5000 Series Undervolting

[The Asgard]

Ahh yeah that's what i thought, havin a tinker with it.

Say i make it to -15 all cores and it crashes how would i know what core is causing it?

You would go to single core and back off each core individually. Depending if its a light application crash, Prime95 should show you which thread is falling over. Usually the Gold and Silver cores should be fine ( Ryzen Master will tell you this ) so you can discount those. It all depends how much time you want to play with it. If you not that bothered, just back the all core value.

Testing is key. Cinebench R23 and 20 Minute Prime 95 with large FFT. Hwinfo to monitor the cores and WHEA errors. Oh and a little book to make notes

 

[Loxa]

You would go to single core and back off each core individually. Depending if its a light application crash, Prime95 should show you which thread is falling over. Usually the Gold and Silver cores should be fine ( Ryzen Master will tell you this ) so you can discount those. It all depends how much time you want to play with it. If you not that bothered, just back the all core value.

Testing is key. Cinebench R23 and 20 Minute Prime 95 with large FFT. Hwinfo to monitor the cores and WHEA errors. Oh and a little book to make notes

I am guessing manual OC best way to get say all cores 4.2-4.5 with 1.25v as an example, this curve optimizer is not really what i am after. Quick change to 10 etc and still cracking at 89c in Scavengers games. It's not the fact i am worried about the temps it is the fan noise is unbearable.

I am just looking for a way to get the voltage as low as possible with stock performance.

 

[BHS1975]

That's definitely on the edge. I would have probably back off to -0.050. LLC could help with stability. Curve Optimiser will give you the biggest gain. The offset is just icing on the cake.

I got is at -0.05 offset now and got the LLC at 1 to keep the vcore down under load. The LLC levels on my ASUS B550 is the opposite of yours with level 1 being the lowest amount of compensation for vdroop.

 

[The Asgard]

So I have settled on PBO Limits disabled, CO all core -30, negative offset -0.1625, LLC Level 5, plus a 150 Boost. This translates up to 4,750 multi core boost and Single core boost 5.1. R23 multi load 1.119 volts. Tweaks seem marginal now and I don’t want to risk an unstable system. It seems a very forgiving chip.

No crashes, no WHEA errors.

 

[The Asgard]

I am guessing manual OC best way to get say all cores 4.2-4.5 with 1.25v as an example, this curve optimizer is not really what i am after. Quick change to 10 etc and still cracking at 89c in Scavengers games. It's not the fact i am worried about the temps it is the fan noise is unbearable.

I am just looking for a way to get the voltage as low as possible with stock performance.

The curve optimiser should do what your after while boosting due to the extra headroom. However if your struggling with heat and don’t mind dropping the frequency down. If you setup Eco mode it will adjust to what you set. 45 watt, 90 watt. Try these settings and it should adjust to the power settings

 

[Loxa]

The curve optimiser should do what your after while boosting due to the extra headroom. However if your struggling with heat and don’t mind dropping the frequency down. If you setup Eco mode it will adjust to what you set. 45 watt, 90 watt. Try these settings and it should adjust to the power settings

A manual overclock of 4.5 all cores with 1.2650v yeilds me decent temps tbh at the moment. (89c to 62c ingame)

 

[pc-guy]

Guys. Seriously setting LLC and using CO is counter intuitive.

What LLC does is to force motherboard to pump more voltages to the core to prevent the Vcore from dropping.

what CO is doing is to actively under volt according to load.

So one algorithm cannot co-exist with the other. For CO to function properly All LLC settings should be left alone.

however CPU voltage offset can work in tandem with CO but I see a problem with using such method as during all core load the vcore might be too low to have stable results as you are offsetting significant amount of vcore (remember vcore is dynamically managed by cpu and CO algorithm).

I strongly recommend people just sticking to using PBO (manually set EDC TDC and PPT to suit their needs) and have CO settings rigorously tested with MT and ST tests.

 

[The Asgard]

LLC on the MSI pumps less volts the higher the level. I wanted the maximum transient voltage to come down. My peak voltage is 1.3 ish down from 1.5. It gives me stability which is why I use it. Each chip is different. LLC is really important when undervolting.

My simple understanding of CO is that it’s just per core VID lookup table against frequency. AMD limited the setting to 30 so I’m using the offset to push it further.

All the settings are tools to play with. There are no rights or wrongs if it gives you the desired results.

PBO Limits adds an extra 12’c to my load temps for very little gain in performance. Yes I can tweak them but I find this method easier and it works for my chip.

Hwinfo, load tools, a notepad and time gives me the most efficient processor possible. That’s my goal. My system is quiet, consumes less power, stable and fast. Seriously!

 

[Aretak]

So I have settled on PBO Limits disabled, CO all core -30, negative offset -0.1625, LLC Level 5, plus a 150 Boost. This translates up to 4,750 multi core boost and Single core boost 5.1. R23 multi load 1.119 volts. Tweaks seem marginal now and I don’t want to risk an unstable system. It seems a very forgiving chip.

No crashes, no WHEA errors.

Have you tested it in OCCT with a rolling stress test? Set up like this:

I thought my 5800X was stable with an all-core curve optimiser offset, and wasn't getting any errors or crashes in other stress tests that load several/all cores, but OCCT quickly dispelled that illusion once set up as so. It's really good for finding instability on a per core basis and will tell you which cores are causing problems. I needed to back off several of mine to stablise them. From what I've gathered elsewhere, it takes a real golden sample to just set and forget a -30 offset on every core. Maybe you have one, but it's worth checking.

 

[The Asgard]

I have used OCCT in past but my general test is Prime, Linpack and Cinebench for getting an indication of load stability. Then general use which consists of Teams, office Apps, Development and Gaming. If I have any issues I back it off or tweak. The machine is used daily for work and so far it’s solid.

 

[pc-guy]

Have you tested it in OCCT with a rolling stress test? Set up like this:

that's a nice feature, i need to update my OCCT.

 

[The Asgard]

So I after a week of testing using the machine about 14 hours of day. Multi loads, lots of idle, games , compiling I have finalised these settings. I did back off the offset due to a single random reboot while idle. Really pleased.
PBO limits disabled
C0 -30 All Core
Offset -0.1375
LLC Level 5
Boost 200
Running 4 stick of 8gb ram at 3333 @ 15,17,17,36 1.32 volts.

So depending on the what's happening it will multi core boost up to 4750. R23 sits around 4,700, Single core boost 5,150.

https://photos.app.goo.gl/EF3adFJbogsg5LCV8

https://photos.app.goo.gl/SEx3bMKjAu6KJSbx9

 

[Joxeon]

Are the benchmark scores in line with those clocks though? as sometimes CO can cause clock stretching.

 

[Loxa]

Whats is a safe MAX voltage on the 5950x for a manual all core overclock (24/7)? 1.3v or

 

[DarkBahamut]

Are the benchmark scores in line with those clocks though? as sometimes CO can cause clock stretching.

You can see in the screenshots that the maximum effective clock was 4613 MHz, so that's what the 'real' clocks/performance are. CO usually is fine, but the negative offset will definitely be causing clock stretching and HWinfo shows it.

 

[The Asgard]

The max clocks are upto for a few cores 5150 and all core can very upto 4700,4750, I think I saw that. With disabling the PBO limits I think the chip is working within certain conditions, probably a lower TDP. With high sustained load the performance will be down with limits disabled. I'm loosing about 1-1.5K from R23 compared with enabling limits. Enabled limits the power goes up, temps go up ( 12'C+ ) which is not what I want. With normal mix load usage I should see good gains for a given watt.

Its interesting trying to make the chip work on the edge of its operating window, its handling some really low voltages.

With a +50 boost and a -0.1125 offset before I installed all my development components I was seeing 22,872 in R23. I am now running a 200 boost.

Trying to tweak the memory but its nothing special, Hynix. Running 3200 , 14,17,17,17 @ 1.35. Its Corsair RGB 3000 goodness, love the lights! Got some 8-Pack B-Die but its only 16 GB.

 

[The Asgard]

Whats is a safe MAX voltage on the 5950x for a manual all core overclock (24/7)? 1.3v or

Single core boost can go up to 1.5 I believe. Before I tweaked my CPU I saw max voltage of about 1.45 ish, depends what the chip is doing really. CO and negative offsets will bring this down.

 

[pc-guy]

All core manual OC shouldn’t go above 1.3V for 24/7 usage 1.25-1.3v. Putting 1.3V under all core workload you are asking a lot from the cpu.

because these CPUs’ EDC is 140A even at 1.3V you are exceeding the package power significantly.

1.25V would be my recommendation for 24/7 which happens to be the save voltage that ryzen clock tuner voltage as well. It is worth noting that even at 1.25V you can see 175w power draw under certain stress tests but generally this setting will give 140w package power so make sure you have a cooling solution that can handle that. 1.1V which is another steady state volt that ryzen clock tuner works to basically ensure the package power doesn’t exceed 140w under all kind of loads but most stress tests you will see around 100w.

 

[Grim5]

Single core boost can go up to 1.5 I believe. Before I tweaked my CPU I saw max voltage of about 1.45 ish, depends what the chip is doing really. CO and negative offsets will bring this down.

When I set my 5950x to stock, voltage at single core max load = 1.45v, voltage at multi core max load roughly 1v +-0.05v

So that's the stock behavior. If I enable PBO, single core voltage is exactly the same, multi core max load voltage jumps up to 1.29v

So as far as I'm concerned 1.29v is safe for overclocking because when I give the CPU free reign to boost as high as it wants to it will go up to 1.29v

Highest I've tested for all core manual overclocking is 1.25v, I reached 4.6ghz all core with that

 

[The Asgard]

The highest voltages you see with Hwinfo are just transient. Sustained core voltages are really low. If you look at my all core Prime95 sustained voltages it’s just 1 volt. General operation is less than a voltage on average. These chips can seriously pull some load hence keeping the voltage down.