Higher Vcore & natural vdrop VS lower Vcore & higher LLC: which is safest/best?

TheBlackHole · Nov 18, 2020

Hi folks,
once found a target Vcore voltage that is sufficient to guarantee full system's stability, which is the best and safest choice to achieve it?

Priority is (for OC 24/7):
1) CPU preservation (no degradation and/or risk of damage, both in the short and long run)
2) efficiency, as the ability to better maintain voltage stability

As per the title, I'd like to know your opinions about which path is best:
A) set a higher Vcore on BIOS (compared to the target) that under full load drop naturally until the desired target voltage (LLC set as "flat")
B) set the a Vcore close to the target on BIOS and then keep it stable by higher LLC level (a somehow aggressive level could be needed)

This interests me especially for fairly high voltages.

Example: TARGET Vcore 1.370, LLC levels 1-8 (1 is the highest, 4 is "flat")

if A): Vcore set to ~1.390-1.400 + LLC 4
if B): Vcore set to ~1.355/1.360 + LLC 2 (with LLC 3 is ~1.368-1.376 under full load according to HWInfo, almost 100% stable but not quite yet - so a more aggressive LLC level is required)

Any feedback would be greatly appreciated! Thanks!!!

The Pook · Nov 18, 2020

as long as you aren't on the most extreme LLC it really doesn't matter too much at those voltages, but you shouldn't aim to get rid of vdroop completely. a bit of droop is better than overshoot.

worth a read:

Load-Line Calibration (LLC) - WikiChip

Load-Line Calibration (LLC) is a mechanism offered to overclockers designed to compensate for large voltage droops when a CPU or GPU is under increased load. The mechanism attempts to compensate for the sudden sagging in voltage by preemptively applying additional voltage. The LLC, which is part...

en.wikichip.org

TheBlackHole · Nov 18, 2020

The Pook said:
as long as you aren't on the most extreme LLC it really doesn't matter too much at those voltages, but you shouldn't aim to get rid of vdroop completely. a bit of droop is better than overshoot.

worth a read:

Load-Line Calibration (LLC) - WikiChip

Load-Line Calibration (LLC) is a mechanism offered to overclockers designed to compensate for large voltage droops when a CPU or GPU is under increased load. The mechanism attempts to compensate for the sudden sagging in voltage by preemptively applying additional voltage. The LLC, which is part...

en.wikichip.org

Thank you for the reply and the useful link!

For my MB, level 2 is the second highest: I thought that going beyond level 3 (level 5 for non MSI boards) was not a great idea, especially on not so great MB like mine, since VRMs are far from the best and I was worried about possible dangerous voltage spikes, etc.
But then I read that many people recommend to use level 6 (on Asus boards, etc.: should be equivalent to level 2 on MSI boards) even applied to high starting Vcore.

If you had to make a choice, which one would you pick between the options A) and B) that I described in my op post?
Is, for ex., 1.390 (as stating Vcore) still completely safe (= no degradation) for 24/7 OC?

Thanks!

shellashock · Nov 19, 2020

There is no such thing as "completely safe" voltages since the act of applying voltage to pass current through the traces will result in varying amounts of electromigration (see Buildzoid's video about chip degradation), but that's besides the point. The main thing you should be worried about is keeping the average voltage under load as low as possible for your desired stability threshold. Using lower levels of LLC will allow for greater minimum voltage (according to oscilloscope) at expense of higher idle voltage; which translates into better stability (see any of Buildzoid's Probinator videos). There are transient response differences between each LLC level as well (board dependent), with the best settings being the one with the lowest difference between average and minimum voltages (according to oscilloscope) while still having reasonable idle/low load voltages.

Realistically, the best LLC setting is usually in the middle to higher levels of LLC (extreme or no LLC is almost never the right choice).

TheBlackHole · Nov 19, 2020

shellashock said:
There is no such thing as "completely safe" voltages since the act of applying voltage to pass current through the traces will result in varying amounts of electromigration (see Buildzoid's video about chip degradation), but that's besides the point. The main thing you should be worried about is keeping the average voltage under load as low as possible for your desired stability threshold. Using lower levels of LLC will allow for greater minimum voltage (according to oscilloscope) at expense of higher idle voltage; which translates into better stability (see any of Buildzoid's Probinator videos). There are transient response differences between each LLC level as well (board dependent), with the best settings being the one with the lowest difference between average and minimum voltages (according to oscilloscope) while still having reasonable idle/low load voltages.

Realistically, the best LLC setting is usually in the middle to higher levels of LLC (extreme or no LLC is almost never the right choice).

First, let me thank you for the informative reply and the links.
I know what electromigration is, but I read an interesting article in which Intel's official team specialized in testing CPU overclocking stated that these CPUs are fine till 1.425V IF:
1) you manage to keep max core temps <80C
2) keep idle core temps <30C
3) keep dynamic voltage mode as well as SpeedStep and C states on

For what they said, it seems that you should not incur in electromigration / chip degradation as long as the aforementioned "rules" are observed. Also, it seems that overshot is way more dangerous than high voltage set in BIOS: for this reason, till now I was very cautious and didn't go past level 3 (level 5 for non MSI boards) which is basically the minimum LLC level that "does something".

For the same reason and for what you wrote too, I'm wondering if it could be an even better choice to raise the starting Vcore in BIOS and set a "flat" LLC (level 4) instead of the level 3 I used till now, so that the risk of serious overshot would become zero and transient response would be the best possible.
Of course I want to keep the voltage as low as possible, but it seems that for 100% stability I need steady 1.370: I tried 1.360 (BIOS) + LLC 3 resulting in 1.368-1.376 under heavy full load...almost nailed it, since the system was able to pass numerous stress tests but failed my "special" test after 10h~ (1 WHEA internal parity error, corrected). That was most likely caused by a sudden (and rare, for what I saw from the tests) vdrop under the minimum stability threshold: to avoid this, I can only:

A) raise LLC level to 2, keeping 1.360V in BIOS
B) raise Vcore to 1.365-1.370 & keep LLC 3
C) raise Vcore to ~1.390(?) & lower LLC to level 4 (flat)

So, among the 3 choices listed above, which one do you think is best for the for the priorities I have described in my op post? (first, preserving CPU from degradation / dangerous voltage spikes)

GeneO · Nov 19, 2020

Electromigration depends on current and temperature. It always degrades the traces on a chip over time no matter what the temperature or current, it is a matter of to what degree. What is behind what Intel is saying is that if you keep withing those guidelines, the chip can last for a reasonable time before the degradation affects function, not that it won't degrade. Also, the smaller the lithography on the chip, the finer the conductive traces and the more susceptible they are to electromigration damage. But the smaller the lithography is balanced by less current needed for operation.

I would caution this. It is really the current and temperature per core that matter. If you run 10 cores at 250A as safe (25A/core), it is does not follow that it is safe to to be running 4 cores at 150A. (37.5A/core). Simplified an maybe unrealistic numbers, but hope you get what I mean. This is something you should be aware of if you are doing per-core multipliers.

TheBlackHole · Nov 19, 2020

GeneO said:
Electromigration depends on current and temperature. It always degrades the traces on a chip over time no matter what the temperature or current, it is a matter of to what degree. What is behind what Intel is saying is that if you keep withing those guidelines, the chip can last for a reasonable time before the degradation affects function, not that it won't degrade. Also, the smaller the lithography on the chip, the finer the conductive traces and the more susceptible they are to electromigration damage. But the smaller the lithography is balanced by less current needed for operation.

I would caution this. It is really the current and temperature per core that matter. If you run 10 cores at 250A as safe (25A/core), it is does not follow that it is safe to to be running 4 cores at 150A. (37.5A/core). Simplified an maybe unrealistic numbers, but hope you get what I mean. This is something you should be aware of if you are doing per-core multipliers.

I get that the chip will not last forever...yeah, I agree that what intel said must be interpreted as "it will last a long time" (by the way, I just finished seeing Buildzoid's video about chip degradation - very interesting!).

I was used to just set fixed, manual Vcore and disable SpeedStep & all C states but after reading that article I decided to set adaptive voltage mode (which works much better than the old "offset" mode) and re-enable SpeedStep & C states: from your understanding / knowledge, you think this is wrong and/or may cause issue with OC?

I ask because you mentioned per-core multipliers, but I did not set that mode in BIOS, its just "All Core" + SpeedStep & adaptive mode / C states: basically, it runs as if it were "a stock" but with a higher Vcore limit (of course, set by me).

And...what is your choice? A), B) or C)? 😄

Thanks!

shellashock · Nov 19, 2020

TheBlackHole said:
First, let me thank you for the informative reply and the links.
I know what electromigration is, but I read an interesting article in which Intel's official team specialized in testing CPU overclocking stated that these CPUs are fine till 1.425V IF:
1) you manage to keep max core temps <80C
2) keep idle core temps <30C
3) keep dynamic voltage mode as well as SpeedStep and C states on

For what they said, it seems that you should not incur in electromigration / chip degradation as long as the aforementioned "rules" are observed. Also, it seems that overshot is way more dangerous than high voltage set in BIOS: for this reason, till now I was very cautious and didn't go past level 3 (level 5 for non MSI boards) which is basically the minimum LLC level that "does something".

For the same reason and for what you wrote too, I'm wondering if it could be an even better choice to raise the starting Vcore in BIOS and set a "flat" LLC (level 4) instead of the level 3 I used till now, so that the risk of serious overshot would become zero and transient response would be the best possible.
Of course I want to keep the voltage as low as possible, but it seems that for 100% stability I need steady 1.370: I tried 1.360 (BIOS) + LLC 3 resulting in 1.368-1.376 under heavy full load...almost nailed it, since the system was able to pass numerous stress tests but failed my "special" test after 10h~ (1 WHEA internal parity error, corrected). That was most likely caused by a sudden (and rare, for what I saw from the tests) vdrop under the minimum stability threshold: to avoid this, I can only:

A) raise LLC level to 2, keeping 1.360V in BIOS
B) raise Vcore to 1.365-1.370 & keep LLC 3
C) raise Vcore to ~1.390(?) & lower LLC to level 4 (flat)

So, among the 3 choices listed above, which one do you think is best for the for the priorities I have described in my op post? (first, preserving CPU from degradation / dangerous voltage spikes)

I bet you're referring to the (excellent) Tom's Hardware article: "Inside Intel's Secret Overclocking Lab". Note that the interviewees never say that the chip will not degrade if you follow these rules; just that "you can make your part last quite a while". IIRC, the adaptive mode they advocated for will have harsher overvoltage spikes (on an oscilloscope) upon load transitions due to the greater current difference but is able to have much lower average voltages at idle; so the overall effect on lifespan (assuming oxide breakdown is not much of a factor for these voltages) should be positive.

I don't have any good sources on oxide breakdown for voltage overshoots during load transients (would love to see examples if anyone has them), but I would mostly be concerned about this when using voltages that already have a good chance of causing oxide breakdown (absolute limits of "safe" voltages at minimum; more realistically the kinds of voltages used for extreme overclocking). The amount of overshoot is for a very short period of absolute time and is unlikely to cause as much damage as having a high average voltage at low load for extended periods of time (say heavy OC for a single core workload).

If I had to pick, I would probably choose B if you could stabilize at 1.365 and A if you need 1.370+ to stabilize at LLC 3. Without seeing the results with an oscilloscope, it really is a crap shoot and you will likely be fine with all three of those options if they really do have the same average voltage at high load. Truthfully, if you are really that concerned about 24/7 OC with minimal degradation risk, you need to ask yourself if you really need that additional multiplier and what voltage it is costing you to reach it. If you can knock 100mV off the stability threshold by reducing the multiplier by 1, you will have EASILY surpassed the lifespan increase ANY of these options could ever hope to give you; let alone the reduced power usage and thermal considerations. Heck, lowering the voltage by 100mV would probably lower your temperatures enough that the lifetime increase of a cooler chip would be within margin of error of the difference playing with (non extreme) LLC settings would give you. All speculation of course, but the point still stands.

Alternatively, play around with adaptive mode like Intel recommends and see what you can achieve. You will probably have a harder time stabilizing in programs that have heavy load cycling like P95 128k due to the harsher load transitions, but it should give you the lifespan improvements you desire (at least going by Intel's engineers).

TheBlackHole · Nov 19, 2020

shellashock said:
I bet you're referring to the (excellent) Tom's Hardware article: "Inside Intel's Secret Overclocking Lab". Note that the interviewees never say that the chip will not degrade if you follow these rules; just that "you can make your part last quite a while". IIRC, the adaptive mode they advocated for will have harsher overvoltage spikes (on an oscilloscope) upon load transitions due to the greater current difference but is able to have much lower average voltages at idle; so the overall effect on lifespan (assuming oxide breakdown is not much of a factor for these voltages) should be positive.

I don't have any good sources on oxide breakdown for voltage overshoots during load transients (would love to see examples if anyone has them), but I would mostly be concerned about this when using voltages that already have a good chance of causing oxide breakdown (absolute limits of "safe" voltages at minimum; more realistically the kinds of voltages used for extreme overclocking). The amount of overshoot is for a very short period of absolute time and is unlikely to cause as much damage as having a high average voltage at low load for extended periods of time (say heavy OC for a single core workload).

If I had to pick, I would probably choose B if you could stabilize at 1.365 and A if you need 1.370+ to stabilize at LLC 3. Without seeing the results with an oscilloscope, it really is a crap shoot and you will likely be fine with all three of those options if they really do have the same average voltage at high load. Truthfully, if you are really that concerned about 24/7 OC with minimal degradation risk, you need to ask yourself if you really need that additional multiplier and what voltage it is costing you to reach it. If you can knock 100mV off the stability threshold by reducing the multiplier by 1, you will have EASILY surpassed the lifespan increase ANY of these options could ever hope to give you; let alone the reduced power usage and thermal considerations. Heck, lowering the voltage by 100mV would probably lower your temperatures enough that the lifetime increase of a cooler chip would be within margin of error of the difference playing with (non extreme) LLC settings would give you. All speculation of course, but the point still stands.

Alternatively, play around with adaptive mode like Intel recommends and see what you can achieve. You will probably have a harder time stabilizing in programs that have heavy load cycling like P95 128k due to the harsher load transitions, but it should give you the lifespan improvements you desire (at least going by Intel's engineers).

Yes, that was the article I read. I too was concerned about the overvoltage spikes that the adaptive mode involves, but I thought that Intel's boys know better and moreover this (adaptive mode) is how the CPU is supposed to work at stock: even though in that case the max Vcore is lower than 1.370 (just as example), if Intel boys said to use this mode and that they themselves set their own CPUs to 1.425V I suppose that this means the CPU should withstand this stress.

Thanks you for your advice: I'm currently trying to achieve the max frequency my CPU is able to reach while keeping temps <80 at full load, hence the 1.370 voltage. But it's not my daily setting: for 24/7 OC, Vcore is set to 1.310V + LLC 3 (according to HWInfo, 1.320-1.328 under full load). I want to find out what is the best mode to achieve a certain voltage not only for the 1.370V profile, but also to eventually modify my daily setting (for ex with: ~1.350 Vcore + LLC 4...could be better?)

What do you think of my current daily profile? It's ok for a 24/7 use and "long lifespan"? (at least 5+ years)

By the way, my temps at full load are ~55-60C (Prime 95 AVX ~70)

Thanks again for your great assistance, I really appreciated!

o1dschoo1 · Nov 19, 2020

Let's use my cpu for example 4960x at 4.6 1.26 vcore medium llc. Run these settings on a stock cooler at 90c and watch the cpu degrade quickly.
Temperature plays a huge part. The cooler the cpu runs the slower electromigration happens and also better your cpu will respond to voltage.

TheBlackHole · Nov 19, 2020

o1dschoo1 said:
Let's use my cpu for example 4960x at 4.6 1.26 vcore medium llc. Run these settings on a stock cooler at 90c and watch the cpu degrade quickly.
Temperature plays a huge part. The cooler the cpu runs the slower electromigration happens and also better your cpu will respond to voltage.

View attachment 2465867

Yeah, I know. I think my temps should be ok...at least I hope.

My current daily OC profile:
4.9 Ghz / 4.4 Ghz RING (probably it can be raised to 4.6, I haven't even bothered to try yet)
Vcore 1.310 (BIOS) + LLC 3 (one level above the 4, that is "flat") ---> 1.320-1328 full load (HWInfo)
Max core temps @ full load: 55-60C (Prime 95 AVX ~70C)

What do you think?

Thanks for your feedback!

o1dschoo1 · Nov 19, 2020

TheBlackHole said:
Yeah, I know. I think my temps should be ok...at least I hope.

My current daily OC profile:
4.9 Ghz / 4.4 Ghz RING (probably it can be raised to 4.6, I haven't even bothered to try yet)
Vcore 1.310 (BIOS) + LLC 3 (one level above the 4, that is "flat") ---> 1.320-1328 full load (HWInfo)
Max core temps @ full load: 55-60C (Prime 95 AVX ~70C)

What do you think?

Thanks for your feedback!

That is perfectly fine. Actually not bad temps at all. Stay as low as you can get temp wise

GeneO · Nov 19, 2020

TheBlackHole said:
I get that the chip will not last forever...yeah, I agree that what intel said must be interpreted as "it will last a long time" (by the way, I just finished seeing Buildzoid's video about chip degradation - very interesting!).

I was used to just set fixed, manual Vcore and disable SpeedStep & all C states but after reading that article I decided to set adaptive voltage mode (which works much better than the old "offset" mode) and re-enable SpeedStep & C states: from your understanding / knowledge, you think this is wrong and/or may cause issue with OC?

I ask because you mentioned per-core multipliers, but I did not set that mode in BIOS, its just "All Core" + SpeedStep & adaptive mode / C states: basically, it runs as if it were "a stock" but with a higher Vcore limit (of course, set by me).

And...what is your choice? A), B) or C)? 😄

Thanks!

I run adaptive, c-states enabled up to C7, and speedshift enabled. No issues here with OC. I have been doing this for awhile. My 8086k at 5.2 GGHz (5.3 without hyperthreading) and 5.1 all-core on my 10700k (1.30v die sense under load).

Don't focus on voltage so much as current and temperature - that is what degrades through electromagnetic. There certainly voltage methods of degradation like breakdown at the junctions that high voltage spikes can bring about though, but that is pretty high bar.

TheBlackHole · Nov 19, 2020

GeneO said:
I run adaptive, c-states enabled up to C7, and speedshift enabled. No issues here with OC. I have been doing this for awhile. My 8086k at 5.2 GGHz (5.3 without hyperthreading) and 5.1 all-core on my 10700k.

Thanks!
This is very precious info, exactly what I was looking for: so your personal experience confirms that what Intel claims (see previous posts) is true. I was a bit worried about the sudden, repeated and HUGE shift in voltages that adaptive mode involves so I was looking for some first hand experience in using it with a highly OC rig.

If you have read the previous posts of this thread, may I ask which do you think is the best way to set a certain voltage and keep it stable, without risking to degrade/damage the CPU? (see my example above, option A), B) or C)

About current and temp: then, Prime 95 is basically a CPU killer - especially with AVX enabled. At the high voltage and clock speed I was testing, the load on my 9600K is around 150W, which should be ~110A (roughly 18,5A per core). What do you think, it's too much even if temps are below 80C?

JackCY · Nov 19, 2020

TheBlackHole said:
Example: TARGET Vcore 1.370, LLC levels 1-8 (1 is the highest, 4 is "flat")

Measured as flat by who? Did you measure the Vcore with an oscilloscope? Anything else is guessing and relying on the unreliable mobo sensors.

if A): Vcore set to ~1.390-1.400 + LLC 4
if B): Vcore set to ~1.355/1.360 + LLC 2 (with LLC 3 is ~1.368-1.376 under full load according to HWInfo, almost 100% stable but not quite yet - so a more aggressive LLC level is required)

Any feedback would be greatly appreciated! Thanks!!!

In the end it doesn't matter, you need the same voltage for stability and it's safer to do it via voltage control rather than forcing LLC overshoot.

Usually the best is somewhere in the middle. You get a little Vdroop but not too much. If that Vdroop even matters to you as modern CPUs have all sorts of voltage regulation, some on mobo entirely, some part of it on mobo and rest on CPU, so while modern mobos may show you in UEFI Vcore it can actually be Vccin as some modern CPUs have their own regulation built in whether that is advertised or not, doesn't have to be a FIVR just LDOs are enough.

TheBlackHole · Nov 19, 2020

JackCY said:
Measured as flat by who? Did you measure the Vcore with an oscilloscope? Anything else is guessing and relying on the unreliable mobo sensors.

In the end it doesn't matter, you need the same voltage for stability and it's safer to do it via voltage control rather than forcing LLC overshoot.

Usually the best is somewhere in the middle. You get a little Vdroop but not too much. If that Vdroop even matters to you as modern CPUs have all sorts of voltage regulation, some on mobo entirely, some part of it on mobo and rest on CPU, so while modern mobos may show you in UEFI Vcore it can actually be Vccin as some modern CPUs have their own regulation built in whether that is advertised or not, doesn't have to be a FIVR just LDOs are enough.

What I mean by "flat" is that, according to MSI and BIOS description, level 4 on my board does basically nothing about Vdroop. And I could see exactly that by sensors report on HWInfo...and NO, I don't have an oscilloscope - I'm just your average Joe - so I must rely on what sensors report.

On the other hand, it would still be useless to have an oscilloscope to determine if level 4 is really "flat", as the other levels (5-8) even favor a slight Vdroop, so in any case they are even less "flat" (not to mention levels 1-3 which favor overshot to varying degrees).

Anyway thank you for the feedback: so, if you read my previous posts on this thread, which option you would choose between A),B) or C)?

Higher Vcore & natural vdrop VS lower Vcore & higher LLC: which is safest/best?

TheBlackHole

The Pook

Load-Line Calibration (LLC) - WikiChip

TheBlackHole

Load-Line Calibration (LLC) - WikiChip

shellashock

TheBlackHole

GeneO

TheBlackHole

shellashock

TheBlackHole

o1dschoo1

TheBlackHole

o1dschoo1

GeneO

TheBlackHole

JackCY

TheBlackHole

Top Contributors this Month

Recommended Communities