Can CPU PL2 be changed with an EFI application (like CFGLock)?

[rafale77]

I will test this as soon as I can on my main machine following your awesome work. I think mine will have quite a bit more headroom.

My plan to test performance improvements:

1. Undervolt. I am targeting stress testing at about -50mV

2. Repaste the CPU HSF.

3. Increase PL2 to 150W (which on my machine seems to be reasonable given my 230W PSU)

 

Thanks for sharing all your meticulous work as usual.

[deeveedee]

@rafale77 You deserve the credit for first mentioning VoltageShift in this thread.  Good luck with your testing.  Looking forward to seeing your method and results.

[WhenMusicAttacks]

On 5/4/2022 at 7:01 PM, rafale77 said:

@deeveedee, congratulations on getting to the PL values! I supposed you have a 150W power brick correct?

My research about the coffee lake i9 lead me to very different conclusions from @WhenMusicAttacks,

Someone ran pretty extensive tests on another forum and concluded the following default power limits:

9900: PL1= 65W, PL2= 225W

9900K: PL1=95W, PL2= 255W

9900T: PL1=35W, PL2= 105W

 

On the Lenovo P330 Vs. M920X, this has explained a number of reports from users experiencing improved CPU performance on the 9900T going from a 90W power brick to 135W. The Lenovo though is able to adjust power limits according to the AC power brick it is plugged to. Not so on the HP. As you know I have switched to the Lenovo P340tiny and have observed a PL2 of 135W on the 10900 which has a default PL2 also of 225W which also explains why there is still a gap in multi core performance between what I am observing and the average for that CPU.

 

That being said, lowering the voltage should also help improve performance by decreasing thermal throttling…

 

You do not understand what you are writing, let alone what i wrote. Of course if your talk about 90w power brick you are referring to the nominal power from the wall, wich then has to be actually fed to the whole machine (ram, ssd, pch) and then to the cpu vrm, with every step losing power depending on the component efficiency. The 9900 will use up to around 120w (upd to 150, depending on avx2 offsets) after the power brick and VRM efficiency losses, the latter being probably massive if ran out of their default 65w design. That's why you can have better result with a 135w power brick, but the VRM is still designed to handle the default limit, and will have issue in the long run, not to mention overheating and lower efficiency the higher you go.

That's exactly why you will have much better results undevolting, keeping the default limits and improing efficiency. You can usually get around -80 on 9900, on average, at unlimited power, with a power limit at 65w and repasting to avoid hotspots you can probably get -100 stable.

The 10900 uses almost 200w, so it's a totally different story, as the pushed 14nm to crazy clocks - just a couple of mhz on the top of the curve around 4.8ghz will result in almost doubling the power usage, that is why they don't do it themselves. Sadly, 10th gen has plundervolt mitigations on by intel microcode so usually software control of voltage is not avaiable. That is why 9th gen notebook are actually better than 10th gen ones.

I undervolt at least 3 machine every single day in my lab. I know what i am talking about. You should NOT touch the power limit on pre built machines, expecially when you can just undervolt anyway.

Edited May 9, 2022 by WhenMusicAttacks

[rafale77]

1 hour ago, WhenMusicAttacks said:

You do not understand what you are writing, let alone what i wrote. Of course if your talk about 90w power brick you are referring to the nominal power from the wall, wich then has to be actually fed to the whole machine (ram, ssd, pch) and then to the cpu vrm, with every step losing power depending on the component efficiency. The 9900 will use up to around 120w (upd to 150, depending on avx2 offsets) after the power brick and VRM efficiency losses, the latter being probably massive if ran out of their default 65w design. That's why you can have better result with a 135w power brick, but the VRM is still designed to handle the default limit, and will have issue in the long run, not to mention overheating and lower efficiency the higher you go.

That's exactly why you will have much better results undevolting, keeping the default limits and improing efficiency. You can usually get around -80 on 9900, on average, at unlimited power, with a power limit at 65w and repasting to avoid hotspots you can probably get -100 stable.

The 10900 uses almost 200w, so it's a totally different story, as the pushed 14nm to crazy clocks - just a couple of mhz on the top of the curve around 4.8ghz will result in almost doubling the power usage, that is why they don't do it themselves. Sadly, 10th gen has plundervolt mitigations on by intel microcode so usually software control of voltage is not avaiable. That is why 9th gen notebook are actually better than 10th gen ones.

I undervolt at least 3 machine every single day in my lab. I know what i am talking about. You should NOT touch the power limit on pre built machines, expecially when you can just undervolt anyway.

 

Though you are right on many points, the 9900 unlimited runs up to >220W, not 80. Under load would casually run up to 160W for short term. This has been well documented and confirmed as I had both the K and the non K CPU and an OC friendly MB. The 9900K PL2 is 255W by default. That’s why at default setting there is very little performance difference between the unlocked and the locked CPU and started getting people to “overclock” the turbo by undervolting. The 10th and 9th Series i9 have very similar power limits. It was also a big outcry about intel CPU being able to consume 3 or 4X the power of their TDP ratings. It started with Coffee Lake Refresh(9th gen).

The VRM comment is very true and is why one needs to be extremely cautious with the PL2 increase… on my Asrock motherboard, the 10900 hits 225W and likewise my 9900 on a Gigabyte board hit 210W easily when benchmarking. It’s not reading from the wall, it’s reading from intel speedstep tool and and HWMonitor “CPU Package Power” reading on MacOS. I can’t comment on notebooks.

Also the power ratings from the power brick is not its power input, if you look closely at their ratings, it is actually it’s power output and the efficiency is already factored in. The largest efficiency loss is from the AC-DC conversion. The step down conversion also generates some loss but much less than what has already been taken by the power brick.

I was discussing hitting 135W with my 10900 on a Lenovo which has a 230W power brick and likely could handle 150W to the CPU leaving 80W for the GPU and all the other accessories. For the HP G5, with the 150W Power brick I would be looking to move to maybe 115W since it doesn’t have GPU. It’s still a far cry from the 225W it takes at default but it’s better than the 80ish I have been observing. I ran this same CPU on an mITX Gigabyte board using the same MacOS partition. The loss in multi core GB5 score is about 25% between the two computers.(7100 Vs 8800).

 

Reading through all the power specifications, for 65W TDP CPUs, the HP machines power only allow for the i3 to run unlimited and the Lenovos can run the i5 unlimited. Anything above it (i7 and i9) will force a lower PL2 than intel’s defaults and while the HPs are fixed by the motherboards, the Lenovos are dynamically adjusting the Power Limits according to what power brick they are connected to.

 

Some references here and here and here

 

[deeveedee]

My previous posts in this thread were reported after testing Big Sur.  I am finding that my attempts to load PikeRAlpha's AppleIntelInfo.kext in Monterey are resulting in system instability (system crashes when kext is loaded).  I haven't spent any time debugging and am just reporting this so that others are aware.

Edited May 9, 2022 by deeveedee

[Slice]

1 hour ago, deeveedee said:

My previous posts in this thread were reported after testing Big Sur.  I am finding that my attempts to load PikeRAlpha's AppleIntelInfo.kext in Monterey are resulting in system instability (system crashes when kext is loaded).  I haven't spent any time debugging and am just reporting this so that others are aware.

I confirm. The kext AppleIntelInfo causes crash of my system so why DumwinDumper contains such warning

Use at your own risk.

Because Piker was not careful in his solutions.

[deeveedee]

I attempted to undervolt my G4 Mini / i5-8600 using multiple versions of VoltageShift from different GitHub repos and found that undervolt failed (voltage offsets before and after were 0mV) in all cases.  I didn't spend any time to diagnose or debug and am just reporting my findings here.  I continue to test on my G4 Mini / i5-8600 running Big Sur.

 

EDIT: I'm not sure which G4 Mini or G5 Mini BIOS version introduced the Intel plundervolt fix.

 

EDIT2: I don't know much about the plundervolt fix, so take this with a few grains of salt:  After reviewing some older threads, I believe that Intel released the plundervolt fix as microcode update 0xCA.  If true, HP released the plundervolt update for the HP EliteDesk 800 G4 Mini with BIOS version 02.10.00 Rev.A (Feb 10, 2020).  This would mean that the last BIOS version to allow undervolting of the G4 Mini would be version 02.09.01 Rev.A released Nov 8, 2019.

 

EDIT3: After downgrading my HP EliteDesk G4 Mini BIOS to version 02.09.01 Rev.A, I am still unable to undervolt using VoltageShift.  I'll wait for someone else who knows more about this to suggest options.

 

EDIT4: Interesting note: zSphereZ's version of VoltageShift reports OC locked and reports PL1 and PL2.  As reported before, I am able to modify PL1 and PL2, but I am not able to change CPU or GPU voltage offset.

 

VoltageShift info (zSphereZ version) (PL1: 65W, PL2 85W)

Spoiler

 

VoltageShift info (zSphereZ version) (PL1: 75W, PL2 125W)

Spoiler

 

@rafale77 - I think you are correct in your suspicion about CFGLock needing to be disabled for undervolting.  It appears that I can change PL1 and PL2 on the EliteDesk Minis, but I won't be able to undervolt.

Edited May 10, 2022 by deeveedee

[rafale77]

@deeveedee, It seems like the OC Lock is a different bit from the CFG Lock and OC Lock is blocking undervolting. I wish there was a similar tool as the one for CFG Lock to find that bit. If found, one could edit it fairly easily with VoltageShift or other tools.

 

Edit: It appears that disabling the OC Lock in the UEFI should be fairly easy to do using the ControlMsrE2 tool in interactive mode as it enables unlocking other BIOS hidden options. I will try to test this next week searching the keyword “overclocking”. This should eliminate the plundervolt fix and enable undervolting.

Edited May 11, 2022 by rafale77

[WhenMusicAttacks]

On 5/9/2022 at 10:41 PM, deeveedee said:

I attempted to undervolt my G4 Mini / i5-8600 using multiple versions of VoltageShift from different GitHub repos and found that undervolt failed (voltage offsets before and after were 0mV) in all cases.  I didn't spend any time to diagnose or debug and am just reporting my findings here.  I continue to test on my G4 Mini / i5-8600 running Big Sur.

 

EDIT: I'm not sure which G4 Mini or G5 Mini BIOS version introduced the Intel plundervolt fix.

 

EDIT2: I don't know much about the plundervolt fix, so take this with a few grains of salt:  After reviewing some older threads, I believe that Intel released the plundervolt fix as microcode update 0xCA.  If true, HP released the plundervolt update for the HP EliteDesk 800 G4 Mini with BIOS version 02.10.00 Rev.A (Feb 10, 2020).  This would mean that the last BIOS version to allow undervolting of the G4 Mini would be version 02.09.01 Rev.A released Nov 8, 2019.

 

EDIT3: After downgrading my HP EliteDesk G4 Mini BIOS to version 02.09.01 Rev.A, I am still unable to undervolt using VoltageShift.  I'll wait for someone else who knows more about this to suggest options.

 

EDIT4: Interesting note: zSphereZ's version of VoltageShift reports OC locked and reports PL1 and PL2.  As reported before, I am able to modify PL1 and PL2, but I am not able to change CPU or GPU voltage offset.

 

VoltageShift info (zSphereZ version) (PL1: 65W, PL2 85W)

  Reveal hidden contents

 

VoltageShift info (zSphereZ version) (PL1: 75W, PL2 125W)

  Reveal hidden contents

 

@rafale77 - I think you are correct in your suspicion about CFGLock needing to be disabled for undervolting.  It appears that I can change PL1 and PL2 on the EliteDesk Minis, but I won't be able to undervolt.

 

Try clearcmos after bios downgrade. I recall HP locked in the november 2021 bios update , maybe try something before that timeframe.

 

On 5/9/2022 at 3:48 PM, rafale77 said:

 

Though you are right on many points, the 9900 unlimited runs up to >220W, not 80. Under load would casually run up to 160W for short term. This has been well documented and confirmed as I had both the K and the non K CPU and an OC friendly MB. The 9900K PL2 is 255W by default. That’s why at default setting there is very little performance difference between the unlocked and the locked CPU and started getting people to “overclock” the turbo by undervolting. The 10th and 9th Series i9 have very similar power limits. It was also a big outcry about intel CPU being able to consume 3 or 4X the power of their TDP ratings. It started with Coffee Lake Refresh(9th gen).

The VRM comment is very true and is why one needs to be extremely cautious with the PL2 increase… on my Asrock motherboard, the 10900 hits 225W and likewise my 9900 on a Gigabyte board hit 210W easily when benchmarking. It’s not reading from the wall, it’s reading from intel speedstep tool and and HWMonitor “CPU Package Power” reading on MacOS. I can’t comment on notebooks.

Also the power ratings from the power brick is not its power input, if you look closely at their ratings, it is actually it’s power output and the efficiency is already factored in. The largest efficiency loss is from the AC-DC conversion. The step down conversion also generates some loss but much less than what has already been taken by the power brick.

I was discussing hitting 135W with my 10900 on a Lenovo which has a 230W power brick and likely could handle 150W to the CPU leaving 80W for the GPU and all the other accessories. For the HP G5, with the 150W Power brick I would be looking to move to maybe 115W since it doesn’t have GPU. It’s still a far cry from the 225W it takes at default but it’s better than the 80ish I have been observing. I ran this same CPU on an mITX Gigabyte board using the same MacOS partition. The loss in multi core GB5 score is about 25% between the two computers.(7100 Vs 8800).

 

Reading through all the power specifications, for 65W TDP CPUs, the HP machines power only allow for the i3 to run unlimited and the Lenovos can run the i5 unlimited. Anything above it (i7 and i9) will force a lower PL2 than intel’s defaults and while the HPs are fixed by the motherboards, the Lenovos are dynamically adjusting the Power Limits according to what power brick they are connected to.

 

Some references here and here and here

 

  

On 5/9/2022 at 3:48 PM, rafale77 said:

 

Though you are right on many points, the 9900 unlimited runs up to >220W, not 80. Under load would casually run up to 160W for short term. This has been well documented and confirmed as I had both the K and the non K CPU and an OC friendly MB. The 9900K PL2 is 255W by default. That’s why at default setting there is very little performance difference between the unlocked and the locked CPU and started getting people to “overclock” the turbo by undervolting. The 10th and 9th Series i9 have very similar power limits. It was also a big outcry about intel CPU being able to consume 3 or 4X the power of their TDP ratings. It started with Coffee Lake Refresh(9th gen).

The VRM comment is very true and is why one needs to be extremely cautious with the PL2 increase… on my Asrock motherboard, the 10900 hits 225W and likewise my 9900 on a Gigabyte board hit 210W easily when benchmarking. It’s not reading from the wall, it’s reading from intel speedstep tool and and HWMonitor “CPU Package Power” reading on MacOS. I can’t comment on notebooks.

Also the power ratings from the power brick is not its power input, if you look closely at their ratings, it is actually it’s power output and the efficiency is already factored in. The largest efficiency loss is from the AC-DC conversion. The step down conversion also generates some loss but much less than what has already been taken by the power brick.

I was discussing hitting 135W with my 10900 on a Lenovo which has a 230W power brick and likely could handle 150W to the CPU leaving 80W for the GPU and all the other accessories. For the HP G5, with the 150W Power brick I would be looking to move to maybe 115W since it doesn’t have GPU. It’s still a far cry from the 225W it takes at default but it’s better than the 80ish I have been observing. I ran this same CPU on an mITX Gigabyte board using the same MacOS partition. The loss in multi core GB5 score is about 25% between the two computers.(7100 Vs 8800).

 

Reading through all the power specifications, for 65W TDP CPUs, the HP machines power only allow for the i3 to run unlimited and the Lenovos can run the i5 unlimited. Anything above it (i7 and i9) will force a lower PL2 than intel’s defaults and while the HPs are fixed by the motherboards, the Lenovos are dynamically adjusting the Power Limits according to what power brick they are connected to.

 

Some references here and here and here

 

I stated that i have no idea about the default AVX offsets - wich is what these numbers come from. I had no idea that locked cpus can still run without avx offset in place, wich is where the 200w usage in small FFT prime 95 avx comes from (very narrow usage example anyway). The power load is also tied to the vrm quality. I tried to run a 9900K on a gigabyte h310 motherboard and the power draw reported for cinebench r15 was 160w (on my z390 was around 120) - wich lasted a couple of seconds before crashing without any undervolt (i had the tdp unlocked in xtu).

Did not know about the power ratign on the brick adapters being relative to their actual output, sure they have to be very efficient as they are fanless and very small.
 

[deeveedee]

@WhenMusicAttacks I did reset BIOS to factory after the downgrade, but did not try an earlier BIOS version. I'll attempt an earlier version and report back.  Thank you.

[deeveedee]

@rafale77 I found my 150W genuine HP adapter and plugged it into my HP EliteDesk 800 G4 Mini / i5-8600.  I am finding that the G4 Mini auto-adjusted PL2 from 85W (with a 90W adapter) to 122W with a 150W adapter.  As you found with your Lenovo, the HP rigs do auto-adjust depending on the power rating of the adapter.  The synthetic performance gain (as measured with GB5) is immediate and is approximately a 10% performance improvement for the multi-core score.

 

HP EliteDesk 800 G4 Mini with 150W Power Adapter: Package Power Limits

Spoiler

 

 

HP EliteDesk 800 G4 Mini with 150W Power Adapter: GeekBench 5 CPU Benchmarks

Spoiler

 

Edited May 12, 2022 by deeveedee

[rafale77]

Wow that is impressive! I wonder why I did not see this on my G5. I wonder also how the BIOS/Motherboard knows what power supply it is connected to. The Lenovo can be explained by the addition of a connector but not on the HP. 

That’s one more thing I want to look at when I get back home. Maybe I don’t have the right model of HP 150W Power supply?

[deeveedee]

@rafale77 I figured out why you didn't observe a performance improvement with your G5 Mini / i9-9900 when upgrading from 90W to 150W adapter - for some reason, the G5 Mini / i9-9900 is not auto-adjusting PL2 when the adapter wattage increases.  That's all I know so far ... looking into this now.  Aside from the obvious difference of G4 vs G5, the other difference is that I reverted the G4 Mini BIOS to an old version.  I don't want to play with my G5 Mini / i9-9900 too much, so I may be limited in what I can determine.

 

EDIT: My 150W adapter is HP 609919-001

 

EDIT2: I bumped PL2 to 135W and am not seeing any improvement in i9-9900 performance.   There's something else going on that I don't understand.

 

GB5 CPU Benchmarks for HP EliteDesk 800 G5 Mini / i9-9900; PL1=65W, PL2=135W

Spoiler

 

Edited May 12, 2022 by deeveedee

[deeveedee]

I performed a few tests with my HP EliteDesk 800 G5 Mini / i9-9900 powered by a 150W adapter.  I was not able to observe any meaningful performance increase over testing with a 90W adapter.  I even tried reverting to BIOS version 02.03.00 (the first BIOS version that splits configuration of Intel ME and AMT.  With both BIOS versions 02.03.00 and 02.13.00, the G5 Mini PL1 / PL2 remained 65W / 90W, unlike the EliteDesk G4 Mini which auto-increased PL2 with the 150W adapter.  I tried manually increasing PL2 to as high as 180W and observed no appreciable performance increase as measured by GeekBench 5's CPU test.

 

If someone has ideas, I'm open to suggestions.  Based on tests so far, it appears that the G4 Mini auto-adjusts PL2 based on the adapter power, but the G5 Mini does not.  Note that the G4 and G5 Mini rigs that I'm testing are both 65W units.

Edited May 13, 2022 by deeveedee

[rafale77]

@deeveedee, I could suggest to install a tool like HWMonitor to look at the actual power sent to the CPU under load… as well as its temperature. It is possible that under GB5 the limitation is purely thermal and therefore our CPU is throttling down before it is able to reach the PL2 value. I have never seen my CPU go beyond 70W on the HP before while it hits 125W on the Lenovo. The G4 may behave differently because of the difference of CPU. It is also possible that the CPU microcode/Power Level table on the G5 is different from what they did on the G4. It seems to me that HP is tried to make some extra dough by maintaining 2 different BIOS on identical hardware and not allowing G4 to have the Coffee Lake Refresh microcodes…

 

With the cooling limitation of the HP Mini (both VRM and CPU), I was expecting the gains to be fairly limited but oddly HP lists models with the 95W TDP “K” CPUs as an option on their website. I feel in comparison that the Lenovos have much more reactive and robust cooling. The fan spins up instantly with temperature and the 65W version of the HSF contact with the VRMs.

 

Edited May 13, 2022 by rafale77

[deeveedee]

I upgraded my G4 Mini BIOS to the current (at the time of this writing) BIOS version 02.19.00 Rev.A and found that the G4 Mini still auto-increases PL2 to 122W when powered by a 150W adapter.  The G4 Mini's auto-adjustment of PL2 based on the adapter wattage was not eliminated in later BIOS versions, so the PL2 behavior difference between the G4 Mini and G5 Mini is still unexplained.

[theroadw]

Using Throttlestop in windows what are the limits showing?
In HP's Zbook G5 there is an (oem added) artificial power limit if using certain newer bios versions and a dual GPU configuration even if the DGPU is turned off.

This can be safely bypassed by Throttlestop or by custom SSDT.

Edited May 15, 2022 by theroadw
added image

[deeveedee]

@theroadw I will have to install Windows, as I'm running only macOS.  I'll report back when I follow your advice.  Thank you!

[rafale77]

I finally got to run some tests on my Lenovo and this is my findings:

First I compiled the zpherez's version of VoltageShift and then launch the info and I got this:

 

 

Interesting that my PL1 and PL2 are already set at 150W. So my only path forward is to reduce my thermals by undervolting and improving cooling.

 

Next, as you can see I am OC_Locked so I used ControlMsrE2.efi to attempt to unlock it.

To do so I ran it in the EFI Shell with the "interactive" argument and changed the search verbiage to "Overclock" and got this:

 

 

 

Ok so now I have two MSR values relatec to Overclocking but unfortunately when I enter either choice 1 or 2, I don't get any output and ControlMsrE2 exits.

I then tried to go read the MSR value using VoltageShift but it crashes when I input the MSR address...

 

@deeveedee have you tried manually changing MSR values? What is the address format?

Edited May 16, 2022 by rafale77

[deeveedee]

@rafale77 If by "manually changing MSR values," you mean using voltageshift to write to specific MSR addresses - then yes, that is how I'm changing PL1 and PL2.  To change PL1 and PL2, I write to MSR address 0x610 as follows:

 

voltageshift write 0x610 <value>

Where <value> is according to my notes here and here.

 

I tried changing PL1 and PL2 to 150w as is the case with your rig, but performance did not change on my G5 Mini, so something else is limiting performance on my G5 Mini.  My G4 Mini continues to show a 10% increase in GB5 multi-core CPU benchmark after powering by 150W adapter vs. 90W with no manual changes to MSR.  I have reported what I think is a G5 Mini BIOS bug to HP and have an open case.

Edited May 16, 2022 by deeveedee

[rafale77]

5 hours ago, deeveedee said:

@rafale77 If by "manually changing MSR values," you mean using voltageshift to write to specific MSR addresses - then yes, that is how I'm changing PL1 and PL2.  To change PL1 and PL2, I write to MSR address 0x610 as follows:

 

voltageshift write 0x610 <value>

Where <value> is according to my notes here and here.

 

I tried changing PL1 and PL2 to 150w as is the case with your rig, but performance did not change on my G5 Mini, so something else is limiting performance on my G5 Mini.  My G4 Mini continues to show a 10% increase in GB5 multi-core CPU benchmark after powering by 150W adapter vs. 90W with no manual changes to MSR.  I have reported what I think is a G5 Mini BIOS bug to HP and have an open case.

 

Thanks, It seems my problem from the screenshots I posted with ControlMsrE2 is that it is unable tor retrieve the IfrVarStore from my Lenovo BIOS. Strangely it means that this tool doesn't work on my machine but @Brumbaer original CFGLock.efi does for the CFG Lock but it doesn't allow me to change other hidden values.

 

So... I ended up brute-forcing with an alternate method using the RU.efi tool and changing the two overclocking options from the offset I got from ControlMsrE2 above. And... it worked! No more OC lock:

 

For testing, I am using CineBench23 and monitoring HWMonitor because that test stresses the CPU is long enough for me to observe the switch from PL2 to PL1.

A couple of observations:

1. Strangely as soon as I unlock the overclocking features, the PL and therefore the clocks appear to decrease: GB5 Multicore goes from 9500+ to 7100+ and so does the Cinebench23 scores down from 12000+ to 8000+

2. I also enabled overclocking features and toggling it doesn't appear to do much

 

I then had a brainfart of trying a 10850K from another machine which theoritically is not supported by this tiny build... but it worked. Now the K CPU reports the exact same PLs from the VoltageShift info but with a different base clock as expected but disabling the OC Lock this time didn't come with a performance penalty/PL decrease:

 

 

3. Just like on the HP G5, The PL values reported by VoltageShift don't appear to be used. I suspect that another set of PLs are being used by the BIOS. I am seeing the 10900 go from 120W to 65W after 28s while it is reporting that both should be 150W on both CPUs.

 

I ran a few benchmarks and indeed, undervolting has yielded some performance improvements. You can see my current voltage offsets above:

GB5 Multi-core has gone from ~9580 to ~9880 or about 3%

Cinebench23 which appears again to be more stressful has gone from 12407 to 13462, an increase of 8%

 

 

Looks like I will be sticking with the 10850K...

 

 

[rafale77]

I used ControlMsrE2.efi in interactive mode to look for the PL1 and PL2 settings and I see Platform PL1 enable and Platform PL2 enable flag which in both case are disabled on my machine.

With the address given by ControlMsrE2 I can easily enable them to see if it enables the PL2 values. It seems to me that the values displayed by VoltageShift are currently somehow overridden. Even the transition timimg between the 2 power limit settings seems (aka. tau) seems to be overriden by the BIOS.

 

I also undervolted my CPU further to -120mV and for the first time am getting Single Core GB5 scores >1400 and Cinebench Multi-Core >14,000. I need to test longer for stability.

I still can't figure out what the Overclocking Features hidden flag in the MSR does...

[WhenMusicAttacks]

On 5/13/2022 at 7:14 AM, rafale77 said:

@deeveedee, I could suggest to install a tool like HWMonitor to look at the actual power sent to the CPU under load… as well as its temperature. It is possible that under GB5 the limitation is purely thermal and therefore our CPU is throttling down before it is able to reach the PL2 value. I have never seen my CPU go beyond 70W on the HP before while it hits 125W on the Lenovo. The G4 may behave differently because of the difference of CPU. It is also possible that the CPU microcode/Power Level table on the G5 is different from what they did on the G4. It seems to me that HP is tried to make some extra dough by maintaining 2 different BIOS on identical hardware and not allowing G4 to have the Coffee Lake Refresh microcodes…

 

With the cooling limitation of the HP Mini (both VRM and CPU), I was expecting the gains to be fairly limited but oddly HP lists models with the 95W TDP “K” CPUs as an option on their website. I feel in comparison that the Lenovos have much more reactive and robust cooling. The fan spins up instantly with temperature and the 65W version of the HSF contact with the VRMs.

 

INTEL POWER GADGET

23 hours ago, rafale77 said:

I used ControlMsrE2.efi in interactive mode to look for the PL1 and PL2 settings and I see Platform PL1 enable and Platform PL2 enable flag which in both case are disabled on my machine.

With the address given by ControlMsrE2 I can easily enable them to see if it enables the PL2 values. It seems to me that the values displayed by VoltageShift are currently somehow overridden. Even the transition timimg between the 2 power limit settings seems (aka. tau) seems to be overriden by the BIOS.

 

I also undervolted my CPU further to -120mV and for the first time am getting Single Core GB5 scores >1400 and Cinebench Multi-Core >14,000. I need to test longer for stability.

I still can't figure out what the Overclocking Features hidden flag in the MSR does...

As i stated multiple times, undervolt was the answer to all your issues from the beginning. Not fiddlign with power adapters or power limits.

[rafale77]

2 minutes ago, WhenMusicAttacks said:

INTEL POWER GADGET

As i stated multiple times, undervolt was the answer to all your issues from the beginning. Not fiddlign with power adapters or power limits.

 

Sorry but I respectfully disagree. I don't think you fully understood what we are trying to do here. It isn't about resolving an issue but rather pursuing performance optimizations.

It all started from people reporting significant performance from Power supply adapter upgrades on the lenovo tinys and reproduced by @deeveedee on the HP G4 mini. Power supplies and Power limit can drive 7-15% performance gains on these, undervolting maybe another 3-8%, may be 10% if lucky with the silicon lottery. What I am currently seeing is that my unit under load, when it is limited at 65W under PL1 doesn't hit thermal limitations yet and the fans are not even at full speed. It appears that the manufacturer has set a hard 150W limit for the hardware but a softer PL2 at 120W which we are struggling to override. With these lower Power limits, the CPUs are performing 15-20% below stock for multicore loads which was the whole point of getting an i7/i9 CPU to begin with.

 

I can't speak too much for the HP because I felt it would hit thermal throttling anyway before raising these power limits could make a difference but the for my Lenovo at least, there is room to get closer to CPU stock performance. We are not overclocking here and though undervolting is definitely helpful, it is far from being the answer and I would like to see what the machine can do with higher power limits which... the manufacturers change according to the power brick feeding the motherboard and consequently are affecting the machine's performance.

 

 

[WhenMusicAttacks]

  

1 hour ago, rafale77 said:

 

Sorry but I respectfully disagree. I don't think you fully understood what we are trying to do here. It isn't about resolving an issue but rather pursuing performance optimizations.

 

 

 

Optimization is not the right term, you are looking for extremes, the optimum - best performance-per-watt - is probably even slightly lower than 65w.

 

You are looking at performance in a very narrow way - as is, a very short burst of cpu intensive load, followed by long pauses, where all the components can cool down.
If you really care about CPU performance for some intensive rendering task, you want a machine that can function under load for prolonged times. That's not going to be achieved with unlocked power limits, as you will find another - potentially more dangerous - limit in the power delivery or cooling system.

Efficiency goes out of the window at higher PL. At 140w you will overheat so much faster than 70w, for a gain of 20% in CPU performance, making it run WORST in long runs as you will exhaust the "cooling buffer" that hte machine stores while idle.

You can read and see graphs here to understand the issue better:

https://www.pizzaundervolt.com/choosing-a-computer-for-audio-production-laptop-2/

You can try yourself runnign a 10 minute r23 loop with different PL and undervolting: lower power limit with good undervolt will give you a better 10 minute run than unlocked power without undervolt, not to mention a 30 min run.

So your gain by unlocking power limit are extremly costly in terms of power, reliability and lifespan of such tiny devices, and limited to slightly higher number on short benchmark runs.

While gains by undervolting are consistent and safe.

Remember undervolt is similar to overclock, as you can get more of it the lower the temps and internal resistance. So you might be able to have -110 at pl65, but only -70 at pl 125.

I fully understand the thirll of unlocking a limit imposed by the machine creator as the reason to do all of this.

But don't lie about this making any sense outside of showing off. These are thermally and electrically constrained machines that will cok themselves even at stock limits if used for serious render tasks. They are made for web browsing and office tasks.

Edited May 17, 2022 by WhenMusicAttacks