Need help: Odd CPU scaling with new Skylake i7

[controlflow]

I'm using a new Dell XPS 15 with a Skylake i7-6700HQ and I am noticing some strange issues in terms of power/CPU management.

I set my power mode to balanced and while on AC power, I ran CPU-Z bench. I got 1452/6716 (Single threaded and multi threaded respectively). I noticed the CPU hitting its max rated frequency (3.5 Ghz) and everything was as expected.

I then disconnected AC power and re-ran the benchmark at which point I got 1079/5311 which is reasonable since the CPU clocks were more conservative while on AC power. I noticed the CPU stayed around 2.5 GHz on AC power while under load.

I then reconnected the AC power expecting the CPU to be able to go back to its full frequency and achieve performance that is about the same as the first result, instead I find that the CPU just gets stuck at 1.69 Ghz or lower. It just cannot go above 1.69 Ghz and I end up scoring about 805/3700.

I also find the following messages in the event log: "The speed of processor 6 in group 0 is being limited by system firmware. The processor has been in this reduced performance state for 71 seconds since the last report."

At this point you cannot get the full AC performance of the system unless you do a full reboot or possibly sleep/wave the system. To me this sounds like some kind of BIOS/Firmware or power management driver issue. Any ideas? FYI there is no throttling going on at any time.

Disabling speed step in BIOS allows the CPU to always maintain full performance but of course that destroys battery life and generates enormous heat.

 

[ShintaiDK]

Try another benchmark. CPU-Z is bugged. Specially if you run it multiple times.

In terms of the system message. Its Dell setting the limitations. It can only be changed by BIOS update, unless this is how Dell wants it to work.

 

[inf64]

CPUz bench is beyond broken. Try running geekbench under same circumstances and see what happens.

 

[controlflow]

Thanks for responding guys. I did try geekbench and I also tried to modify my windows power plan to keep the CPU at 100% at all times.

Sure enough, in both cases the CPU just can't break 1.69 GHz once you unplug and re-plug the AC connector. It really does seem like a BIOS issue. I first noticed that something was wrong with my CPU performance when I was compiling some code and it was taking much longer than it had earlier, that's what caused me to investigate. I hope Dell knows about this and patches it soon.

Anyone have any ideas on how to go about reporting this? Trying to talk to their tech support people has been a waste of time, they simply don't understand what I'm telling them. They just tell me to return the laptop and get another one, obviously that won't help...

 

[Burpo]

Had some luck with Throttlestop 8 Beta 5 on Skylake @ overclock.net (don't know how it works with laptops).

http://www.overclock.net/t/1570313/skylake-overclocking-guide-with-statistics/3420#post_24559503

 

[unclewebb]

ThrottleStop 8.00 beta 5
https://www.sendspace.com/file/p1q40a

ThrottleStop is still a work in progress but it should work OK with a Skylake laptop. Post some pics if you have any problems.

Make sure BD PROCHOT is not checked, check off Set Multiplier and set that to the maximum value and click on the Turn On button to take ThrottleStop out of Monitoring mode.

 

[controlflow]

Thanks for the useful posts on ThrottleStop. This tool is pretty handy and it seems it lets me control C1E, SpeedStep, and Multipliers on the fly. Even after the CPU gets stuck at 1.69 GHz, I can control the multiplier and bring it back to full performance, however it doesn't scale anymore, CPU frequency doesn't scale with load. Anyway to get it to do that? Disabling BD PROCHOT itself isn't useful in stopping the firmware from capping the CPU at 1.69 Ghz. I'm going to spend some more time messing with these settings but it really looks like Dell is just going to have to fix this BIOS.

I must say other than that, this new laptop is quite impressive. Love the aluminum and carbon fiber construction, the smooth glass trackpad, the 100% Adobe RGB IGZO 4k display, GTX 960M GPU, Skylake CPU. All nicely done but it is kind of a deal breaker if Dell can't get the BIOS done correctly.

 

[unclewebb]

...however it doesn't scale anymore, CPU frequency doesn't scale with load.

Old school thinking was that you had to reduce the MHz to save power when idle but modern CPUs use various C States to reduce power consumption when idle. When a CPU core enters C7, it is disconnected from the voltage rail and its internal clock speed is zero. Using monitoring software that only shows MHz can be misleading when a CPU is idle.

Some smart guys at Berkeley discovered that slowing a CPU down doesn't save you as much energy as you would think it does. A slow CPU is an inefficient CPU and instead of saving energy, it can consume more energy as it plods along for a longer period of time compared to processing what needs to be processed quickly and then getting back into one of the low power C States.

http://www.cs.berkeley.edu/~krioukov/realityCheck.pdf

What does the ThrottleStop C States data show? Some of the new Skylake U CPUs use deeper C States like C8, C9 and C10.

Here is a mobile 4700MQ and the 4 cores are averaging 99% of the time in C7 where MHz and voltage are completely meaningless.

High or low makes no difference when a core is in C7 so forget about idle MHz and concentrate on improving C State residency time.

 

[controlflow]

The reported package and IA core power is lower on idle when I set the multiplier to 8 as opposed to 35, even though at 35 the CPU is mostly in C7.I think lowering the CPU frequency on idle still does help with idle battery life. Otherwise wouldn't technologies like SpeedStep be obsolete?

 

[unclewebb]

The reported package and IA core power data that all software reports is not based on actual power consumption. This number is calculated within the CPU and its only purpose is for management of the Intel Turbo Boost feature. At idle, it is a meaningless number.

Have a look at the percentage of time my CPU is spending in C7 and compare that to your CPU. Stop worrying about MHz or reported power consumption. You will save far more power and extend battery run time by finding ways to get your CPU to spend more time in C7.

 

[controlflow]

At idle my CPU spends a pretty good amount of time in C7 as well.

I'm going to just set the CPU to multiplier of 35 and make sure BD PROCHOT is disabled and use the system for a few days and see what happens to the battery life.

Any way to get all of this to run on each reboot of the OS? Maybe add ThrottleStop to startup? How are you guys doing it?

 

[controlflow]

BTW unclewebb thanks for posting the interesting study on power optimization. I was pretty convinced that you had to bring down your voltage as much as possible and bring the CPU clock all the way down during idle to maximize battery life but I guess this actually seems to be the better way to do it:

"The most immediate implication of these two observations is that they explain anecdotal evidence for a model of computation known as “race to sleep.” This idea holds that the most energy-efficient way of scheduling a computation is to put all hardware into the highest-performance state and race to completion as quickly as possible. Once finished, the hardware should drop to very low power modes – potentially ACPI sleep states or even powered off. The explanation for this fact draws on both the constant power and the linear power-performance tradeoff."


Do you have any more information about the C7 or lower power states? What kind of power consumption does a core actually have in C7?

 

[unclewebb]

A couple of interesting things I noticed is that Skylake CPUs support new C States like C8. It looks like Dell has disabled C8 in the bios. They also seem to be limiting your package C States to only C3. It's not good when Intel pours millions of dollars into R&D so their CPUs will consume less power and then manufacturers do not bother using these new features. They have also locked the Package C State register in the bios so users can not use software to adjust this and access the deeper power saving C States.

If you want to find out more about this, can you run the Dump tool by Dufus. It will read the contents of all of your CPU registers and it will create a file called Dump.txt with this info. You could either send it to me in a message, post it here or Copy and Past the data to www.pastebin.com and then post or send me a link to this info. I am just curious to see what Dell is doing.

Dump
https://drive.google.com/open?id=0B0dpSo9k93jDX1Fpa1FpTmo1Qk0

In core C6 / C7 / C8; the core itself is disconnected from the voltage rail so it is basically consuming as little power as possible. There is a small amount of RAM on the chip that is used to save the state of the core before it turns itself off. When the core is needed again, it reloads its previous state from this on chip RAM. This RAM needs a small amount of power but the core itself doesn't need any. The deeper C States turn off more and more of the CPU when it has nothing to do.

The datasheet shows a maximum power consumption of 11 Watts when in Package C7 and a maximum of 2 Watts in Package C8.

http://www.intel.com/content/www/us...ktop-6th-gen-core-family-datasheet-vol-1.html

C8 monitoring should be added to ThrottleStop in the near future. You could contact Dell and ask them why they are not using this new C State. If C8 was available, battery run time should be significantly improved when doing light duty stuff.