Results and questions

[temp123]

First of all, I want to thank the author for this amazing utility!
I didn't have so much fun since the early 2000's when I was frying my memory at 4V with 2-2-2-5 latencies
I wish there was a comprehensive guide explaining all the parameters in details but a bit of googling solves the problem.
Here's my situation:
I have i7-9750H on Eluktronics Mech-15G2Rx and I'm quite happy with it.
But I'm not satisfied with the battery life so I've decided to thinker with it a bit.
Turns out that pumping up is much easier than toning down.
The processor itself seems to be quite good and I was able to undervolt it quite a bit.
You can see the results and settings on the screenshot.
What surprised me is how hot is the cache memory and it's effect on the TDP since, for example, reducing the voltage for cache from -140 to -190 reduced TDP from 68W to 60W while reducing the core from -250 to -300 (not stable) had almost no effect on it.
But I'm not working on anything very computationally demanding so this is not such a big deal for me, and I would keep processor's TDP much lower for gaming since 1660ti is quite power hungry as well and it has a much bigger impact on the performance. When you have a laptop you have to make a compromise.
The main goal is to make a battery saver profile as efficient as possible.
First, I've replaced my Windows 10 Home edition (that comes preinstalled) with Win10 LTSC 2021 to reduce the amount of unnecessary MS crap to a minimum.
Then I tinkered with the services a bit more and I would say I did a good job.
However, something in the Speed Shift implementation seems to make a difference in the C0% when processor runs at max (0) and reduced (200, see screenshots)
At 0, C0% is at 0.1-0.2%, but at 200 it goes up to 08-09% and if fluctuates quite a lot. All the parameters are exactly the same, except the Speed Shift.
Why is this happening?
Try as I might I was not able to reduce this.
Also, on the Windows Home edition the idle TPD would go as low as 0.4-0.5W, but here it never goes bellow 1.6W.
Looking forward to your advice!

Cheers!

 

[DarkDreams]

C0% is relative (as the % indicates). So if the cpu runs at reduced clocks (800Mhz) from speed shift, it takes a larger percentage of the remaining clock cycles to do the same background work than when it runs at full 4Ghz+.

 

[unclewebb]

Good work.

An R20 score of 3183 puts you in the top 10 on HWBot.

Venky`s Cinebench - R20 score: 3387 cb with a Core i7 9750H

The Core i7 9750H @ 4490MHzscores getScoreFormatted in the Cinebench - R20 benchmark. Venkyranks #915 worldwide and #1 in the hardware class. Find out more at HWBOT.

hwbot.org

I cannot ever remember seeing a 9750H that was 100% stable at a cache offset undervolt of -190 mV. Does the TS Bench 960M test crash or report any errors?

reducing the core from -250 to -300 (not stable) had almost no effect on it.

There is always a certain point where using a really big negative number for the core offset does not accomplish anything. I have never done any testing but somewhere around -75 mV to -100 mV more than the cache seems to be the practical max. If you use -150 mV for the cache then about -225 mV to -250 mV for the core is usually the point where if you go further, there will be no further advantage. These are only voltage requests. The CPU will just ignore some ridiculous requests. I have seen some have the core set to an offset of -1000 mV. If that was actually being used the CPU would instantly crash.

The C0% that ThrottleStop reports is not the same as CPU load. C0% measures what percentage of time a CPU core is in the active state actively working on a task. A fast CPU can process background tasks much faster compared to a slow CPU. When you use a Speed Shift EPP setting that slows the CPU down to 800 MHz, the CPU will need to spend more than 5.5 times as much time in the C0 state to complete a background task compared to running the CPU at almost 4500 MHz like you are doing. When Windows is lean without a bunch of bloatware, having a fast CPU is not a bad thing. My laptop is plugged in 99% of the time so I always run it at full speed.

When comparing different Windows versions, did you check to see if they are both using the same low power C states? One slightly different driver version can accidentally disable one of the low power package C states. Your screenshot shows that the CPU is being blocked from using anything greater than Package C3. That might be why there is a difference.

Keep in mind that the power consumption data reported by Intel processors is not actual measured power consumption. It is only an estimation. I do not know how accurate this number is after you start playing around with the voltages. I find watching the CPU temperature is a good alternate method to confirm if your undervolt adjustments are working as they should.

The main goal is to make a battery saver profile as efficient as possible.

I rarely use a laptop on battery power so I do not have any suggestions. I have never had access to a Speed Shift capable laptop. I am still living in the 4th Gen era, pre Speed Shift. I know that using a large EPP value will slow the CPU down when lightly loaded but I do not know how much power during general use that will actually save. A slow CPU is an inefficient CPU so slowing a CPU down does not always save as much power as some people think it does. A fast CPU gets tasks done quickly. This allows cores to spend a bigger percentage of time in one of the low power C states like C7. What works best, fast vs slow, will depend on what sort of load you normally run while on battery power. Best not to assume anything without doing some testing.

I always set IccMax to the max for both the core and the cache just to avoid any throttling.

I want to thank the author for this amazing utility!

You are welcome. I like using ThrottleStop too. It is like a Swiss army knife for a wide variety of Intel CPUs. ThrottleStop 9.6 still works on the older Core 2 Duo based CPUs. I quit writing documentation years ago. I found that many people that most need documentation never bother reading the manual. I prefer hands on learning. Randomly checking and unchecking various ThrottleStop boxes and doing some Google searching is the best way to learn ThrottleStop.

 

[temp123]

OK, first the good news.
Mentioning that it could be a bad driver made me do a bit of searching and it turns out that many people have a problem like mine. Because of their Realtek LAN card. It never occurred to me that M$ generic driver might not be enough! How smart am I?
Anyway, as soon as I've installed the proper driver the problem was solved (screenshot).
Cores now go directly to C7. Whether they can go to other states I don't know but this does the job.

I cannot ever remember seeing a 9750H that was 100% stable at a cache offset undervolt of -190 mV. Does the TS Bench 960M test crash or report any errors?

There's a ton of errors and this is only for benchmarking. I was just curious to see how much I can push it and how much of a difference in TDP does it make.
Surprisingly enough, this benchmarking showed me that 100% CPU utilization can mean different things:
1) TSBench consumes around 5-7W less than Cinebench R20 even though they are both at "100%"
2) As such you might get different warnings when you run one or the other (one can give you no warnings but the other will and vice versa)
3) For long term stability nothing beats Prime95 but as a short test Intel Processor Diagnostic Tool does a very good job. Even if you pass both Cinebench and TSBench (960M) it will easily crash on Intel Processor Diagnostic Tool, especially on the graphic tests.

There is always a certain point where using a really big negative number for the core offset does not accomplish anything. I have never done any testing but somewhere around -75 mV to -100 mV more than the cache seems to be the practical max. If you use -150 mV for the cache then about -225 mV to -250 mV for the core is usually the point where if you go further, there will be no further advantage. These are only voltage requests. The CPU will just ignore some ridiculous requests. I have seen some have the core set to an offset of -1000 mV. If that was actually being used the CPU would instantly crash.

Now that you mention it I see that processor doesn't pay attention to core settings at all. The default voltage is 1.15-1,17V and it decreases only in the amount set by the cache parameter. If it was paying attention to core settings, the voltage (when testing) would be <0.9V but it's actually 0.98V. Does iGPU have a separate volt rail as well? If everything goes on the same rail how is it possible to change the settings separately for each component?

The C0% that ThrottleStop reports is not the same as CPU load. C0% measures what percentage of time a CPU core is in the active state actively working on a task. A fast CPU can process background tasks much faster compared to a slow CPU. When you use a Speed Shift EPP setting that slows the CPU down to 800 MHz, the CPU will need to spend more than 5.5 times as much time in the C0 state to complete a background task compared to running the CPU at almost 4500 MHz like you are doing.

I guess I should have thought of that myself

I always set IccMax to the max for both the core and the cache just to avoid any throttling.

That's an easy thing to do when you're on desktop but on laptop everything counts

An R20 score of 3183 puts you in the top 10 on HWBot.

Now that I know I might bother to make an account to register it. But there's only 39 people on the list so being in Top 10 is not that hard

Thx for the suggestion and feel free to add any other advice that you might have.

Cheers!

 

[unclewebb]

Realtek LAN card

Good work doing some detective work. Nice to see the package power back down to 0.5W. Laptops running Windows have traditionally had a bad reputation for battery run time compared to similar Apple laptops. Maybe Microsoft should have put a little more effort into the default generic drivers that it installs.

Package C7 is the important one. Some of the other deeper package C states are only used on the low power U series CPUs and often times only when the CPU is in sleep or connected standby mode.

it will easily crash on Intel Processor Diagnostic Tool

Good to know. All stress testing tools work different parts of the CPU. I like the TS Bench as it is a quick test and it does not use any AVX instructions like the majority of stress tests use these days. A stable computer should be able to run any test you throw at it without crashing or reporting any errors.

TS Bench consumes around 5-7W less than Cinebench R20 even though they are both at "100%"

That helps confirm my theory that different testing tools are not all using the CPU cores in the same way. A CPU running an AVX instruction consumes more power compared to the same CPU not running an AVX instruction.

doesn't pay attention to core settings at all

Using different offset voltages for the core and the cache seems to be a trick that lowers voltages when the CPU is running AVX instructions. The VID voltage reported by monitoring software might not accurately report what actual voltage is going to the CPU. Cinebench has proven to be a good test when running different offset voltages. Many laptops with 8th and 9th Gen CPUs report an advantage when doing this. The Intel GPU and the iGPU Unslice are on a different voltage rail compared to the core and cache.

Setting IccMax really high does not tell the CPU to start flowing current like a wild river. High IccMax just prevents any random throttling due to the IccMax current limit being set too low. If you want to limit power consumption, reduce the turbo power limits. This will cause fewer problems than IccMax being set low.

being in Top 10 is not that hard

Some of the guys in the top 10 are the type of people that are willing to dedicate their life to finding a way to get a few more Cinebench points. Being in the top 10 is better than you realize. The guys at the top are using every trick in the book like disabling background tasks and running Cinebench at a higher Windows priority. (hint hint!)

 

[temp123]

Some of the guys in the top 10 are the type of people that are willing to dedicate their life to finding a way to get a few more Cinebench points. Being in the top 10 is better than you realize. The guys at the top are using every trick in the book like disabling background tasks and running Cinebench at a higher Windows priority. (hint hint!)

I know that overclocking can be quite competitive.
Used to do that myself as well 20 years ago. But everything was much simpler and straightforward back then, there weren't so many parts integrated in the processor itself so you didn't have to bother with all of the settings like with this one. I had to relearn a bunch of things to do this properly.

Setting IccMax really high does not tell the CPU to start flowing current like a wild river. High IccMax just prevents any random throttling due to the IccMax current limit being set too low. If you want to limit power consumption, reduce the turbo power limits. This will cause fewer problems than IccMax being set low.

This is a good thing to know, thx for the info.

One more thing.
As you can see in the first screenshot, the processor was able to go to its lowest frequency (800Mhz( without a problem.
But now it's goes only to 1.2Ghz (like in the last screenshot) even if I set Speed Shift EPP to 255.
No idea why.
Any suggestions?

 

[unclewebb]

But now it's goes only to 1.2 GHz

That is interesting. Speed Shift allows the CPU hardware to decide what speed it should run at. After enabling the C7 package C state, your CPU has decided that it is no longer necessary to slow down to 800 MHz so it runs at a minimum of 1200 MHz instead.

Everyone thinks that running the CPU as slow as possible is best but at the low end you can reach a point where CPU efficiency decreases. When all of the C states are properly enabled, slowing the CPU down to a crawl is not necessary. I guess you will have to trust Intel on their decision to use 1200 MHz instead of 800 MHz for the minimum when package C7 is enabled.

If you want to force your CPU to run at 800 MHz for testing purposes, try setting Speed Shift Min and Speed Shift Max to 8 in the TPL window.

 

[temp123]

I wanted to see what's the lowest possible TDP when using Movies & TV app (which is by far best optimized player for Windows even though it's as simple as it gets with no advanced options) so I can squeeze the maximum amount of time on battery while traveling but, as you've said, looks like Intel did some smart programing when it comes to power consumption. There's no difference in TDP at 800Mhz and 1200Mhz. Surprisingly enough even 800Mhz is enough for the app to work without a problem, hardware decoding works like a charm.