Optimal CPU Tuning settings for i9 MBP to stop VRM throttling / Explanation of Apple's Engineering F

[Petetastic]

So after the first post in this thread, can we finally admit that Apple--not Intel, not Adobe, Apple--screwed up, and can we hold Apple accountable for this? Can we agree as a community that they don't get a pass on this? That they actually need to fess up to this mistake and fix it timely and without ********?

By the way, I placed a 2.2ghz 2018 MBP. I'm not a Mac hater. I just believe in corporate responsibility.

Edit: Though tbh I'm having second thoughts now given that Apple never admits mistakes ever.

 

[winterny]

Since these settings don't survive reboots , will they survive wake from sleep?

They should survive a sleep.

 

[Elektrofone]

@winterny If Apple were to fix this issue with a software update do you think the i7 would fair better than the i9? Currently have an i9 and trying to figure out if I should return and swap for i7.

 

[winterny]

@winterny If Apple were to fix this issue with a software update do you think the i7 would fair better than the i9? Currently have an i9 and trying to figure out if I should return and swap for i7.

I don't have a 2.6 i7 nor do I know anyone who does, so I am purely speculating, but ...

It depends on the workload.

For sustained workloads, I think the 2.6 i7 will probably be faster (though I'm still hoping someone will post a screenshot of some back-to-back cinebench tests after making this adjustment.

For burst workloads, especially tasks which normally complete in under 10 seconds, the i9 will probably still win.
[doublepost=1532399872][/doublepost]

Failure

https://www.reddit.com/r/macbookpro...s_for_i9_mbp_to_stop/?st=JJZ1REWT&sh=7513fe6a

Yep.

It maxed out the title length, and edits for the title don't seem to be allowed ... oops

In any way, F means Failure, so I think the title still works
[doublepost=1532399925][/doublepost]

The time values seem so random , how on earth did you calculate those?

I think intel explains it best:

PKG_PWR_LIM_2_TIME: x = PKG_PWR_LIM_2_TIME[55:54] y = PKG_PWR_LIM_2_TIME[53:49] The timing interval window is Floating Point number given by 1.x * power(2,y). The unit of measurement is defined in PACKAGE_POWER_SKU_UNIT_MSR[TIME_UNIT]. The maximal time window is bounded by PACKAGE_POWER_SKU_MSR[PKG_MAX_WIN]. The minimum time window is 1 unit of measurement (as defined above).

Tongue firmly in cheek.

 

[ESA]

I don't have a 2.6 i7 nor do I know anyone who does, so I am purely speculating, but ...

It depends on the workload.

For sustained workloads, I think the 2.6 i7 will probably be faster (though I'm still hoping someone will post a screenshot of some back-to-back cinebench tests after making this adjustment.

For burst workloads, especially tasks which normally complete in under 10 seconds, the i9 will probably still win.
[doublepost=1532399872][/doublepost]
Yep.

It maxed out the title length, and edits for the title don't seem to be allowed ... oops

In any way, F means Failure, so I think the title still works
[doublepost=1532399925][/doublepost]
I think intel explains it best:

PKG_PWR_LIM_2_TIME: x = PKG_PWR_LIM_2_TIME[55:54] y = PKG_PWR_LIM_2_TIME[53:49] The timing interval window is Floating Point number given by 1.x * power(2,y). The unit of measurement is defined in PACKAGE_POWER_SKU_UNIT_MSR[TIME_UNIT]. The maximal time window is bounded by PACKAGE_POWER_SKU_MSR[PKG_MAX_WIN]. The minimum time window is 1 unit of measurement (as defined above).

Tongue firmly in cheek.

Could you please make a YouTube video where you slowly do all the steps? So we can´t fail. Then we all can send you some bucks on paypal.

 

[Chancha]

This thread needs to be pinned, if not making the macrumors front page. I was quite split as to whether or not to consider the i9 before seeing this. The confirmation of VRM, a hardware chip on the mother board being the culprit, and that it is prone to failure resulting a possible shorter life span of the MBP is enough to help a decision. In fact I am waiting to see if insufficient VRM cooling is present on all 2018 models, in which case I will avoid buying any of them.

 

[leman]

I think intel explains it best:

PKG_PWR_LIM_2_TIME: x = PKG_PWR_LIM_2_TIME[55:54] y = PKG_PWR_LIM_2_TIME[53:49] The timing interval window is Floating Point number given by 1.x * power(2,y). The unit of measurement is defined in PACKAGE_POWER_SKU_UNIT_MSR[TIME_UNIT]. The maximal time window is bounded by PACKAGE_POWER_SKU_MSR[PKG_MAX_WIN]. The minimum time window is 1 unit of measurement (as defined above).

Can you point me to the relevant document and/or header? All I find in Intel docs is partial bit layout of the register, but nothing concrete...
[doublepost=1532413818][/doublepost]

So after the first post in this thread, can we finally admit that Apple--not Intel, not Adobe, Apple--screwed up, and can we hold Apple accountable for this?

Apple has misconfigured the CPU power management, there is no doubt about it. Intel on the other hand has released a CPU that can and will draw much more power than it spec suggests.

Case in point: https://www.notebookcheck.net/Some-...nning-any-faster-than-a-Core-i7.317268.0.html This shows that the only laptops capable of truly supporting the potential of the i9 are 3.5-5kg gaming behemoths. IMO, this is still completely misleading. The i9 is simply a marketing move my Intel, and not the one that benefits the user. Of course everyone wants to have that (the specs are incredible after all), but you essentially need a desktop cooling solution to accommodate it. For us users, it would be much better if Intel released a top-tier i7 CPU (like they always did prior to now), with more reasonable clocks around 2.7/4.4 or so, and marketed the i9 as a 60W part for enthusiast laptops. Then nobody would expect to use it in the thin and light category.

 

[kotlos]

Apple has misconfigured the CPU power management, there is no doubt about it. Intel on the other hand has released a CPU that can and will draw much more power than it spec suggests.

Case in point: https://www.notebookcheck.net/Some-...nning-any-faster-than-a-Core-i7.317268.0.html This shows that the only laptops capable of truly supporting the potential of the i9 are 3.5-5kg gaming behemoths. IMO, this is still completely misleading. The i9 is simply a marketing move my Intel, and not the one that benefits the user. Of course everyone wants to have that (the specs are incredible after all), but you essentially need a desktop cooling solution to accommodate it. For us users, it would be much better if Intel released a top-tier i7 CPU (like they always did prior to now), with more reasonable clocks around 2.7/4.4 or so, and marketed the i9 as a 60W part for enthusiast laptops. Then nobody would expect to use it in the thin and light category.

i9 was no marketing move by Intel. It was meant to be used in "3.5-5kg gaming behemoths" that many people want to use. It is also unlocked which means it is allowed to use much more energy than the TDP, as every other unlocked CPU does.

It was Apple's decision to stick it into a case that clearly doesn't support it. From all the laptops that have the i9, the MBP performs the worst. And on top of it, they didn't go through the trouble of not only redesigning the cooler, but not even optimizing the power settings.

Don't forget that intel also released an i7, which by the way also throttles in the current MBP.

So who did really pull a marketing move? Apple clearly messed this up and we fell for it. Lets hope they fix it somehow.

 

[Nozuka]

I feel like MR should put these findings from the OP on the frontpage, so that more people understand what and why it is happening.

 

[ESA]

If I understand this right, if Volta will make an update with undervolt-support. Then we can do all the necessary things in the first post right? Except the thing that you get full speed the first seconds.

 

[M.Rizk]

If I understand this right, if Volta will make an update with undervolt-support. Then we can do all the necessary things in the first post right? Except the thing that you get full speed the first seconds.

They can’t do undervolting. Apple has it locked since the 2016 MBP. They both do power limiting and that’s it.

 

[ESA]

They can’t do undervolting. Apple has it locked since the 2016 MBP. They both do power limiting and that’s it.

Ok, then I hope there will be a "video for dummies" version of this first post haha

Is there any comparison yet with the 2.6 model that use the same configurations?

 

[maflynn]

They can’t do undervolting. Apple has it locked since the 2016 MBP. They both do power limiting and that’s it.

Is it Apple or is the CPU locked?

 

[M.Rizk]

Is it Apple or is the CPU locked?

I think it’s Apple. Undervolting has to do with the VRM not the CPU itself as far as I know.

 

[drvelocity]

I haven't seen any other results in this thread from using these settings other than mine which tanked my i9 into the 850 range for Cinebench, so about a 10% loss. Have we confirmed that this is actually working for anyone else?

 

[winterny]

I haven't seen any other results in this thread from using these settings other than mine which tanked my i9 into the 850 range for Cinebench, so about a 10% loss. Have we confirmed that this is actually working for anyone else?

As I was working through this, I validated my work by having a friend test on his machine as well, and you can see him posting about it here on reddit:
https://www.reddit.com/r/macbookpro...u_tuning_settings_for_i9_mbp_to_stop/e2wp0jl/

Someone who I don't know, also posted their results on reddit:
https://www.reddit.com/r/macbookpro...u_tuning_settings_for_i9_mbp_to_stop/e2wt504/

Is there any chance that you have some other app open which is activating your GPU while you are doing the Cinebench test? (You can see if it is activated by looking at the Energy tap of Activity Monitor)

There have also been a couple of posts, like yours, which are not getting good results. I don't work for Apple, and obviously have access only to a limited number of machines for testing. Based on the posts of people having worse performance, it seems that with manufacturing variability, some machines are made worse than others, and would require settings which limit power draw even beyond what my settings do. Your machine seems to be one of them. Post a screen capture of Intel Power Gadget while running Cinebench back-to-back a few times, both "at stock", and with the adjusted MSR settings.

 

[Nozuka]

I feel like MR should put these findings from the OP on the frontpage, so that more people understand what and why it is happening.

Nevermind. I guess news is out thanks to reddit and 9to5mac ...

https://9to5mac.com/2018/07/24/macbook-pro-throttling-fix/

 

[The Mercurian]

As I was working through this, I validated my work by having a friend test on his machine as well, and you can see him posting about it here on reddit:
https://www.reddit.com/r/macbookpro...u_tuning_settings_for_i9_mbp_to_stop/e2wp0jl/

Someone who I don't know, also posted their results on reddit:
https://www.reddit.com/r/macbookpro...u_tuning_settings_for_i9_mbp_to_stop/e2wt504/

Is there any chance that you have some other app open which is activating your GPU while you are doing the Cinebench test? (You can see if it is activated by looking at the Energy tap of Activity Monitor)

There have also been a couple of posts, like yours, which are not getting good results. I don't work for Apple, and obviously have access only to a limited number of machines for testing. Based on the posts of people having worse performance, it seems that with manufacturing variability, some machines are made worse than others, and would require settings which limit power draw even beyond what my settings do. Your machine seems to be one of them. Post a screen capture of Intel Power Gadget while running Cinebench back-to-back a few times, both "at stock", and with the adjusted MSR settings.

The silicon lottery in action ?

 

[winterny]

The silicon lottery in action ?

There are always going to be silicon lotteries.

In this case, you are hoping for three different variables to be in your favor:
1) Good VRM
2) Good thermal paste job
3) Good Watt/HZ ratio on the CPU.

 

[drvelocity]

Indeed I confirmed with gfxcardstatus that the system was in integrated GPU mode before opening cinebench.

With OP's settings I still get a lot of throttling, whereas with Volta set to limit power to 45W (turbo boost still active) I get a solid 3+ Ghz. Graph below shows both tests, the left half is with OP's settings and the right half with Volta active. Everything is back to back so other than activating Volta halfway through all other variables should be relatively consistent.

 

[MacRS4]

I don't think the settings are surviving sleeps from the look of it, unless I'm misunderstanding.

Before Sleep
./voltageshift read 0x610
RDMSR 610 returns value 0x4283e8001b8188

After Sleep
RDMSR 610 returns value 0x4283e800dd8320


Am I doing it wrong. Wouldn't surprise me. I've been wrong all day.

 

[prh8]

There are always going to be silicon lotteries.

In this case, you are hoping for three different variables to be in your favor:
1) Good VRM
2) Good thermal paste job
3) Good Watt/HZ ratio on the CPU.

I just got a i9 and I'm wondering, what would be the best way to diagnose whether I've got a good one or a bad one? So far, paying attention to Power Gadget while I do my normal workloads, it seems pretty good, but I don't know enough to determine for sure. FWIW I'm not too concerned about graphics intensive loads, I'm a software dev. Only graphics needs for me are powering 2 4k monitors

 

[kotlos]

Indeed I confirmed with gfxcardstatus that the system was in integrated GPU mode before opening cinebench

But when cinebench runs it forces the switch to the dGPU correct? @winterny The same with you right?

I tested again my setup at a low temperature room (<70F) on a cold metal flat surface, and still get throttling with your settings. When I reduce the initial boost to 7 seconds I get 1050 but only on the first run after that it drops.

So either I have a bad paste job, or/and more inefficient CPU/VRM.

Now the question is: Should I start modding my laptop with repasting & thermal pads, or should I go exchange it for another model and test my luck?

 

[skywalk423]

If you still need someone to post what RDMSR 610 returns on a new i7 2.6Ghz machine (just got mine):

RDMSR 610 returns value 0x4383e800dd8320

Second digit is a 3 instead of a 2. If I wanted to try your tunings, I assume I would leave the second digit as-is?

EDIT: Rebooted and retested the same read... it now matches your string (0x428). Weird.

 

[Stux]

So, I've got some extra datums to add to the thread.

I've been testing with Prime95 SmallFFTs, 12 threads.

I found that with my external LG5K TB3 display unplugged I had to use a 47.5W limit to get a stable clock speed with no throttling.

ie

0x4283E8001B817C

Meanwhile, when I plugged in the LG5K, I found I further had to reduce the avg to 43.5W (ie 0x15C in order to have a stable steady state.

And yes... I stepped down in 0.125W increments.

With the 43.5W setting (0x4283E8001B815C), I find that if I unplug the LG5K display while running the torture test then the CPU will actually start to cool!

It seems the use of the TB3 display adds an extra 4W to the package... which is about 10% of the power budget.... and it also translates into about 10% loss of Mhz. from 2.7 -> 2.5

Sigh. Good thing there are 50% more cores still

But yeah... beginning to seem a lot like an i7.
[doublepost=1532451127][/doublepost]

Hello all,
Edited to add:

If you want timing values shorter than 7 seconds ....
6 Seconds: 99
5 Seconds: 59
4 Seconds: 19
3.5 Seconds: D7
2.5 Seconds: 57
2 Seconds: 17
1.5 Seconds: 95
1 Second: 15

I'm playing with the PL2. I'm finding by setting it lower, I can sustain high boost clocks without overheating the CPU before it hits steady state... but 10 seconds is not a long enough window... And I can't work out how to derive the values

Any chance you can post the values from 10..20?

28 is too long

FWIW, this is what i'm using right now

0x428235001B815C

70.625W Boost, 43.5W steady, 8s window.


And an example of boosting at various cpu utilizations... then going to steady state. There's no VRM or thermal throttling. Its worth pointing out that p95 uses AVX instructions and so clock speeds get reduced by a few buckets.