, I just listened to BuildZoid's latest video (almost two hours worth) so you don't have to. While you're obviously free to watch BuildZoid's latest video, and develop your own takeaways, I thought I'd summarize some of my own for your consideration. Incidentally, if you have an MSI motherboard with a Raptor Lake CPU then this video is worth watching first.

The latest video is a deep dive into the topic of Intel's current power delivery architecture, both the VRM and the Loadline mechanism. I didn't expect to get much from this video, but instead I found it actually gave me the all important context that I had been searching for, i.e. why does undervolting Raptor Lake feel so different from older generations?

------------------NEW TEXT------------------
The primary reason for using a loadline in modern systems is to try to reduce the severity of voltage spikes (overshoot) when going from high to low output current (i.e. high processor workload to low processor workload) by allowing a certain amount of Vdroop (voltage drop due to electrical resistance in the components/wires/traces) to occur before trying to stabilize the voltage.

The loadline setting [i.e. MSI’s CPU Loadline Calibration Control], normally representing a mΩ (milliohms) value, determines how much the output voltage decreases when the current increases (CPU workload). This is derived from Ohm’s Law U = R*I. The drop in output voltage is calculated as loadline * Iout (output current). For example a loadline of 1 mΩ and output current of 100A, dU = 0.001 Ω * 100A = 0.100V. At 1.300V set-point output voltage, when loaded with 100A the output would really be 1.300 – 0.100 = 1.200V.

In general, the benefit of a “droopy” loadline (i.e. MSI mode 8, ASUS mode 1) is that it reduces the size of voltage spikes in terms of measured peak voltage. In theory, lowering voltage spikes can lead to an extended life expectancy for a CPU. However, the downside of a droopy loadline is that it can allow the voltage to drop too low, which can lead to one or more types of instability.

In general, the benefit of a “flatter” loadline (i.e. MSI mode 1, ASUS mode 8) is that it can be better at handling workloads with lots of transient behavior (e.g. games??), especially when manually overclocking Core Ratios to higher than Intel factory spec., and to some degree also help to maintain a stable voltage during heavy all-core workloads (i.e. high current (amps) workloads). However, as previously hinted at, the downside of a flatter loadline is those pesky voltage spikes when an all-core workload stops or is interrupted, perhaps when a user moves the mouse or when a Windows background process kicks off. These constant high-voltage peaks can in theory be damaging. Another downside of a flatter loadline is that most commercially available motherboard VRMs are not quick enough to moderate the voltage up and down as required to maintain the goal of a steady voltage. This limitation can result in its own level of instability.

Think of the situation when the rear of your car steps out and you find yourself in an oversteer situation. The correct action is to counter-steer to correct the slide. However, if you are not fast enough, you will cause an overreaction. Each time you are late in your corrective inputs, the “oscillation” gets bigger and bigger (often referred to as a “tank slapper”) until you finally lose total control and the car goes into a spin. Well, the same sort of thing can happen with the motherboard’s VRM loadline mechanism. When the VRM sees that a voltage spike is occurring, it interrupts the voltage output from the VRM. Then when it sees the voltage dropping too low, it resumes the voltage output. However, if the VRM is not quick enough in it’s efforts to maintain a consistent “flat” voltage, you get the potential for larger and larger oscillation, eventually leading to a crash or other instability.

The trick when tuning the Intel loadline is to find that happy median between not too droopy and not to flat. And in many cases, it will depend upon how you use your PC.
------------------END OF NEW TEXT--------------

I'll start out with a little factoid that BuildZoid shared late in the video. To demonstrate how outdated Loadline mechanisms really are (i.e. AC_LL + LLC settings), BuildZoid pointed out that Nvidia's GTX980 Ti (launched in mid-2015) actually had a better power management system than Intel's latest generation of CPUs. Basically, it was during the design of Nvidia's GTX900 series that the Loadline mechanism was retired in favor of a clock-stretching mechanism. The reason for this was that the Loadline mechanism was created way back when CPU's and GPU's had a current draw of around 100 amps, and not the 300 - 400 amps of today's CPUs and GPUs. Note: The Loadline approach was originally invented to overcome the issue that capacitors were just way too slow to react to the rapidly changing power demands of modern CPUs.

So, yes...the whole issue we appear to be struggling with when it comes to Raptor Lake, especially the newer i7's and i9's, is, in a word, CURRENT. When it comes to stability, current is everything. It has a direct and instantaneous affect on voltage. And it's reliable voltage that a CPU needs to remain stable.

The problem with the antiquated Loadline mechanism is that it cannot react to rapid changes in current, especially when those changes exceed what was originally anticipated. We all know by now that the way the loadline works is to "assist" the CPU in adjusting the voltage upwards in preparation for current induced Vdroop. This was fine in the old days, when a CPU would have a maximum current range of 0 - 100 amps, but today's CPU's have an operating range of 0 - 300+ amps. The only way a Loadline mechanism can be effective is to always assume the worst case scenario for current draw, which again in our case means 300+ amps.

At this point, it's important to interject the next most significant revelation shared by BuildZoid: Basically, All all-core workloads must be treated as if they could turn into the worst type of workload for generating maximum Vdroop at any moment. As BuildZoid puts it "The voltage compensation of the Loadline mechanism must assume that, at any time, the specific workload that initially woke up all the cores could morph into a Prime95 Small FFTs vdroop beast. As a result of this inability to predict the future, the voltage must be increased to such a degree that even the worse kind of Vdroop, caused by the worse kind of current, can still be handled by the CPU." This is why we see such seemingly disproportional effects of AC_LL on peak voltages. And why it's so important to lower the stock/default AC_LL value.

To illustrate what BuildZoid just said, do this: Check the Peak VID for Cinebench R23 and compare it to the Peak VID for Prime95 Small FFTs. You should find that they match. And not just with this generation of Intel CPUs, either. Now take a look at what happens to Vcore and Peak Power. They don't match...because the all-core workloads are very different.

To hammer this point home, consider this: Any time your interactions with your PC cause all the cores to activate, regardless of workload, you are going to see the Loadline prepare the CPU for the maximum Vdroop. Again, it does this by pushing VID requests sky high! And this is why lowering CPU Lite Load is so effective at bring down both voltages and temps. CiTay was right all along!

So, to recap. The rapid increase in current demands of modern high core-count CPUs has exposed a fundamental flaw in the Loadline approach. It has basically outlived its usefulness and should be retired by Intel asap. It should go the way of the dinosaurs, just like Raptor Lake! Let's see what Intel does with Arrow Lake and beyond. My guess is that Loadline tuning will either become unnecessary, or will no longer be offered to the end user. CEP, and its clock-stretching ability, will effectively phase out the loadline.

Incidentally, BuildZoid suspects that it took a long time for Intel to realize that high VID requests were killing CPUs because the motherboard vendors had actually done us a favor when they deliberately chose to undervolted via the CPU Lite Load mechanism (as the default out-of-the-box configuration). This had the effect of lowering AC_LL, and therefore the VID requests.

[By the way, BuildZoid does not think that the other side of that coin, i.e. motherboard vendors having unlimited power limits by default, was as much a problem as the overvolting caused by high AC_LL values. This conclusion is based upon the observation that the rapidly degrading CPUs in Minecraft servers had very conservative power limits, but would have likely implemented the equivalent of the 0x125 Intel defaults from day one. Again, resulting in high AC_LL values and high VIDs. A very interesting hypothesis.]

However, when Intel rolled out the 0x125 microcode to "fix" the instability of i9's (that could not run stable out of the box) they inadvertently exposed the issue of high VID requests. Now that the motherboard vendors had adopted the new Intel Defaults, there was nothing to artificially hold down the excessive VID requests caused by Intel’s Loadline approach. Intel had fixed one problem but possibly caused an even bigger one. It's why CiTay had to write his second guide, the one addressing high temps caused by the new microcode. And perhaps this is why Intel rushed to release a voltage limit of 1.55V with their next microcode rollout (0x129).

BuildZoid has a very tantalizing theory there, and one that he thinks is deeply troubling. I quote (again, I'm paraphrasing), "I can't believe it took Intel so long to figure this out and implement a VID request limit. The issue of high current and the flaw in the Loadline design had to have been known from the very beginning."

I’m mindful of the fact that in this phase of the Raptor Lake saga, we’ll want to keep things as actionable as possible. So what does all of this mean to the RPL initiated?

My Conclusions based upon BuildZoid Recommendations
(1) People on the 0x125 microcode should move on to 0x129 (or newer versions) as fast as possible, especially if you are using the standard Intel Defaults mode.

(2) People on the newer BIOS versions should consider the new Intel Defaults mode (with its default settings) as being targeted at those (especially i9 owners) who were suffering from out-of-the-box instability due to too little voltage. For the vast majority of Raptor Lake owners, these settings do not necessarily apply to you, with perhaps the exception being the recommended power limits.

(3) Peak amps (IccMax) should be limited accordingly, with perhaps 307A being a good limit for the power hungry CPUs. [Note: It's likely that transient/extremely short current spikes can exceed this figure.]

(4) CiTay’s approach of advising people to lower CPU Lite Load and power limits is more important than ever.

(5) Clock-stretching, which dynamically lowers operating frequency to match a drop in voltage, will likely replace Loadline tuning at some point in the future. Therefore, enabling CEP should be part of your plan, if at all possible, so you get used to tuning with it enabled. It might also have some benefits - as seen by Charonme’s recent testing. See Level Four Knowledge.

(6) A LLC setting in the middle of the range (4,5,6) will likely offer more upsides than downsides. For example, it makes it easier to lower the AC_LL setting and do it while keeping CEP enabled. It also offers the potential for higher manual negative offsets because all-core Vdroop will be less significant.
Note: Vassil_V’s case study (see post below for link) already includes a comparison of LLC=6 versus LLC=Auto.

(7) If you implement a manual all-core overclock/max core ratio then lowering AC_LL could be even more important (than manual negative voltage offsets) due to how all the cores will boost as if they are one. [EDIT: Statement still needs to be verified.]

(8) Don’t disable default C-States - may cause high VIDs outside of Windows.

(9) Don’t use Windows High Performance plan (unless you absolutely cannot live with Balanced mode due to latency issues) because it may increase the time that all-cores remain active, and has the potential to significantly impact average voltages and general wear-and-tear.
Note: There are ways to manually tune Windows to create a hybrid mode with some of the benefits of both Balanced and High Performance modes. This is likely a better approach to address perceived latency issues.

(10) In theory, a LLC setting that is more aggressive (e.g. 4,5,6) than the Intel Defaults (LLC=Auto=8), when combined with a lower AC_LL, should result in:
  • >Lower all-core VIDs
  • >Slightly higher all-core Vcore
  • >Slightly higher all-core temperatures / power usage
  • >Lower non all-core VIDs (especially when manually undervolting)
  • >Lower non all-core Vcore
  • >Lower core temperatures for non all-core workloads (e.g. games)
  • >Highest voltage (due to transient spikes) will occur when current is falling, which I think is less of a degradation concern than high voltage (due to elevated AC_LL) when current is rising.
In my personal opinion, the above configuration should prove more suitable for normal daily activities (e.g. web browsing, Email, YouTube, Word, Excel, gaming, etc.) while having a relatively small detrimental effect on heavily multithreaded applications.

Tip: Recent user trends are clearly showing that people who override the LLC setting seem to prefer mode 5 and mode 6. I think these two modes offer a good balance of Vdroop (for all-core workloads) and healthy manual undervolting potential for gaming workloads. Access our Raptor Lake Survey Settings “database” to see the full range of settings that others have chosen.

EDIT: Another voltage related future direction for Intel was recently introduced with the Arrow Lake architecture. Read about it here.
 
Last edited:
[LEVEL FOUR KNOWLEDGE]

The main focus of the next links are discussions and oscilloscope investigations of CEP‘s behavior. These are definitely edge-conditions and most users can skip this level unless they’re truly interested in watching the on-going CEP debate. As one of the conclusions of the post above, I do still see a CEP-type mechanism as the future direction for Intel. Perhaps we‘ll learn more about this direction from studying Arrow Lake.

Charonme's in depth testing of how CEP enabled vs. disabled affects system behaviour in subtle ways (e.g. voltages, temperatures, power, core frequencies):
enabling CEP results in better performance at particular settings
Related: 14700K CB R23 scores curve vs. power limits between 15W and 260W
Related: Is it safe to disable IA CEP and IA CEP 14 GEN and use lite load with 0x129
 
Last edited:
[LEVEL FIVE KNOWLEDGE]

How to tune your DC_LL setting to improve the accuracy of the VID reporting (and CPU Package power) in applications like HWInfo64.
<placeholder>

Voltage Bible TL/DR Bullet List
<placeholder>

Stability Testing Guides
<placeholder>
 
Last edited:
Nice write up.

Looking forward to more.
Thanks Arctucas! I thought it might be useful if I tried to consolidate all this stuff into one place and then tried to present it as various levels of "rabbit hole" depth. I hope this format helps someone new to the forum. If nothing else, it might serve as a "guide to the guides."
 
[LEVEL SIX KNOWLEDGE: Application Tips]

HOW TO CREATE CUSTOM SENSORS IN HWINFO64

See the attached instructions.

EXAMPLE: The effective clocks in HWInfo64 are ultimately more useful to track than the regular core clocks because, in addition to the BIOS settings that influence voltage, power consumption and core frequencies, they are also very sensitive to background processes and C-States. For run-to-run variance in benchmark scores, etc., your average effective clocks tell a more complete story.

However, for Intel's new Big.Little core architectures (e.g. Alder Lake, Raptor Lake, Arrow Lake) it's far more useful to see the Average Effective Clocks reported separately for P-Cores and E-Cores rather than the standard Average Effective Clock sensor that combines these two IA Core types into one number. And you might just find that extra granularity of two separate averages more useful when comparing your run-to-run performance with all-core applications like Cinebench R23. The attached guide covers how you can use the "custom sensors" feature of HWInfo64 to achieve this goal. Be warned, however, that you should probably make a backup of the Windows Registry before you start making changes to it. You will also need to close and restart HWInfo64 for the changes to take effect.

[EDIT: I’m still evaluating the above suggestion to see if the averages of the custom sensors are negatively affected by your HWInfo64 polling period, and a few other things. So, still a work-in-progress.]

HOW TO FIX AN APPLICATION THAT USES E-CORES AND IGNORES P-CORES

If you have an application that is using E-Cores when it should be using P-Cores, try using an application to assign the worker threads to only the P-Cores. One such application is CoreDirector. See attached instructions.
 

Attachments

  • HWInfo64 How to create Custom Sensors.pdf
    540.7 KB · Views: 162
  • CoreDirector.pdf
    282.2 KB · Views: 24
Last edited:
Below is an example of how one user tried to find (read "guess") the impedance values associated with each rung on the LLC ladder. This information is important as it makes tuning the AC_LL and DC_LL values much easier, especially if you plan to keep IA CEP enabled and happy! More on that later...

View attachment 194990
Table A - LLC Impedance relationship for MSI Z690 Edge motherboard
Hey there -- I'm the person that generated that table over in this thread on overclock.net. I wish it was all about guessing, but it took a over day of experimentally dialing in ACLL=DCLL under full load to generate VRVOUT as close as possible (usually <~.003V) to VID for every LLC level. I still have all my original notes/measurements from two years ago, in fact.

Many kudos to you for this guide as well as the other work being posted on here. Things have really come a long way in the last two years.
 
Last edited:
Hey there -- I'm the person that generated that table over in this thread on overclock.net. I wish it was all about guessing, but it took a over day of experimentally dialing in ACLL=DCLL under full load to generate VRVOUT as close as possible (usually <~.003V) to VID for every LLC level. I still have all my original notes/measurements from two years ago, in fact.

Many kudos to you for this guide as well as the other work being posted on here. Things have really come a long way in the last two years.
Wow! That was some great original work you and others did on that thread. It was a goldmine of loadline insight! And I know that several of our regular MSI Forum members also contributed in one way or another.

Thank you for popping in to check us out. I definitely appreciate the thumbs up! I still have a lot to learn about the behavior of the loadline - and the various relationships - so I still consider this guide as a work-in-progress affair. However, I wanted to leave as many breadcrumbs as possible (and as quickly as possible) for others to follow. It was also a good place for me to accumulate my findings for my own benefit.

If you're planning on sticking around these parts for a while, I will certainly look forward to learning from any wisdom you have to share with us. I hope you don't mind that I already borrowed from your past work. :-) Cheers!
 
I'll certainly be around, but you, Vassil_V, citay, and several others have really carried the banner. The guides you guys have been kicking out are fabulous and what has always been needed by the MSI user community which has, for a long time, suffered from a lack of clear technical documentation/specifications. In fact, I complained about this back in my thread years ago. :cry:
 
I'll certainly be around, but you, Vassil_V, citay, and several others have really carried the banner. The guides you guys have been kicking out are fabulous and what has always been needed by the MSI user community which has, for a long time, suffered from a lack of clear technical documentation/specifications. In fact, I complained about this back in my thread years ago. :cry:
I couldn’t agree with you more. The work that Citay and Vassil_V have done (and many others who contributed) has really helped my own knowledge immensely. And I know they have helped thousands of other people who were left in a very uncomfortable place when Intel dropped the ball. Collectively, I think we managed to give Raptor Lake owners a personal sense of control over the situation. Hopefully, people would agree with that statement. I think we managed to put the fire out just in time. I mean, citay's guide has already reached twice as many people than you can pack into Wembley Stadium at a Taylor Swift concert - or maybe The Rolling Stones if you're old like me!
 
Last edited:
I can't agree with you more. The work that Citay and Vassil_V (and many others who contributed) have done has really helped my own knowledge immensely. And I know they have helped thousands of other people who were left in a very uncomfortable place when Intel dropped the ball. Collectively, I think we managed to give Raptor Lake owners a personal sense of control over the situation. Hopefully, people would agree with that statement. I think we managed to put the fire out just in time. I mean, citay's guide has already reached twice as many people than you can pack into Wembley Stadium at a Taylor Swift concert - or maybe The Rolling Stones if you're old like me!
Well, I was born in the late '60s, so The Stones work for me. I'm also a retired electrical engineer, so I get frustrated with the lack of official documentation. If it wasn't for the underappreciated enthusiasts out there, the situation would be way worse. The fact that an MSI rep stated that their LLC impedances were some sort of company proprietary information at one point boggles my mind. There is zero competitive advantage to keep that information secret, and it only serves to frustrate a community that is filling in all the gaps left by Intel and the board manufacturers (MSI certainly isn't the only one) that has led to the current meltdown -- literally, meltdown.

The latest BIOS jacked default Normal Auto mode to 18(!) for my 13600k on my MPG Z690 EDGE WIFI DDR4. I don't use Normal Lite Load, but I certainly took a look at all the defaults after a CMOS clear too see what changed. That's higher voltage than what I first saw on the board two years ago when I already knew it was too much. That's the whole reason I went down the LLC/ACLL/DCLL path back then. It's crazy to me that the situation is basically worse now as a result of all the degradation.
 
I definitely agree with you that modern life is made so much more difficult due to the lack of good technical documentation. Thank god we have YouTube to help us out some. Personally, I actually enjoy documenting things, especially if I’m learning as I go. I assume that makes me an oddity in this world. Case in point: Last year, I started documenting MSI Afterburner - which I had used for a decade or more. Quite frankly, I was blown away by all the things I never knew it could do.

To illustrate how things have changed, I recently found a booklet that came with one of the old Tex Murphy games, Under the Killing Moon (1994). It was a really nice large print 25+ page booklet with everything from setup and troubleshooting advice to a printout of key assignments. Now, those were the good old days of PC ownership! Same goes for my old TVs and AVRs from 2000 - 2014. They came with 100+ page books, packed full of setup tips and technical specifications.
 
Last edited:
I had all the Tex Murphy games; I remember them well. And the best manual included with a game in history was the hard covered clip-binder manual for Spectrum Holobyte's Falcon 3.0 in 1991. Just amazing. Heck, I still remember when they put schematics inside the backs of all electronics to help make repairs. Anyway, I'll stop derailing this very useful thread. :-)

Edit: The Mods have graciously allowed me to update my username to match my old one over on overclock.net.
 
Last edited:
Back
Top