5950x at stock, single thread Prime95 fatal rounding error on core 1

[str]

I was initially impressed with my B0 stepping 5950x but it turns out there is an issue with it so I looked into testing with CoreCycler.

There is no warranty because I bought it second hand in good faith so I'm looking to fix it while still being able to maximise it's potential.

It's in a Phanteks Shift 2 with Shift 1 Air mesh side panels and custom water cooling on the CPU and GPU. The motherboard is the Asus B550 ITX with 64GB DDR4. I have two 120mm rads with the fans configured to result in a coolant Delta-T of less than 5c when the CPU is under load and the GPU is idling.

I made a diagnostics prototype to learn more and so far the evidence is that when only core 1 is active (meaning CCD2 is idle/cool) the CPU temp is low enough to cause it to boost beyond it's capabilities resulting in Prime95 ending with a fatal rounding error.

Anyone fixed theirs or have an idea what I should do about it?

 

[jigger]

Run at least two instances of prime 95

 

[str]

Run at least two instances of prime 95

Yep, I already tried loading other cores at the same time to cause the problem to disappear. The problem appears to be an edge case single core boost fault.

 

[str]

I managed to get useful data from the last run where it failed. It wanted to boost to 5GHz at 1.469v and the graph shows the effective clock was on it's way up but evidently it wasn't stable.

The problem appears to be specifically a situation where the core speed is boosted then a brief idle period at 3.6GHz then it gets boosted again but the core voltage hasn't increased enough to support the boosted clock speed.

 

[welshrat]

Can you not use a positive offset on just that core of say +5 and see if that eliminates it ?

Should all be accessible from the bios

 

[str]

Can you not use a positive offset on just that core of say +5 and see if that eliminates it ?

Should all be accessible from the bios

Yep, I want to learn about that, I don't know much about the configuration options yet. I went the diagnostics route as I couldn't find discussions on the root cause of this type of issue.

 

[welshrat]

I suspect you have a bit of bad CPU but it may not be a problem if adding a tiny bit of voltage to that core solves it.

 

[Muzzza]

Just as a side note, I find having 64gb on my 5950x means it doesn't perform at as high a clocks as others.

 

[str]

I suspect you have a bit of bad CPU but it may not be a problem if adding a tiny bit of voltage to that core solves it.

I will figure out how to add voltage like you said and see how it goes.

Yep it seems to be bad but the previous owner didn't know about it as he had configured an offset of -0.2v and was happily using it like that.

I don't know enough to know if I'm wrong or not but I thought because it is the preferred core when single core boosting occurs, and I've seen it boost over 5.1GHz, it'll boost even higher if I increase the core voltage because it thinks it can.

Also, people have reported this type of issue albeit via issues like lockups/resets/WHEA errors rather than it being specifically reproduceable and many have been able to RMA it back to AMD and received a brand new replacement which presumably means it isn't an easy fix.

I was thinking once I had created a diagnostics prototype I could learn how to achieve a decent all cores overclock.

Just as a side note, I find having 64gb on my 5950x means it doesn't perform at as high a clocks as others.

Thanks. I will install 32GB when I know more about the configuration side.

 

[str]

I managed to get the HWInfo64 data sampled at 200ms without preventing the single core boosting fault. I don't know how reliable the data is but presumably it is useful.

Could be interesting that the Core VID readings are weird for example core 0 at 4GHz had 1.469v at the same time core 1 at 5.05GHz had only 1.431v which was when Prime95 failed.

Also saw a YT video mentioned on here for configuring the BIOS options which looks to be what I need to make progress.

 

[str]

It didn't take much to learn how to do CO +5 core 1 and the Core VID data has improved. No failure so far with Prime95. Core VID is now minimum 1.469v when core 1 boosts to 5GHz and if still boosted the next sample is 1.475v or 1.481v.

 

[str]

It now single core boosts to 5250 with Boost Override CPU at 200 and CO +5 on core 1 is Prime95 stable. I maxed out some of the DIGI+ VRM settings but still it needed the +5 on core 1.

I also configured ~4.4GHz all core (CB23 ~28K pts) with PPT 395 (~226 actual max), TDC 160 (board max) and EDC 148 (lowest for the target clock speed).

Other BIOS settings are at defaults with the exception of a CPU VID minimum negative offset, enabling Resize BAR Support and DOCP for the XMP 3600 configuration.

I'm impressed with this 5950x again.

 

[Demon]

It now single core boosts to 5250 with Boost Override CPU at 200 and CO +5 on core 1 is Prime95 stable. I maxed out some of the DIGI+ VRM settings but still it needed the +5 on core 1.

I also configured ~4.4GHz all core (CB23 ~28K pts) with PPT 395 (~226 actual max), TDC 160 (board max) and EDC 148 (lowest for the target clock speed).

Other BIOS settings are at defaults with the exception of a CPU VID minimum negative offset, enabling Resize BAR Support and DOCP for the XMP 3600 configuration.

I'm impressed with this 5950x again.

Sounds good!

I'm sorry to say whilst I keep buying Asus motherboards, I have found their out of box settings (clear CMOS) to sometimes be stupid.. My 3900X had a stupidly high vCore set by the motherboard as default, I 'assumed' there is a standard AMD value that should be applied when that is set to 'auto'... but it seems not. Each manufacturer likes to have different default values for allsorts to eek a bit more out of things, so the concept of 'stock' IMO doesn't really apply!

 

[str]

I know what you mean. Some of the readings in HWINfo64 were also confusing.

I tested some things with Ryzen Master and every time it required a reboot to change settings, I had to go into the bios because Ryzen Master had changed the CO option to Auto when I needed it on Per Core. I wasn't able to figure out how to set CO +5 in Ryzen Master.

I also saw a post recently saying the peak core voltage reading in Ryzen Master and the corresponding entry in HWInfo64 are the reliable core voltage readings.

It seems evident enough that my setup is running at decent temperatures and stable so I'm reasonably confident it's safe to use.

 

[tamzzy]

Does AMD RMA need proof of purchase? Whilst the 5950x remains purchasable, there's a cheeky way to do it ofc. Ethical not so much

 

[str]

I was too inexperienced to know how bad it was to max out the motherboard with TDC at 160A for all-core activities.

I also investigated the previous owners configuration of VDDCR offset -0.2v with the rest of the BIOS settings at defaults. It actually causes significant negative clock stretching with all-core target clocks at 4.5GHz but the effective clock is close to 800MHz lower at ~3.7GHz and the single core boosted target clock is around 5GHz but again it has negative clock stretching resulting in the effective single core boost clock being ~500MHz lower.

Evidently simply undervolting a 5950X CPU needs to be avoided.

I also hadn't fully tested core 3's 5250MHz max clock speed and of course it isn't stable. It rarely goes above 5050MHz but it's guarenteed to eventually crash when it does.

It soon became evident that an optimal setup might be to limit the motherboard VRM usage to 80% which means reducing the max PPT/TDC/EDC by 20%.

For my Asus B550 ITX board I now have PBO at 180/128/152 and a side effect of doing so is the "default" CPU core voltage (VDDCR) lowers by around 0.08v which is actually a bonus because it now runs cooler especially when single core boosting. Another side effect is the single core boost was limited to ~4.7GHz however using a very small negative VDDCR offset of -0.03125v helps it to boost reliably to ~4.9GHz again.

The all-core target clock and effective clock are now the same at ~4.1GHz max 65c (custom water cooled) when running CB23 = ~27K pts and the single core boost is also without negative clock stretching with an effective clock speed of ~4.9GHz max 60c when running CB23 = 1613 pts.

I also reduced the maximum PBO clock speed to 4.925GHz with a negative CPU boost clock override. It made sense because that speed is close enough to the upper limit of this overclocking system and now there is no need to use CO +5 for core 1 to fix the Prime95 failure at 5GHz+.

Does AMD RMA need proof of purchase? Whilst the 5950x remains purchasable, there's a cheeky way to do it ofc. Ethical not so much

I'm tempted to buy a new one for myself and install this one in my daughters PC. She is learning programming and it would be a decent upgrade from a 1700X.

 

[str]

I ordered the brand new 5950X which got me motivated to figure out how to extract more useful data from the faulty 5950X so I can compare it to the new one.

The motherboard did not actually supply these voltages but it's what each core requested for each clock speed:

Has anyone else managed to extract information like this?