QuickCPU Power saving - My settings (Relevant for 6xxx/7xxx/8xxx/9xxx/10xxx processors)

Keljian

[H]ard|Gawd
Joined
Nov 7, 2006
Messages
1,279
There have been some questions about how I managed to get the power usage of my 9900k down so low. What follows is what I've discovered over the last few weeks

Recently I lost my job (covid!), I'm a reasonably heavy user in terms of time on my computer, and I looked at how I could save on power on my desktop.

I've got a 9900k, per my sig. It is not known as the most power efficient/frugal cpu in the world, I never bought it for that reason. What I didn't realise is that when it isn't working hard, it can be extremely efficient.

The first thing I did was go into the bios and enabled speedshift, explicitly enabled all possible c-states (C10 and below) and set a negative offset for the processor voltage.

The second thing I did was learn a bit about Speedshift, and how it and Hardware Duty Cycling worked/what they are.

I looked for ways of checking that these were enabled (they weren’t) and then ways of enabling them.

I found that I could change the power plan variables manually to enable them, but this was fiddly and tedious.

Then I discovered Quick CPU. QuickCPU is a basic, open source piece of software that allows you to modify settings in your windows power plan, monitor use of the processor, and alter nearly every power setting the processor has.

I set up my plan specific to my needs of the computer, and have not noticed any slow-downs, hitching, stuttering or other ill effects in what I do.

Now this does not mean I'll ace benchmarks with the machine, but it'll still be within a few percent of the top speeds, while being a lot more frugal on power.

Specifically how I set my computer up is as below, most of the options are defaults.


Capture.JPG

Capture.PNG

Capture3.JPG



Capture1.JPG


One somewhat welcome side effect is that the room the computer is in is much cooler as a result of this.

I hope that this helps people a bit, and I'll edit this post to add stuff/make it more readable if there is enough interest. Feel free to ask questions - I'll do my best to answer them
 
Last edited:

PhaseNoise

2[H]4U
Joined
May 11, 2005
Messages
3,306
Thanks for the pointer to this tool. I love lean little utilities in this era of "everything is an electron app". Also, I fixate on efficiency for no real good reason.
 

kirbyrj

Fully [H]
Joined
Feb 1, 2005
Messages
27,115
I appreciate the screenshots of the settings pages. Gives me more info to work with. I was getting some blue screens when I tried to implement some of this though. Still tweaking to find something stable for me ;).

What does the "Turbo Boost Index" and "Frequency Scaling Index" do and why do you have them set at 60%?
 
Last edited:

Keljian

[H]ard|Gawd
Joined
Nov 7, 2006
Messages
1,279
Turbo boost index specifies when the load on the processor should trigger turbo, ie. 60% means when the load shoots up to 40% or greater at the base frequency, trigger turbo. The other frequency (under 40% in my case) is the base clock frequency which is 3.6ghz on the 9900k.

Frequency scaling index is the same but for the base clock frequency rather than turbo.

For completeness

Core parking index refers to the number of cores that are allowed to be parked 100% means all cores are to stay unparked, 10% means 90% of the cores can be parked.

I find it weird that they have done it this way and inverted the numbers, but it works

As to why I set them this way? Because those seem like reasonable targets, and I haven’t experienced any performance drops
 
Last edited:

kirbyrj

Fully [H]
Joined
Feb 1, 2005
Messages
27,115
A couple questions:
How do I make a custom power plan? It looks like you made your own.

If I screw up a setting in my balanced plan by not paying attention, how can I recover the original settings to start over? For example, core parking is now a regular feature of my balanced plan for some reason.
 

kirbyrj

Fully [H]
Joined
Feb 1, 2005
Messages
27,115
DOH! That makes more sense. I was looking in QuickCPU to create it and fix it. Forgot about just telling Windows to do it :oops:.
 

Keljian

[H]ard|Gawd
Joined
Nov 7, 2006
Messages
1,279
For point of reference, I've used my Eaton UPS to measure power consumption at idle vs load. Note that this is with monitors, amps and other gear:

System per sig

Idle with default Balanced power plan: 137W
Idle with all of the above implemented: 119W

Might not seem like much, but it all adds up. I'll connect the pc to the UPS solo a bit later.

I did notice two cores were asleep while I was playing AC:Origins last night.
 

PhaseNoise

2[H]4U
Joined
May 11, 2005
Messages
3,306
Efficiency is actually a really cool metric. Top performance always gets the accolades, and rightly so I suppose, but what you get out of a watt is also really cool.

I do a lot of embedded systems work, and what I can get for a watt is often as much a constraint as perf/$.
 

Keljian

[H]ard|Gawd
Joined
Nov 7, 2006
Messages
1,279
Efficiency is actually a really cool metric. Top performance always gets the accolades, and rightly so I suppose, but what you get out of a watt is also really cool.

I do a lot of embedded systems work, and what I can get for a watt is often as much a constraint as perf/$.

As you know, I do embedded systems work too - performance per watt is super important to me (I'm working on a product that uses coin cells at the moment). Plus, why should I pay for power that I don't need to?
 

Keljian

[H]ard|Gawd
Joined
Nov 7, 2006
Messages
1,279
Without the whole shebang:
  • 62-70W with the balanced plan.
  • 44-48W with the updated power plan
That's a delta of about 12-22W, or 1/3 less with the new plan - as always, ymmv

Notes:
  • I have 5 total fans in the setup, though they are running slow being that it isn't producing much heat
  • it's powering my keyboard and mouse (which have LEDs... ) , plus webcam.
  • Bluetooth enabled
  • Xbox controller dongle attached
  • System setup per my sig
 
Last edited:

kirbyrj

Fully [H]
Joined
Feb 1, 2005
Messages
27,115
I have been trying to pay more attention to this in my AMD system sitting here. It's a little trickier though as there are two numbers in Ryzen Master: CPU Power (7-10W while browsing generally and 2-3W with nothing running) and SOC Power (pretty steady 12W).

I don't have a way to monitor the wall (might just pick up a kill-a-watt to have an idea).
 
Last edited:

Keljian

[H]ard|Gawd
Joined
Nov 7, 2006
Messages
1,279
I have been trying to pay more attention to this in my AMD system sitting here. It's a little trickier though as there are two numbers in Ryzen Master: CPU Power (7-10W while browsing generally and 2-3W with nothing running) and SOC Power (pretty steady 12W).

I don't have a way to monitor the wall (might just pick up a kill-a-watt to have an idea).

Quickcpu has two figures too. I have noticed an increase in power consumption in Quickcpu (by about 2W at idle) having moved to running the cache at 47x vs 43x

Note however that more is actually implemented on the ryzen cpu (eg things like usb3 and sound) which intel doesn’t have on their cpu, but has on the chipset
 

Keljian

[H]ard|Gawd
Joined
Nov 7, 2006
Messages
1,279
5150Joker

Here are the Davinci results as requested - Note it is the 60fps/4k version of the movie. Apex legends is still downloading. I noticed a peak voltage of 1.272v but it was for a split second, same with temps. Download the attached file if you like. Maximum wattage was 154W - unfortunately the forum isn't letting me post the entire 4K+ screenshot, but here's the HWinfo64/Quickcpu readout about a minute into the compression
 

Attachments

  • Capture.JPG
    Capture.JPG
    450.7 KB · Views: 0
Last edited:

Keljian

[H]ard|Gawd
Joined
Nov 7, 2006
Messages
1,279
@5051Joker Apex legends as requested - note the cores parked (even in gameplay) - new run of HWinfo for the peaks
 

Attachments

  • Apex legends.JPG
    Apex legends.JPG
    345.4 KB · Views: 0

5150Joker

Supreme [H]ardness
Joined
Aug 1, 2005
Messages
4,538
5150Joker

Here are the Davinci results as requested - Note it is the 60fps/4k version of the movie. Apex legends is still downloading. I noticed a peak voltage of 1.272v but it was for a split second, same with temps. Download the attached file if you like. Maximum wattage was 154W - unfortunately the forum isn't letting me post the entire 4K+ screenshot, but here's the HWinfo64/Quickcpu readout about a minute into the compression

Thx but one minute isn’t really sustained. With speedshift I bet your cpu kept cycling below 5 GHz during the encode to maintain a voltage/temp threshold. That’s why every overclocker disables it.
 

Keljian

[H]ard|Gawd
Joined
Nov 7, 2006
Messages
1,279
Thx but one minute isn’t really sustained. With speedshift I bet your cpu kept cycling below 5 GHz during the encode to maintain a voltage/temp threshold. That’s why every overclocker disables it.

No, no voltage/temp threshhold, it just cycled below because of lack of instantaneous load. Note peak temps were still below 80C, no way it was thermally/voltage throttling . I'll run Davinci for the entire encode if you prefer.

Thing is, you don’t need the processor to be running at 5ghz constantly, and it is not designed to do that.

It is one thing to say the 9900k is a hot processor, another to say “it’s a hot processor when you’re running it at maximum speed constantly”. You don’t redline an engine constantly when you’re driving at 10kph/mph, why should you do that when you’re running a chip? You also don’t need a fire breathing v8 to drive granny down to the shop (notice all those cores sleeping in the apex legends shot?)

It should be abundantly clear by now that my preference is not to do that. I don’t want my rig to burn power unnecessarily.
 
Last edited:

5150Joker

Supreme [H]ardness
Joined
Aug 1, 2005
Messages
4,538
No, no voltage/temp threshhold, it just cycled below because of lack of instantaneous load. Note peak temps were still below 80C, no way it was thermally/voltage throttling . I'll run Davinci for the entire encode if you prefer.

Thing is, you don’t need the processor to be running at 5ghz constantly, and it is not designed to do that.

It is one thing to say the 9900k is a hot processor, another to say “it’s a hot processor when you’re running it at maximum speed constantly”. You don’t redline an engine constantly when you’re driving at 10kph/mph, why should you do that when you’re running a chip? You also don’t need a fire breathing v8 to drive granny down to the shop (notice all those cores sleeping in the apex legends shot?)

It should be abundantly clear by now that my preference is not to do that. I don’t want my rig to burn power unnecessarily.

When you're encoding, it should be pegging all cores at 100% at all times pretty much and never be thermally cycling. You're going through p-states which is what speedshift does and it is never sustained. I could go turn on C-states and P-Shift and do the same thing but it would be disingenuous.
 

Keljian

[H]ard|Gawd
Joined
Nov 7, 2006
Messages
1,279
When you're encoding, it should be pegging all cores at 100% at all times pretty much and never be thermally cycling. You're going through p-states which is what speedshift does and it is never sustained. I could go turn on C-states and P-Shift and do the same thing but it would be disingenuous.

Urm, no. No code is going to peg cores at 100% constantly, not even the best multithreaded code..

You would need a loop that fits in cache that just goes over and over stuff, generally that is not very useful

I propose a different test, we pit one 9900k against another, one with all this stuff enabled, vs another that doesn’t. I would bet money they are within a few percentage points of each other performance wise, but the one without pstates and cstates enabled will be hotter and use more power
 
Last edited:

Keljian

[H]ard|Gawd
Joined
Nov 7, 2006
Messages
1,279
So I have been looking at inexpensive ways of reducing the amount of noise my PC puts out when under load. I spent considerable time looking at new cases, before stumbling on a really easy way of controlling fans and monitoring load: Argus Monitor.

While it is registered software (USD$19.99/3 years of updates), it's a lot less expensive than buying a new case or cooler. After carefully adjusting fan profiles now the machine whisper quiet, I'm very impressed. It's also allowed me to tune based on averages eg. cpu temperature over 10 secs, as opposed to just cpu temperature, which is more valuable as it means the fans aren't pulsing.

It also shows the effect of speedstep graphically which is kind of nice.

1592265279858.png


And CPU power (note this starting a handbrake encode)

1592266761212.png
 
Last edited:

kirbyrj

Fully [H]
Joined
Feb 1, 2005
Messages
27,115
I just want to note that I installed this again in my 10700 (non-k) machine and it works great with the 10th gen CPU's also (unsurprising as they are still Skylake based).

Just the C states I have to play around with. I can't configure them in the program as it gives me an error. Probably needs an update for the new platform.
 

kirbyrj

Fully [H]
Joined
Feb 1, 2005
Messages
27,115
Well, I'll have to tweak it some more. I'll get random reboots when I have QuickCPU running, but when I leave it stock, it is rock solid. Right now I have it very similar to what is posted above with the frequency scaling index, turbo boost index, and core parking index. Maybe I'm trying to be too aggressive?
 

Keljian

[H]ard|Gawd
Joined
Nov 7, 2006
Messages
1,279
Well, I'll have to tweak it some more. I'll get random reboots when I have QuickCPU running, but when I leave it stock, it is rock solid. Right now I have it very similar to what is posted above with the frequency scaling index, turbo boost index, and core parking index. Maybe I'm trying to be too aggressive?

Going to need to know a lot more about your setup to help.

What voltage/offset?
Load line?
Turbo frequency?
Motherboard?
Cache speed?

etc etc
 

kirbyrj

Fully [H]
Joined
Feb 1, 2005
Messages
27,115
Well, I'm running a 10700 (non-k) at stock voltage, no offset. Turbo set to 225W for both PL1/PL2, but I never get that high. Usually I top out around 160W and even that's a stretch for most things.
I didn't adjust the LLC.
BCLK is at 102
Cache is 43 multipler (so 4.4Ghz effective based on BCLK)
RAM is at 3670Mhz
Motherboard is Asus Strix Z490-G

It's not under load that it resets. It's more of an idle thing. I'll walk away and come back to find that the computer restarted. Maybe a core parking issue?
 

Keljian

[H]ard|Gawd
Joined
Nov 7, 2006
Messages
1,279
Well, I'm running a 10700 (non-k) at stock voltage, no offset. Turbo set to 225W for both PL1/PL2, but I never get that high. Usually I top out around 160W and even that's a stretch for most things.
I didn't adjust the LLC.
BCLK is at 102
Cache is 43 multipler (so 4.4Ghz effective based on BCLK)
RAM is at 3670Mhz
Motherboard is Asus Strix Z490-G

It's not under load that it resets. It's more of an idle thing. I'll walk away and come back to find that the computer restarted. Maybe a core parking issue?

So I’m going to ask obvious things:

  • Have you installed the latest inf drivers for the chipset?
  • Are you running the latest bios?
  • Have you tried it without messing with the bclk?
  • What is the configured peak turbo speed?
  • Have you tried it without hardware duty cycling?
  • Can you isolate it to core parking (disabling speedshift and HDC)?
  • What is running in the background?
  • Is cache downbin enabled? (Guessing not, maybe trying to run the cache faster than the processor is an issue?)
 

kirbyrj

Fully [H]
Joined
Feb 1, 2005
Messages
27,115
Yes on the latest INF and bios.

No, I haven't tried it without the BCLK adjustment, but it is perfectly stable for days on end without QuickCPU running. I can try it though.
Peak turbo speed is ~4.9Ghz single core, ~4.7Ghz all core.
I'll try it without the hardware duty cycling
Not necessarily isolated to core parking. That was just a guess.
I never checked the cache downbin, but it looks like it does downclock with CPU load in CPU-Z.

I'll play around with it more. It's just tough to diagnose because I'm not using the computer actively when it happens. The only thing I've noticed is that when I'm not running QuickCPU it is perfectly stable, so I'm guessing it's a setting in there. I just have to enable things one at a time to see if I can isolate it.

Starting a test right now. Put the core parking index to 100% and am going to see what happens over the next day or so.

Edit: It locked up again. So I guess my initial core parking theory is wrong. Onto another setting! I'm going to try Hardware Duty Cycling.

Edit2: Well, Hardware Duty Cycling is the early leader in the clubhouse for causing the freezing/reboots. I disabled it, and my computer has been running all day with a mixed workload (mostly idle which is where it was rebooting). Let's see how it looks tomorrow morning.
 
Last edited:

Keljian

[H]ard|Gawd
Joined
Nov 7, 2006
Messages
1,279
So I've been away for a bit, which gave me time to play with different settings, I feel I've hit on what could be good for the majority of 9900ks/10700ks of the same calibre of mine.

1596338998054.png

45/46% turbo boost and 10% core parking seem to be approximate values which work well, as in, if you use 44% and 7% it won't make much difference.

1596339036547.png


Note the core ratios (5.1ghz/51) and the long/short duration power limits/times. This allows up to 8 cores to boost to 5.1ghz for a short period before levelling out at 4.9-5.0 under excessively heavy AVX loads.
This means fans and temps don't go nuts after prolonged loads (note I'm running a D15). I got these numbers from this chart on GamersNexus.
An increase to 165-170W for the long duration power limit will give marginally higher clocks, albeit at much higher temperatures, and the times could be potentially increased with water cooling as time to max increases.
In the bios I have set the uncore to a maximum of 4.7ghz, but with downbin so that it reduces when we're idling.

1596339069043.png


Min allowed performance of 2.9-3.5ghz allows good general purpose power.

New version of Quickcpu is out with a power plan manager, if anyone's interested, and I've attached my power plan settings which could be used to make your own or be implemented directly.
 

Attachments

  • My Custom Plan 1Report-2020.02.8-14.06.24.pdf
    107.1 KB · Views: 0
  • My Custom Plan 1-Backup-2020.02.8-14.07.32.pow.zip
    1.9 KB · Views: 0
Last edited:

jyi786

Supreme [H]ardness
Joined
Jun 13, 2002
Messages
5,562
The question I have about this QuickCPU program is whether or not it will give me the ability to control minimum CPU clock speeds (and thereby minimum voltages). The reason why I ask this is because I have the opposite problem of most.

My CPU on auto voltage and using the FIVR can remain super rock solid while abusing it....but it has instability and crashes at IDLE. Even using vdroop mitigations, the speeds and voltages drop way too far down, causing instability at idle or programs that require low CPU usage (such as simply watching videos).

I was thinking, if I can control the minimum CPU clock speed, I could thereby prevent the voltage from dropping so low. It seems that QuickCPU will allow me to do that. Can anyone confirm?
 

Keljian

[H]ard|Gawd
Joined
Nov 7, 2006
Messages
1,279
The question I have about this QuickCPU program is whether or not it will give me the ability to control minimum CPU clock speeds (and thereby minimum voltages). The reason why I ask this is because I have the opposite problem of most.

My CPU on auto voltage and using the FIVR can remain super rock solid while abusing it....but it has instability and crashes at IDLE. Even using vdroop mitigations, the speeds and voltages drop way too far down, causing instability at idle or programs that require low CPU usage (such as simply watching videos).

I was thinking, if I can control the minimum CPU clock speed, I could thereby prevent the voltage from dropping so low. It seems that QuickCPU will allow me to do that. Can anyone confirm?

You don't have the opposite problem of most, it's an issue that still keeps popping up. There are voltage settings in QuickCPU, I haven't used them though. I don't think that (in particular) is supported.
 

jyi786

Supreme [H]ardness
Joined
Jun 13, 2002
Messages
5,562
You don't have the opposite problem of most, it's an issue that still keeps popping up. There are voltage settings in QuickCPU, I haven't used them though. I don't think that (in particular) is supported.

Good to know I'm not alone. I scoured and did real heavy research on it, but there's nothing definitive out there on how to combat it, other than to run a manual voltage at all times.

I'm cool with QuickCPU not giving me the ability to control voltage, but I saw a slider which said something to the effect of "minimum CPU clock". My 9980XE using FIVR on auto voltage would drop to 0.83GHz and around 0.65V at idle and would randomly crash. I was thinking I could use the minimum CPU clock control in QuickCPU to force my 9980XE to run at a faster minimum speed, circumventing the extreme droop.
 

Dan_D

Extremely [H]
Joined
Feb 9, 2002
Messages
57,209
Efficiency is actually a really cool metric. Top performance always gets the accolades, and rightly so I suppose, but what you get out of a watt is also really cool.

I do a lot of embedded systems work, and what I can get for a watt is often as much a constraint as perf/$.

The thing is, 10w or whatever doesn't actually make much of a difference on your electric bill.
 

PhaseNoise

2[H]4U
Joined
May 11, 2005
Messages
3,306
The thing is, 10w or whatever doesn't actually make much of a difference on your electric bill.
For sure. My comment is not about utility costs on a no-limits desktop.

In my work, it often means getting by with a smaller battery, a smaller psu, or simpler cooling / power delivery designs in general. Those are BOM items which affect cost per unit. If I ask for a bigger battery, something has to give.
 

Dan_D

Extremely [H]
Joined
Feb 9, 2002
Messages
57,209
For sure. My comment is not about utility costs on a no-limits desktop.

In my work, it often means getting by with a smaller battery, a smaller psu, or simpler cooling / power delivery designs in general. Those are BOM items which affect cost per unit. If I ask for a bigger battery, something has to give.

But more power is always the solution. :)
 

Keljian

[H]ard|Gawd
Joined
Nov 7, 2006
Messages
1,279
The thing is, 10w or whatever doesn't actually make much of a difference on your electric bill.
10w for 8 hours a day = 80wh
80wh x21 days (assume work days per month) = 1680wh, or 1.68kwh = 20kwh/year

1 kWh here is 25-28c, so let’s say 26.5, so $5.30/year on average.

I saved somewhere in the order of 2-3x this (let’s say 12.50?) per year.

But that doesn’t take into account the savings in power from speedshift, hardware p states, parking and dynamic power, which probably is significantly higher when (for instance) I am gaming or encoding or compiling or running something else that is intensive.
 

learners permit

Limp Gawd
Joined
Jun 15, 2005
Messages
360
10 watts x1/3 total population of 2.25 billion would be substantial if one were to think a bit globally. I've tuned my cpu down to about 60 watts max under my normal everyday usage. I hope to see video cards under 100 watts in my lifetime but that's not looking very likely at this point in time.
 

Keljian

[H]ard|Gawd
Joined
Nov 7, 2006
Messages
1,279
10 watts x1/3 total population of 2.25 billion would be substantial if one were to think a bit globally. I've tuned my cpu down to about 60 watts max under my normal everyday usage. I hope to see video cards under 100 watts in my lifetime but that's not looking very likely at this point in time.
When playing games, or doing moderately intensive work, the chip I have doesn’t see the high side of 60W for more than the briefest of moments. I’m prepared to forgo a little bit of power momentarily to get a boost in responsiveness
 
Top