[Latte]

Updated the profiles; reverted to P-State as I had performance issues with autonomous mode (lower CPU-z scores)

My experience so far with your balanced v5 plan:

• major lags, as soon I start e.g. AIDA or CPU-z bench
• temporary mouse freezes, when I got load
• significant performance drops in AIDA and CPU-z bench
• AIDA bench latency of 68.3ns - had aprox. 61ns with windows balanced and 63ns with v4 balanced

- tried AIDA several times with v5 balanced, but full AIDA test always freezes | single latency bench works in about 10x the time I need with other plans

Spoiler: cpu-z benches v5 balanced <> windows balanced

Do I have to adjust something into bios maybe? C States, DF States? It's enabled for now. Maybe increasing min. cpu state in profile?

 

[ManniX-ITA]

My experience so far with your balanced v5 plan:

• major lags, as soon I start e.g. AIDA or CPU-z bench
• temporary mouse freezes, when I got load
• significant performance drops in AIDA and CPU-z bench
• AIDA bench latency of 68.3ns - had aprox. 61ns with windows balanced and 63ns with v4 balanced

- tried AIDA several times with v5 balanced, but full AIDA test always freezes | single latency bench works in about 10x the time I need with other plans

Spoiler: cpu-z benches v5 balanced <> windows balanced

View attachment 2475298
View attachment 2475299

Do I have to adjust something into bios maybe? C States, DF States? It's enabled for now. Maybe increasing min. cpu state in profile?

It's weird... please try changing the min. cpu state up.

Except the AIDA 64 latency I don't have any of the other issues.

 

[Latte]

It's weird... please try changing the min. cpu state up.

Except the AIDA 64 latency I don't have any of the other issues.

That's weird indeed! With min. cpu state set to 100 it's better with cpu-z values in MT, but unfortunately the rest behaves the same.

 

[ManniX-ITA]

That's weird indeed! With min. cpu state set to 100 it's better with cpu-z values in MT, but unfortunately the rest behaves the same.

Updated again, let me know.
I've disabled core parking as probably it's the cause of these weird issues.
Also reverted to autonomous mode; Ultimate now scores right with CPU-z, the Balanced less but should still be very reactive and without unwanted CCD spikes.
AIDA64 latency should be right now, make a check.

 

[sharkzer]

Is the ultimate performance power plan safe to use 24/7 with PBO?

 

[Latte]

Updated again, let me know.
I've disabled core parking as probably it's the cause of these weird issues.
Also reverted to autonomous mode; Ultimate now scores right with CPU-z, the Balanced less but should still be very reactive and without unwanted CCD spikes.
AIDA64 latency should be right now, make a check.

First things first - thanks a lot for your effort and your time!

Your new balanced v6 is on a good way, related to bench performance:
Hints:

• images show following:

1. 30 seconds idle warm-up
2. CPU-Z bench
3. AIDA bench (visible in image)
4. CPU-Z bench (visible in image)
5. 4 minutes idle

• similar results in CPU-Z & AIDA (within margin of error)
• AIDA bench took aprox. 4 minutes longer, I don't know why!

Spoiler: Balanced v6 v Windows Balanced

My observations:

• not as snappy as windows balanced or your balanced v4 (v4 was the snappiest till now by far)
• every app takes significantly longer to open it, seems a litte bit laggy
• temps & voltages are nice
• good performance
• vCore is more fluctuating than v4 => lower average!

My order so far (snappiness):

1. v4
2. windows balanced
3. v1
4. v6
5. ...

My order so far (performance):

1. windows balanced
2. v4
3. v1
4. v6
5. ...

 

[kairi_zeroblade]

Can I have a say on this?? well eventually some would have an issue mainly because you are using numbers that aren't proportional to the number of cores that a user might use this one..so end result is either he losses/cripples performance or increase the heat and power draw..

Ironically for smaller core counts than what you use/have the percentages on how many cores should be online/offline/parked should be re-computed/adjusted/calibrated proportionally to avoid issues..

I am using a 5800x and upon loading your profile it goes haywire, and upon evaluation and comparison to my personal profile your values are too high/out of proportion on my setup..

here is my personal profile stats..

when low usage/load/utilization my system is set to park 8-10 cores..when there is a light burst of load I only allow 2 cores to wake up immediately..if the currently running unparked/online cores aren't still enough then that is where the rest of the offline/parked cores (all) wake up and help..your profile still lacks this kind of logic as you didn't played enough with the other core parking options also your load thresholds are all messed up..Core Parking isn't just a power saving feature, its a feature meant to efficiently use/train multi core setups to run effectively without the penalty on power draw vs latency..(the logic aims to maximize single core boosting algorithms for workloads that are within the thresholds set on the parameters instead of bruteforcing the workload to all cores, if this feature is calibrated perfectly, you get an effective and efficient all rounder system)

memory latency on AIDA64 is 57.1ns running 3800mhz RAM CL16 with 1900FCLK..despite Core Parking enabled, I still have no performance impact on my benches, on CPU-Z I usually get 650/6842 scores..(non AVX test) this profile as well is the one I universally use when gaming and benching, I don't like having to touch many stuff when I need to do something..Ryzen's architecture already has the knobs to tame itself down and make it perform the way you want it..

 

[ManniX-ITA]

Is the ultimate performance power plan safe to use 24/7 with PBO?

Yes it's safe, it's more power hungry in idle and low load

First things first - thanks a lot for your effort and your time!

Your new balanced v6 is on a good way, related to bench performance:
Hints:

• images show following:

1. 30 seconds idle warm-up
2. CPU-Z bench
3. AIDA bench (visible in image)
4. CPU-Z bench (visible in image)
5. 4 minutes idle

• similar results in CPU-Z & AIDA (within margin of error)
• AIDA bench took aprox. 4 minutes longer, I don't know why!

Spoiler: Balanced v6 v Windows Balanced

View attachment 2475391
View attachment 2475392

My observations:

• not as snappy as windows balanced or your balanced v4 (v4 was the snappiest till now by far)
• every app takes significantly longer to open it, seems a litte bit laggy
• temps & voltages are nice
• good performance
• vCore is more fluctuating than v4 => lower average!

My order so far (snappiness):

1. v4
2. windows balanced
3. v1
4. v6
5. ...

My order so far (performance):

1. windows balanced
2. v4
3. v1
4. v6
5. ...

Thanks for the comparison!

I really don't know why AIDA takes longer... will check if it's the same for me.

Yes that is a bit more laggy is as intended; vCore goes lower and power consumption is quite low.
The goal is to fix the temp spikes and so far that's the best I could do.

The test against the balanced should be monitoring the CCD temperatures with the HWInfo graph.
With the balanced a low usage on 1 or 2 cores will spike the temp very high; this will make spike the CPU die temp up and down.
Typically having Chrome a lots of background apps open as below:

v6:

Balanced:

I will try to make it snappier but so far either it starts consuming more or the spikes comes back.

I like a lot the snappiness of the v4 as well and will probably make a specific profile on it with the aim to get max snappiness and performances.

Can I have a say on this?? well eventually some would have an issue mainly because you are using numbers that aren't proportional to the number of cores that a user might use this one..so end result is either he losses/cripples performance or increase the heat and power draw..

Ironically for smaller core counts than what you use/have the percentages on how many cores should be online/offline/parked should be re-computed/adjusted/calibrated proportionally to avoid issues..

I am using a 5800x and upon loading your profile it goes haywire, and upon evaluation and comparison to my personal profile your values are too high/out of proportion on my setup..

here is my personal profile stats..
View attachment 2475396

View attachment 2475397

when low usage/load/utilization my system is set to park 8-10 cores..when there is a light burst of load I only allow 2 cores to wake up immediately..if the currently running unparked/online cores aren't still enough then that is where the rest of the offline/parked cores (all) wake up and help..your profile still lacks this kind of logic as you didn't played enough with the other core parking options also your load thresholds are all messed up..Core Parking isn't just a power saving feature, its a feature meant to efficiently use/train multi core setups to run effectively without the penalty on power draw vs latency..(the logic aims to maximize single core boosting algorithms for workloads that are within the thresholds set on the parameters instead of bruteforcing the workload to all cores, if this feature is calibrated perfectly, you get an effective and efficient all rounder system)

memory latency on AIDA64 is 57.1ns running 3800mhz RAM CL16 with 1900FCLK..despite Core Parking enabled, I still have no performance impact on my benches, on CPU-Z I usually get 650/6842 scores..(non AVX test) this profile as well is the one I universally use when gaming and benching, I don't like having to touch many stuff when I need to do something..Ryzen's architecture already has the knobs to tame itself down and make it perform the way you want it..

Sure, jump in

I didn't test enough the Core parking; don't like the added latency and while it's not reflected in benches it has an hit on the snappiness.

I've set those values cause they looked liked they were working better with that specific combination of autonomous mode, heterogeneous and boost policy.
It's a kind of alchemy cause the AMD power management it's still terrible like with the 3000, if not worse.

The aim was to reduce the spikes and keep temperatures and power consumption in check.
Didn't work out and parking looks like doesn't work properly for anyone.

Maybe I'll make a specific profile with parking but as you said the main problem is that you can only define percentages and between my 5950x and a 5600x there's too much difference in core count.
The boosting is sadly not positively affected by parking, it works exactly the same as when the cores are just sleeping.
Also the temperature doesn't seem to improve that much if not at all; hard to say with the 5950x.

The good side of the profiles in autonomous mode is that the behavior should improve with new AGESA, if AMD decides to improve it.

 

[ManniX-ITA]

AIDA64 Balanced, 2m 50s:

AIDA64 v6, 2m 45s:

Took a bit less for me...
Maybe there's something in the BIOS profile or the AGESA (mine is 1.2.0.0).
Being an autonomous profile makes it more sensitive to the settings.

What about the Ultimate profile, same issue?

 

[Latte]

AIDA64 Balanced, 2m 50s:

View attachment 2475474

AIDA64 v6, 2m 45s:

View attachment 2475475

Took a bit less for me...
Maybe there's something in the BIOS profile or the AGESA (mine is 1.2.0.0).
Being an autonomous profile makes it more sensitive to the settings.

What about the Ultimate profile, same issue?

Yes, maybe it's the bios version. I use F31 for now - waiting for stable F33 I guess. For now I didn't used the ultimate plan, because for gaming the balanced one is quite nice.

 

[kairi_zeroblade]

Sure, jump in

I didn't test enough the Core parking; don't like the added latency and while it's not reflected in benches it has an hit on the snappiness.

I've set those values cause they looked liked they were working better with that specific combination of autonomous mode, heterogeneous and boost policy.
It's a kind of alchemy cause the AMD power management it's still terrible like with the 3000, if not worse.

The aim was to reduce the spikes and keep temperatures and power consumption in check.
Didn't work out and parking looks like doesn't work properly for anyone.

Maybe I'll make a specific profile with parking but as you said the main problem is that you can only define percentages and between my 5950x and a 5600x there's too much difference in core count.
The boosting is sadly not positively affected by parking, it works exactly the same as when the cores are just sleeping.
Also the temperature doesn't seem to improve that much if not at all; hard to say with the 5950x.

The good side of the profiles in autonomous mode is that the behavior should improve with new AGESA, if AMD decides to improve it.

1. There is no latency impact, Core Parking isn't the same as Hotplugging..the parked cores are not offline, dead or sleeping, they are online but are just idling and tagged as "parked"..The scheduler will make up for that via utilizing turbo boost..(if you have just properly set the thresholds on your profiles, as it is right now, sadly its a mess..)

2. There is an option there where you messed up so bad..you set up parameters for Power Efficiency Class 1 Processors but you never assigned any cores for it..

3. Autonomous mode has nothing to do with boost policy..boosting is affected by workload and the thresholds you put and how the AGESA code reacts/works..

4. Generally speaking, not all options for core parking works with Ryzen, those Heterogenous options don't work since Ryzen isn't built with HMP (Heterogenous Multi Processing) on mind..those options will only work on big.LITTLE architecture (mobile SOC's), MS added those in since there is a version of Win10 for mobile SOC's..

5. I don't see the point where the spikes come in..please don't equate less spikes = more power savings/better performance or anything..it doesn't co-relate..what you should be focusing (and that is what these profiles are all about) is the frequency behaviour (on boost and off boost)..we can argue the whole day long how flat you can get on your voltage spikes but that doesn't really equate to any real world stuff..

I highly advice you review your profile and set the proper thresholds for it to properly work..

here is a run from mine, as you can see the frequency that has been captured is even low..that means I am in the state of low power or some of my cores are parked..and as you see there is 0 impact on memory latency.. (those are what my sticks can do best, they aren't godly binned)

As I said above, if thresholds and Core Parking is properly setup you have a very efficient system..10 cores parked, and 6-8 cores go active on light to medium workload (includes Gaming) and there's no negative impact..

 

[ManniX-ITA]

1. There is no latency impact, Core Parking isn't the same as Hotplugging..the parked cores are not offline, dead or sleeping, they are online but are just idling and tagged as "parked"..The scheduler will make up for that via utilizing turbo boost..(if you have just properly set the thresholds on your profiles, as it is right now, sadly its a mess..)

2. There is an option there where you messed up so bad..you set up parameters for Power Efficiency Class 1 Processors but you never assigned any cores for it..

3. Autonomous mode has nothing to do with boost policy..boosting is affected by workload and the thresholds you put and how the AGESA code reacts/works..

4. Generally speaking, not all options for core parking works with Ryzen, those Heterogenous options don't work since Ryzen isn't built with HMP (Heterogenous Multi Processing) on mind..those options will only work on big.LITTLE architecture (mobile SOC's), MS added those in since there is a version of Win10 for mobile SOC's..

5. I don't see the point where the spikes come in..please don't equate less spikes = more power savings/better performance or anything..it doesn't co-relate..what you should be focusing (and that is what these profiles are all about) is the frequency behaviour (on boost and off boost)..we can argue the whole day long how flat you can get on your voltage spikes but that doesn't really equate to any real world stuff..

I highly advice you review your profile and set the proper thresholds for it to properly work..

View attachment 2475514

here is a run from mine, as you can see the frequency that has been captured is even low..that means I am in the state of low power or some of my cores are parked..and as you see there is 0 impact on memory latency.. (those are what my sticks can do best, they aren't godly binned)

As I said above, if thresholds and Core Parking is properly setup you have a very efficient system..10 cores parked, and 6-8 cores go active on light to medium workload (includes Gaming) and there's no negative impact..

1. I will check but from what I've seen there is a latency impact and it's variable, on my benching install which is very clean is more noticeable.

2. I don't know which one are you referring to

3. Of course it has and it's macroscopic; just test the behavior with different policies and autonomous mode enabled/disabled.

4. Again test it, you are making assumptions. I never considered these options cause I've made the same assumption in the past, thought they were for big.LITTLE. But it's not like that.

5. The goal of this profile is to reduce the temp spikes (which has an impact on power savings, why shouldn't) without destroying the snappiness and having a big impact on performances and max boost clock. The CCD temp spikes are causing spikes on the CPU die which is what usually people are setting the CPU fan speed which means consequently the fan is ramping up and down like hell when you are just browsing a web page.

I don't understand, if you have such a beautiful and perfect profile why don't you share it? Do you have it published somewhere?

 

[kairi_zeroblade]

1. I will check but from what I've seen there is a latency impact and it's variable, on my benching install which is very clean is more noticeable.

2. I don't know which one are you referring to

3. Of course it has and it's macroscopic; just test the behavior with different policies and autonomous mode enabled/disabled.

4. Again test it, you are making assumptions. I never considered these options cause I've made the same assumption in the past, thought they were for big.LITTLE. But it's not like that.

5. The goal of this profile is to reduce the temp spikes (which has an impact on power savings, why shouldn't) without destroying the snappiness and having a big impact on performances and max boost clock. The CCD temp spikes are causing spikes on the CPU die which is what usually people are setting the CPU fan speed which means consequently the fan is ramping up and down like hell when you are just browsing a web page.

I don't understand, if you have such a beautiful and perfect profile why don't you share it? Do you have it published somewhere?

1. your latency is from the dual CCD construction of the 5900x and 5950x..(I have a working solution on that)

2. This is why throwing just random numbers on every option on your profiles is pointless and pretty much messes things up..

4. been there, done that already even with my 9th gen Intel setup I have no issues on my side with autonomous mode..its just again the ridiculous values you pitch in without properly accounting load thresholds..I tried all the Heterogenous options and none of them work with my properly calibrated load thresholds..

5. the reduce in temp spikes has nothing to do with how the scheduler would react on certain loads..again, you aren't making sense, since those 2 doesn't co-relate..you are the only one destroying the balance/snappiness by just throwing non-sensical values and not properly accounting/calibrating load thresholds properly..(the fans are another story on this as it co-relates to the temps perfectly)..also if you are solely basing the reaction of the scheduler to temps then you are again doing it wrong..the pursuit of performance has always lead to additional heat..

By default balance already provides snappiness with a little bit of power savings in mind..for a desktop, Heat was never a factor to consider to begin with..

Before you should brought this thread up, you could have spent a week or 2 to read on what the options exactly mean and do..(I won't go further on leading you to where you should read, nor dump mine, I have never been fond of people using my stuff and the same as you do, adjust it for them..)

I have been emphasizing for you to review your numbers on the options on your profile(not to start an argument)..I know your intentions are for the good but seriously if you are going to release something like these it should have been something really thought of and has been calibrated properly, nicely..I would have wanted to help others about this but it would seem I was to hi-jack the thread post..

 

[ManniX-ITA]

1. your latency is from the dual CCD construction of the 5900x and 5950x..(I have a working solution on that)

2. This is why throwing just random numbers on every option on your profiles is pointless and pretty much messes things up..

4. been there, done that already even with my 9th gen Intel setup I have no issues on my side with autonomous mode..its just again the ridiculous values you pitch in without properly accounting load thresholds..I tried all the Heterogenous options and none of them work with my properly calibrated load thresholds..

5. the reduce in temp spikes has nothing to do with how the scheduler would react on certain loads..again, you aren't making sense, since those 2 doesn't co-relate..you are the only one destroying the balance/snappiness by just throwing non-sensical values and not properly accounting/calibrating load thresholds properly..(the fans are another story on this as it co-relates to the temps perfectly)..also if you are solely basing the reaction of the scheduler to temps then you are again doing it wrong..the pursuit of performance has always lead to additional heat..

By default balance already provides snappiness with a little bit of power savings in mind..for a desktop, Heat was never a factor to consider to begin with..

Before you should brought this thread up, you could have spent a week or 2 to read on what the options exactly mean and do..(I won't go further on leading you to where you should read, nor dump mine, I have never been fond of people using my stuff and the same as you do, adjust it for them..)

I have been emphasizing for you to review your numbers on the options on your profile(not to start an argument)..I know your intentions are for the good but seriously if you are going to release something like these it should have been something really thought of and has been calibrated properly, nicely..I would have wanted to help others about this but it would seem I was to hi-jack the thread post..

Sorry, until I see it for me the latency hit still doesn't have a solution

I don't throw random numbers in of course and you still don't get the purpose of this profile.
Except maybe the Class 1 parameters which I'm filling anyway but those are working only for heterogeneous architecture and doesn't make sense for Ryzen AFAIK.

I don't have any idea on what you are referring to about load thresholds.
That heat is not a factor for a desktop is highly questionable.
Anyway reducing heat is the main goal for this profile, it should be easy to get it!
All the parameters are calibrated toward that goal.

It's a work in progress, I'm not making a business out of it.
I don't have problems with thread hi-jacking or whatever, if someone finds them better than the Windows default then I'm happy.

If you don't want to share your work and knowledge with the community fine, it's up to you.
I don't like being only a leecher, I do like to give back.

I'm also not going to spend too much time on this, it's a power plan...
There's no damage in experiment a bit and there's a lot of people like you that are not fond of dumping their stuff for others, which is sad.

I'm always up to critics and sometimes arguments too, feel free to chime in.
Thanks for letting me know better is possible but I'd have appreciate if you were more constructive!

 

[kairi_zeroblade]

Anyway reducing heat is the main goal for this profile, it should be easy to get it!
All the parameters are calibrated toward that goal.

This is an excerpt from MS Techdocs:

Parking a core, by itself, does not save power. Parking a core alters the behavior of the scheduler to target threads at other cores. This allows the parked core to stay idle more (decreasing its power consumption), at the cost of placing additional work on unparked cores (increasing their power consumption). Whether or not this tradeoff results in a more or less efficient system is highly dependent on the processor.

so if your goal is to reduce temps its neither helping..if your goal is to lower power consumption its also not helping..the thing is, Core parking makes use of those single threaded boosts logic to effectively balance workload (this would make sense since on Ryzen the most efficient CCX's are already identified) on cores that would run efficiently rather than using more cores, that means more boosting activity (which is not bad, and is why you are getting those spikes frequently)..

I would be happy to help, so lets begin with the proportions you are setting under core parking parameters..for the number of CPU (count) there is a proper way of assigning values accordingly based on the number of cores (actual ones, not the threads), respectively:

For 12 cores CPU it should be 8%, 16%, 25%, 33%, 40%, 50%, 60%, 68%, 80%, 88%, 95%, 100% ​

For 8 cores CPU it should be 13%, 25%, 38%, 50%, 63%, 75%, 88%, 100%.​

For 6 cores CPU it should be 17%, 34%, 50%, 67%, 84%, 100%.​

For 4 cores CPU it should be 25%, 50%, 75%, 100%.​

For 2 cores CPU it should be 50%, 100%.​

​

this is why using your profile sometimes results in loss of snappiness because your proportions are way off..windows scheduler is too dumb to make a check by itself on the number of core counts..

This is why there should be a separate profile for 6-8 cores and 12-16 cores..

 

[ManniX-ITA]

@kairi_zeroblade

But that's why I've disabled it; it's not helping at all with the CCD temperature.
Maybe on a high core count like the 5950x it could help with power consumption.
The problem is that if even a single core is running with a load higher than 30-50% the whole CCD temperature will jump immediately to 50-60c.
So at the end is doing quite the opposite of what I'm trying to achieve.
It's good for responsiveness as more workload is directed to a single core which has more headroom to boost
But bad for the temperatures when there's low load; a single core boosting will spike the temp while 2-3 sharing low load will not.

Also I have no idea where these thresholds should be set
I can define only one percentage for the parking thresholds not 12.

 

[kairi_zeroblade]

It's good for responsiveness as more workload is directed to a single core which has more headroom to boost
But bad for the temperatures when there's low load; a single core boosting will spike the temp while 2-3 sharing low load will not.

what kind of High temperatures??? do you monitor your temps on a per second basis..overall for me proper calibration reduced temps..idle temps are around 29-31c just browsing, I don't see jumps that would make mine go 60c on a very light to moderate workload..I never expected the drop on temps since cores are just parked and not offline..and my idle frequency is even around 3.2ghz..unlike yours you pulled it way down (which negatively effects performance, the longer it takes to switch from low freq to high freq results in latency itself)

Like I said YMMV..mine is a single CCD and yours is dual CCD..

 

[ManniX-ITA]

what kind of High temperatures??? do you monitor your temps on a per second basis..overall for me proper calibration reduced temps..idle temps are around 29-31c just browsing, I don't see jumps that would make mine go 60c on a very light to moderate workload..I never expected the drop on temps since cores are just parked and not offline..and my idle frequency is even around 3.2ghz..unlike yours you pulled it way down (which negatively effects performance, the longer it takes to switch from low freq to high freq results in latency itself)

Like I said YMMV..mine is a single CCD and yours is dual CCD..

You need a bloated Windows install, yours is probably too much clean
And yes the dual CCD can make a big difference.
I've posted the CCD HWInfo graph earlier, check it out.

I'm not setting the lowest frequency, I let Windows or the CPU decide with a 100% min state.

 

[Latte]

I gave the Balanced v8 a try and WOW...this one is

• snappy,
• no temp spikes anymore (doesn't matter how many tabs, twitch, etc.),
• nice voltages, VID & vCore are idling now properly
• no lag
• nice & partly improved CPU-Z & AIDA scores
• (but stuttering in gaming - doesn't matter, I use windows balanced for gaming)

This is a No-Brainer for desktop usage! THX a lot @ManniX-ITA

Now I will test your ultimate plan for gaming ;-) Will post results soon.

 

[ManniX-ITA]

I gave the Balanced v8 a try and WOW...this one is

• snappy,
• no temp spikes anymore (doesn't matter how many tabs, twitch, etc.),
• nice voltages, VID & vCore are idling now properly
• no lag
• nice & partly improved CPU-Z & AIDA scores
• (but stuttering in gaming - doesn't matter, I use windows balanced for gaming)

This is a No-Brainer for desktop usage! THX a lot @ManniX-ITA

Now I will test your ultimate plan for gaming ;-) Will post results soon.

Nice, you're welcome!

I didn't expected to perform in gaming but also not that would stutter... will check it.
But probably it's a sacrifice that can't be spared...
I use the Ultimate for gaming indeed with Process Lasso switching it.