5800x (and other Zen 3 chips) PBO settings/Temperature fix

[Mussels]

Oh i missed the DOCP/XMP stuff
Like everyone else on ryzen, you just need to raise your SoC voltage.

XMP raises DRAM voltage and sets the timings, but ryzen CPU's memory controller are only guaranteed to 3200 - above that, you're likely to need to raise the SoC voltage (especially for 4+ ranks)

1.15v is the higest i've needed, 1.10v is enough for 99% of people to max out what they can do (3600-3800 w/ 4 ranks) if the RAM is capable

 

[BetrayerX]

Why do you have llc on if you have PBO enabled?

Just read somewhere (youtube guide I guess) it is more stable. And since all I can do is speak from my own evidence, it has worked fine.

 

[Falcon216]

curve undervolt goes unstable at idle, not at load

You may have more luck with the LLC controls instead of the offset, as the CPU performance is based on requested voltage (VID) which offsets adjust while LLC does not

I keep my system on 24/7, and haven't had any issues with any no/low load/ idle so far.

LLC has no effect (and it shouldn't) on performance or temps as long as you aren't pushing it too high and going over the requested vcore. See attached pics - one with (Asus') LLC3 and one on AMD's Auto spec. Notice how the delivered voltage to the cores is exactly the same (or would be on a longer average), but the voltage pushed out of the VRMs is necessarily higher. Notice also my CPU in R23 is basically unaffected.
I set my LLC to provide a 20-30mv vdroop, nothing more, nothing less. In theory the only thing this effects in the physical world is some extremely minor efficiency gains for the VRMs, and in the virtual world the VID requested by the chip/platform. Again, as you can see, delivered vcore is practically the same and what I didn't include is that performance is the same 4675Mhz.

your Vcore looks kinda weird. But it's hard to tell as you've scrambled the order of all the HWInfo entries. What BIOS are you running?

Ok, I reset my hwinfo layout for you. Dunno why but here it is. BIOS 4403 w/ latest chipset drivers.

you're completely wasting your time with Prime95 Blend.

first I have ever heard that in 15 years. For me, and my experiences, if it passes P95 blend while I'm away at work, it's a functional CPU setting. simple as. This chip passed 12 hours of blend and a further 9 hours of AVX2. It works exactly as I need with no errors. That's all that matters.

VID is not useful. SVI2 TFN should only be relied on if you're under significant load (whether single core or otherwise). On the Prime-Pro you don't have another reliable die sense Vcore to rely on ("Vcore" from the Richtek controller is BS).

I have a few multimeters and can probe hardpoints if you really want, but the board sensors are accurate. What are you trying to say

And just for fun here's a "cheater bench" of r23.
No other open apps except fancontrol, fans full blazing, 16C ambient (with AC running), a true OS idle, and R23 task priority set to High (it defaults to below normal), CPU receiving 1.25V from 1.281 delivered for ~4730Mhz average.

I think my settings, setup, testing, and understanding of the process works?

 

[tabascosauz]

I keep my system on 24/7, and haven't had any issues with any no/low load/ idle so far.

LLC has no effect (and it shouldn't) on performance or temps as long as you aren't pushing it too high and going over the requested vcore. See attached pics - one with (Asus') LLC3 and one on AMD's Auto spec. Notice how the delivered voltage to the cores is exactly the same (or would be on a longer average), but the voltage pushed out of the VRMs is necessarily higher. Notice also my CPU in R23 is basically unaffected.
I set my LLC to provide a 20-30mv vdroop, nothing more, nothing less. In theory the only thing this effects in the physical world is some extremely minor efficiency gains for the VRMs, and in the virtual world the VID requested by the chip/platform. Again, as you can see, delivered vcore is practically the same and what I didn't include is that performance is the same 4675Mhz.

View attachment 267046View attachment 267047


Ok, I reset my hwinfo layout for you. Dunno why but here it is. BIOS 4403 w/ latest chipset drivers.

first I have ever heard that in 15 years. For me, and my experiences, if it passes P95 blend while I'm away at work, it's a functional CPU setting. simple as. This chip passed 12 hours of blend and a further 9 hours of AVX2. It works exactly as I need with no errors. That's all that matters.

I have a few multimeters and can probe hardpoints if you really want, but the board sensors are accurate. What are you trying to say

And just for fun here's a "cheater bench" of r23.
No other open apps except fancontrol, fans full blazing, 16C ambient (with AC running), a true OS idle, and R23 task priority set to High (it defaults to below normal), CPU receiving 1.25V from 1.281 delivered for ~4730Mhz average.

View attachment 267061

I think my settings, setup, testing, and understanding of the process works?

So have you probed Vcore your physical CPU then?

"Vcore" that comes from the board's PWM is less reliable than SVI2 TFN, it's not die sense. You can probe the VRM all you want, that's not the Vcore actually getting to your CPU. SVI2 TFN can be funky at idle depending on setup, but it's the closest we have to a die sense.

You don't have to get defensive over Prime95... ? I didn't say it's a useless test, I said it's useless for undervolt testing. You're just running Blend all core. You don't test Curve Optimizer stability with all core. Run some OCCT core cycle function or the corecycler script and you'll see what I mean. You need as high as possible single core clocks to test curve optimizer settings, and all core just runs some of the lowest possible clocks because that's how boost works.

 

[BetrayerX]

And just for fun here's a "cheater bench" of r23.

What is that, or how do you do it? 1st time I see that term.

 

[DemonicRyzen666]

What is that, or how do you do it? 1st time I see that term.

Taskmanager-> details->right click on CinebenchR23-> set prority -> realtime prority, Close taskmanager Click Run multi core bench. Not see it render on screen suddenly show score at the end.

 

[BetrayerX]

Taskmanager-> details->right click on CinebenchR23-> set prority -> realtime prority, Close taskmanager Click Run multi core bench. Not see it render on screen suddenly show score at the end.

Went from ~15500 to ~15900 ... so 400pts approx. Thanks for the information.

 

[Falcon216]

You don't have to get defensive over Prime95... ?

 

[The King]

Taskmanager-> details->right click on CinebenchR23-> set prority -> realtime prority, Close taskmanager Click Run multi core bench. Not see it render on screen suddenly show score at the end.

If you running Benchmate then you can set this to Realtime permanently so you don't have to go into TM all the time to set it. Just be very careful not to set it to Realtime and do a 10 minute run because you wont be able to stop it or use your PC until the run finishes

 

[Mussels]

LLC works because the CPU thinks it's getting the higher VID value so it doesn't clock down

It thinks its getting 1.20v so it runs at the clock associated with 1.20v - even if LLC makes it receive 1.15v
This is why voltage offsets hurt performance since they reduce VID after a certain threshold


as for the P95 thing, CPUs are different now. Intel for example has different multiplier settings for AVX workloads since they throttle so hard from the heat on them
Modern hardware is different, that's why R23 is commonly used since it's a fully threaded AVX workload and error sensitive

 

[Deleted member 185158]

that's why R23 is commonly used since it's a fully threaded AVX workload and error sensitive

It's just not really that bad of a stress test. It's consistent and reliable way of measuring cpu render performance.

But can you get it to pass some y-cruncher or linpack extreme. Prime95 AVX2 or Intel Burn test maybe?

(I'm promoting burn testing )

 

[Gmr_Chick]

Got a 5600(non X) a few weeks ago and am just wondering what settings to use in Ryzen Master if I want a balance between good performance and temps? Running my CPU on a B550 PG Velocita with a 360mm be quiet! Pure Loop 2 FX AIO.

 

[tabascosauz]

Got a 5600(non X) a few weeks ago and am just wondering what settings to use in Ryzen Master if I want a balance between good performance and temps? Running my CPU on a B550 PG Velocita with a 360mm be quiet! Pure Loop 2 FX AIO.

helluva cooler for 5600 lol

Temps will always be great. Stock 76W so you can try running the regular 65W 88/60/90 PBO limits to see if you gain any clocks in all-core work.

Open up PBO boost override to +200MHz - if 5600G are any indication 4650MHz single shouldn't be too hard just by raising Fmax like that. Then dial up curve optimizer to see if you get more clock improvements in all-core.

 

[pf100]

first I have ever heard that in 15 years. For me, and my experiences, if it passes P95 blend while I'm away at work, it's a functional CPU setting. simple as. This chip passed 12 hours of blend and a further 9 hours of AVX2. It works exactly as I need with no errors. That's all that matters.

What he's saying is that Prime95 is useless for testing curve optimizer because curve optimizer only effectively comes into play under light loads or idle. You can set curve optimizer so that the cpu always crashes 100% of the time at idle and the cpu will still pass a thousand hour Prime95 test.

 

[HD64G]

Got a 5600(non X) a few weeks ago and am just wondering what settings to use in Ryzen Master if I want a balance between good performance and temps? Running my CPU on a B550 PG Velocita with a 360mm be quiet! Pure Loop 2 FX AIO.

78W power limit, +200Hz override and -15 in curve optimizer and you are game! Only RAM tuning is needed after that to get the max. If you want to oc hard go for 90W of max power limit with PBO.

 

[Mussels]

Got a 5600(non X) a few weeks ago and am just wondering what settings to use in Ryzen Master if I want a balance between good performance and temps? Running my CPU on a B550 PG Velocita with a 360mm be quiet! Pure Loop 2 FX AIO.

otehrs gave you good advice, but I personally like to run PBO on auto and disabled and record the values in HWinfo under R23 load and then aim in between them

It's just not really that bad of a stress test. It's consistent and reliable way of measuring cpu render performance.

But can you get it to pass some y-cruncher or linpack extreme. Prime95 AVX2 or Intel Burn test maybe?

(I'm promoting burn testing )

It's consistent, easy to source and run - and it uses AVX.
AVX is key with modern CPU's since for a long time people didnt test with it, and then when games used AVX their overclocks went to absolute shit
Witcher 3 used it (lightly) and would either make systems unstable or force them to downlock with the AVX offsets, leading to a lot of people complaining about the game being unstable, buggy, blahblahblah


Unlike other programs because it's just the one test with an optional time limit you get a lot less user errors since some people would just fire it up without configuring first and assume alls well

 

[Deleted member 185158]

otehrs gave you good advice, but I personally like to run PBO on auto and disabled and record the values in HWinfo under R23 load and then aim in between them


It's consistent, easy to source and run - and it uses AVX.
AVX is key with modern CPU's since for a long time people didnt test with it, and then when games used AVX their overclocks went to absolute shit
Witcher 3 used it (lightly) and would either make systems unstable or force them to downlock with the AVX offsets, leading to a lot of people complaining about the game being unstable, buggy, blahblahblah


Unlike other programs because it's just the one test with an optional time limit you get a lot less user errors since some people would just fire it up without configuring first and assume alls well

Yeah, Y-cruncher is a lot more intense and consistent and easy to run also.

But I'd be willing to bet that your CB R23 stability testing won't be stable for a Y-Cruncher. At least this is the experience I have, seems quite a bit hotter than CB R23.

 

[The King]

Yeah, Y-cruncher is a lot more intense and consistent and easy to run also.

But I'd be willing to bet that your CB R23 stability testing won't be stable for a Y-Cruncher. At least this is the experience I have, seems quite a bit hotter than CB R23.

Y-cruncher is also way harder on the RAM and VSOC compared to R23 which is mainly focused on the CPU

Its a great test to find out quickly if your RAM OC and VSOC is unstable. Should run at least 3-4 cycles back to back 2.5B some run it for hours to check system stability.

 

[Deleted member 185158]

Y-cruncher is also way harder on the RAM and VSOC compared to R23 which is mainly focused on the CPU

Its a great test to find out quickly if your RAM OC and VSOC is unstable. Should run at least 3-4 cycles back to back 2.5B some run it for hours to check system stability.

I'd like to see just a pass at 1b from a lot of braggers I've seen running around.

Now don't get me wrong, I know this isn't an indication of gaming stability. Which is the backing argument of running CB R23 and calling it gaming stable, cause AVX.

Nothing wrong with a burn in. Heat cycles are good for some thermal compounds like AS5 too!

 

[Falcon216]

heat cycles are good for some thermal compounds!

If Arctic could make an AS5 update I'd buy it in a heartbeat. I'll gladly trade a couple C for a "apply it once, forget it for 8 years" reliability. Thermal Grizzly makes fantastic OC enthusiast products, but those products just aren't as dependable as a "lesser performing" TIM.

I'd like to see just a pass at 1b

I still think motherboards are a huge culprit in behavior anomalies. While I understand a Ryzen 5000 should -not- work stable at -30 CO all core, my setup doesn't work -the way it's supposed to- unless I have it set to -30 CO.
I've been having a hard time looking for absolute top-end benchmark run results, as such I have zero real statistics to compare the quality of my 5800X with.
With that said, my chip just cannot do R23 at 4.95Ghz. I attempted a 14C ambient run at (fully reset/manual control) 1.365v (SVI2 input) and it ran for a whole 7 seconds before a full system crash. WHEA suggests a cache-level fault (i.e. too low voltage for the entire CCD), but I'm suspicious of temp-limit shutdowns due to the op-temp of 88C, who can even know what the real hotspot spikes are.

And to reiterate some previous posts - I still haven't had any idle/low load/light MT issues with my tested setting of 135PPT/100TDC/125EDC at -30CO. I'd expect a real world scenario of multiple RAM heavy browser + network multitasking + virtualization usage + mixed-load gaming to create SoC related errors, as I've seen in literally every PC pushed "too hard" with concurrent low/light MT RAM heavy loads in the past 15 years of my experience. These issues have not happened (yet).

 

[The King]

Run Y-cruncher 2.5B at least 4 cycles with -30 all core if that can pass then you know your CO is good.

If not then its unstable. Many pass R23 with -30 all core 9/10 setups wont pass Y-cruncher with -30 all core.

 

[Falcon216]

Hmm but what if I don't need a 101% y-cruncher stable system?
Where's the functional limit for a more real, typical use case?

 

[The King]

Hmm but what if I don't need a 101% y-cruncher stable system?
Where's the functional limit for a more real, typical use case?

Simply put that means your system is not stable with your CO settings and could crash when the RAM or VSOC is put under stress under normal usage like gaming etc.

Running R23 for hours does not mean your system is stable at all. Has I pointed earlier its only stresses the CPU and not much else. So good CPU test bad system stability test.

I can run -30 All core in R23 but if I run that while gaming after a few minutes reboot or BSOD.

 

[tabascosauz]

Hmm but what if I don't need a 101% y-cruncher stable system?
Where's the functional limit for a more real, typical use case?

Unstable = unstable..........

However, if you don't want to spend too much time, you can run corecycler on just default settings (ie. not editing config file). Let it cycle through all the cores 2 or 3 times (about 6 minutes per run per core), and if error free then it should be good enough for games and daily tasks. My full testing routine for CO takes like 10 hours for 2 cores........so sometimes I'd just prefer doing default config. Didn't encounter noticeable daily stability issues with it on 5900X, 5700G (x2) and now 5800X3D.

GitHub - sp00n/corecycler: Stability test script for PBO & Curve Optimizer stability testing on AMD Ryzen processors

But like I said, it's more "should" than "will", and I'd still get in some more comprehensive testing down the line.

 

[Dr. Dro]

How bad is the EDC bug on recent AGESA (newer than 1.2.0.3 C) hitting you guys? It's a major nuisance to me, like, an itch I can't scratch off. My CPU could really use the extra juice, it makes PBO worthless.