5800X PBO & CO

[Yaayuh!]

What’s your Curve Optimizer and PBO settings?

Just diving into one (from a 3900X).

What’s the best way to find and tweak the weaker cores?

 

[VersionMonkey]

What’s the best way to find and tweak the weaker cores?

I used the corecycler script https://forums.overclockers.co.uk/posts/34947981 for my 5950X. I would guess it'll also work for a 5800X

 

[Yaayuh!]

I used the corecycler script https://forums.overclockers.co.uk/posts/34947981 for my 5950X. I would guess it'll also work for a 5800X

Thanks, i'll take a look.

 

[Yaayuh!]

I'm finding there are several variables I need to pin down first. Namely Vdroop setting and PBO limits to either Auto (default) or Motherboard.

Might take some time.

 

[Yaayuh!]

Core 6 is making my eye twitch…

 

[GregI]

@Yaayuh! Is that stock boost clocks without PBO? That's pretty insane for a 5800x. My 5950x stock only hits 5Ghz on 2 cores... rest 4.9-4.925.

Some info for core cycler. I did a decent amount of testing with it and I recommend 2 setting changes:

1: Set the program to ycruncher (via config file) and never look back. This should be default imo and the dev should make it so. In my testing it will boost approx 25-50 Mhz higher than prime95 and it will also make sure your CPU/BIOS is playing correctly with voltages.

Here's a 50 hour Prime95 run: https://i.imgur.com/sYeDaLE.png

In 50 hours it never once maxed out the voltage at 1.500v (1.494v max - sv12 tfn). I did another 20 hour run to verify: https://i.imgur.com/ilr4cUp.png

After a couple hours, ycruncher not only got me higher boosts to check stability, it also got me to 1.500v to ensure the vcore cap was working as expected: https://i.imgur.com/HaTkwJA.png

I know it says 15 hours but it got to 1.500v quick (3 hours I think).

2: In the config file there is an option to stop the test if an error is detected. If you don't enable it the test just continues and skips that core each iteration. Unless you're reading the output continually you can miss this. Easier to spot on a 5800x but on a 5950x there are so many cores to look at you get lost lol.

 

[Yaayuh!]

Is that stock boost clocks without PBO? That's pretty insane for a 5800x. My 5950x stock only hits 5Ghz on 2 cores... rest 4.9-4.925.

Nope, that's PBO enabled, limits on Auto (Default), +150MHz and -15 CO.

Cinebench gives me 4.625GHz -4.675GHz constant across all cores. Which is higher than my 3900X single core boost ever achieved

Stock I get most cores boosting to 4.85GHz.

Some info for core cycler. I did a decent amount of testing with it and I recommend 2 setting changes:

1: Set the program to ycruncher (via config file) and never look back. This should be default imo and the dev should make it so. In my testing it will boost approx 25-50 Mhz higher than prime95 and it will also make sure your CPU/BIOS is playing correctly with voltages.

Here's a 50 hour Prime95 run: https://i.imgur.com/sYeDaLE.png

In 50 hours it never once maxed out the voltage at 1.500v (1.494v max - sv12 tfn). I did another 20 hour run to verify: https://i.imgur.com/ilr4cUp.png

After a couple hours, ycruncher not only got me higher boosts to check stability, it also got me to 1.500v to ensure the vcore cap was working as expected: https://i.imgur.com/HaTkwJA.png

I know it says 15 hours but it got to 1.500v quick (3 hours I think).

2: In the config file there is an option to stop the test if an error is detected. If you don't enable it the test just continues and skips that core each iteration. Unless you're reading the output continually you can miss this. Easier to spot on a 5800x but on a 5950x there are so many cores to look at you get lost lol.

Yeah, I'm just using core cycler to find the weaker cores at the moment, so short runs. Along with Cinebench and CPUz for checking performance and ensuring I'm not clock stretching and RealBench for general stability.

I'll come back to core cycler once i've found the rough sweet spot. But in general, I've got to the point where it becomes fiddly to get more performance, so I might just leave it.

It's interesting how PBO limits effect CO. With standard default limits, pretty much somewhere between -15 and -18 is the limit. But push the PBO limits and you can start using up to -30. I've decided that the former is the easier to deal with as there's less variables to contend with.

 

[Audioboxer]

Small tip for OP and anyone else, in HWINFO settings turn on snapshot CPU polling for more realistic recording of your frequencies

You can then run boosttester to see what your cores are actually boosting to

https://github.com/jedi95/BoostTester/releases

It runs an incredibly light load, per core, in a loop to make each core boost as high as it can. This will be recorded correctly in HWINFO with Snapshot CPU Polling enabled.

You then want to make sure your Core clock and Core effective clock maximum are pretty much the same.

As you can see from above cores do not magically boost to exactly the same frequency like what HWINFO may show without snapshot polling. Think of it like the "requested frequency" isn't what your cores might actually reach. So HWINFO might show it requested 5050mhz or even 5100mhz for every core, but what they actually hit is under that.

I've got a +50 AUTO OC above and as you can see only a couple of my better cores manage 5.1ghz. The default max boost for a 5950x is 5050mhz.

CCD1 is normally always better than CCD2 on Ryzen CPUs with 2 CCDs.

As for PBO, remember to take your time to properly test the stability of each core. Sadly it can take many hours but it's honestly needed

That's my cores on a 5950x. Yes, Core5 you read right. 6 hours in testing that core on its own on a loop brought an error at -2. It's one of my best cores so not too surprised at the result, you tend to find your better cores of which there are 4 on a 5950x can't go as low. But -1 is still quite the outlier from the rest.

95% of the time anyone saying they are -30 all core stable are talking nonsense. Just because you can boot a game and play it for a few hours does not mean your cores are stable. Especially since gaming can be very light on the CPU.

Corecycler as posted above is really good to run, OCCT is quite good as well but limited to 1 hour at a time in the free mode. So I mostly recommend Corecycler.

Y-cruncher is the best at testing overall CPU stability once you think you've got a curve locked in. It hammers the CPU on multicore loads. It also puts load on the IMC so it will hit your RAM too.

 

[Mesai]

I just set PBO2 -25 all core and let it figure it out. I'm sure I could get better results with a more granular config but for work and gaming I likely wouldn't see any difference.

 

[Yaayuh!]

Thanks @Audioboxer, very helpful!

What PBO power limits are you using with those curves? The power limits change what's required on curve, wildly, mainly TDC it seems.

 

[Audioboxer]

Thanks @Audioboxer, very helpful!

What PBO power limits are you using with those curves? The power limits change what's required on curve, wildly, mainly TDC it seems.

Above for the core boosting example is just PBO enabled but on AUTO. It's something like 142/95/140. Very light ST boosts will be highest with PBO reigned in a bit. I wanted to show best case example above with core boosting. Even under this scenario I am not getting 5050 on all cores (AMD default max for a 5950x), let alone the 5100 I'm pushing for with AUTO OC +50hz.

My daily is 270/168/220. This produces a CB23 score of 31200+.

PBO is all about balance between ST and MT. The more you try to push MT, for example, for running CB23, the lower your ST score might be.

Cooling obviously plays a role as well. Majorly so with CB23. Competitive OCers will be opening all the windows in the house, moving the PC next to an open window and doing CB23 runs in the middle of winter because even a degree or two drop in ambient can result in hundreds of points. Just keep that in mind if you see any really good CB23 scores! lol

Lastly, something you can play around with is CPU telemetry. On my MSI bios it's CPU VDD full current and offset. You have to be mindful that what telemetry does is trick the CPU into thinking its power draw is underreported so it boosts higher. It's not dangerous, but it can lead to instability if you don't redo your curve and PBO settings.

I use 150ma for full current and 145ma for offset.

For example here if you look at the max wattage it says 208w. This is not correct. Under HWINFO if we were to check PRD, power reporting deviation, under CB23 load, it would likely be around 80%. This means its underreporting by around 20%.

20% of 208 is around 42, which would be correct. The real power draw CB23 can hit with a 5950x is as high as 250w and even a bit beyond.

CPU telemetry is very useful for pushing frequency higher on chips, especially if you can cool it. FIT limits are still observed, the CPU isn't going to unsafe voltages or anything. It's just being tricked into thinking its power draw is a bit lower so it allows a higher frequency to be requested. In most cases this is stable because these chips can boost better than what AMD lets them. Many of the OCers out there posting the highest scores in benchmarks will be using telemetry.

 

[GregI]

How long should I be running core booster to make sure the boost clocks are at their max obtainable? My Core 05 has hit 5049 Mhz which is ~50 Mhz higher than core cycler/y cruncher but my other cores are slightly lower by 75 or so (comapred to ycruncher). Only tested for a couple minutes though.

 

[Audioboxer]

How long should I be running core booster to make sure the boost clocks are at their max obtainable? My Core 05 has hit 5049 Mhz which is ~50 Mhz higher than core cycler/y cruncher but my other cores are slightly lower by 75 or so (comapred to ycruncher). Only tested for a couple minutes though.

An overnight with all cores tends to be a good idea. Then if it's OK what you want to do next is go here

Add every core apart from your best 4 if 2 CCD or best 2 if 1 CCD to coresToIgnore in the config.ini for Corecycler. Then test them overnight. Yes, give those best cores alone the same overnight you did for them all.

Unfortunately it's a slightly time consuming process but it's best to be safe. Errors can sometimes pop up hours in on a single core, especially if you're moving the curve +/-1 to get the most performance you can.

The first run with all your cores overnight will give a good indication and catch most issues. The focussing on your best 2/4 cores is really for when you know the rest of the weaker cores are likely fine where they are but you want to micromanage the best cores. They tend to be the ones that cannot go as high as the weaker cores, hence me ending up with -1 on one of mine, let alone -16 on another. Compared to my weaker cores getting near or right to -30.

As for finding out what your best cores are, HWINFO will tell you or for an even easier visual representation, AMD Ryzen Master

Gold star means best core, silver circle means 2nd best.

Do keep in mind that Ryzen Master likes to name the cores 1-16, whereas the BIOS and Corecycler name them 0 to 15. So in terms of the above image, C03 in Ryzen Master is really core 2 in the BIOS/Corecycler.

 

[GregI]

Sorry I was referring to boost tester.

This is what it gives me: https://i.imgur.com/ZQFGp3s.png

I had higher boosts on some cores with ycruncher but then again I wasn't using cpu snapshot then so tough to do a direct comparison.

 

[Audioboxer]

Sorry I was referring to boost tester.

This is what it gives me: https://i.imgur.com/ZQFGp3s.png

I had higher boosts on some cores with ycruncher but then again I wasn't using cpu snapshot then so tough to do a direct comparison.

Ah right, usually a couple of cycles is enough, so maybe 5 minutes? Even a minute or two is probably enough.

And yeah, CPU snapshot is telling you the actual frequency, not just the frequency requested. So it's not a valid comparison.

With a +100 AUTO OC I can get my cores to all display 5100 in HWINFO if left long enough and not using CPU snapshot. As you can see above even at +50 AUTO OC some of my cores barely get to 5050 let alone actually hitting 5100.

Effective core clock is usually still decent at telling you the actual frequency even with CPU snapshot turned off, but it's best just to turn it on.

 

[Scougar]

Watching for updates and reading with interest. Just got a 5800X B2 Stepping, so keen to have a little fun in my ITX case lol. (LFii cooler)

 

[luismenendez]

Watching for updates and reading with interest. Just got a 5800X B2 Stepping, so keen to have a little fun in my ITX case lol. (LFii cooler)

nice, keen to see your results!

 

[Scougar]

nice, keen to see your results!

Will run the whole thing through the stability checker later. I thought it was only one iteration of each core, turns out I was wrong lol. My bios does auto-volts as well, not sure how much I trust it.