Crotchety old man needs help overclocking 5800X

[Marco2G]

Hi everyone

 

Frankly I don't know where to go from here. I have tried overclocking this thing and the results make no sense to me anymore.

 

Used to think I was good at this sort of thing but I had my 4th gen i5 for like 8 years. I think I was  left behind :D.

 

My setup:

Ryzen 5800X

MSI X570 MAG Tomahawk Wifi

Noctua NH-D14

Corsair Vengeance LPX 4000MHz 16-16-16-36

NZXT H510

 

Now my initial idea when I bought this was to basically try and get the FCLK to 2000MHz. Granted, I didn't have high hopes for that. I thought if that doesn't work, I'll try 1900 and 3800 MHz on memory and tighten timings instead.

 

When I started fiddling with the settings, I used OCCT to benchmark and test but I think that tool is lying to me.

 

 

Anyway, I went through a lot of settings and whatever I did, it made my benchmarks worse. I noticed that Average CPU Die temp was glued to 90C so I thought it's probably cooling. I had liquid metal in there at that point but I think the issues come from the case and where it was located (in a nook in my desk).

 

Long story short, the absolute best results were achieved with the machine outside that nook and a Curve Optimizer setting of -25, XMP memory profile and an FCLK of 1900. Reached 15592 multithread points in cinebench.

 

The very first time I had this combo, I saw idle temps in the 30C range and under load in cinebench I hardly broke 80C. I did some more testing trying CPU Core voltage offsets in both directions and always, ALWAYS CB scores got worse (CB 23 by the way). So I went back to my best setup (had the profile saved) and now idle is 40 to 50C and in cinebench I reach 85C again.
 

The funny thing is it seems every snippet I read about what is the most optimal config just doesn't hold true for me. 4000MHz RAM and 1900 FCLK is better performance than 3800 MHz and 1900 for example even though having memory and FCLK the same is supposed to get rid of overhead synchronisation... I was also not able to tighten timings... perhaps that would have required more voltage? But which? SoC? DRAM? Both?

Also When I tried getting to FCLK 2000 a few weeks ago, I could fiddle with memory voltage as much as I wanted, stability never got better... and that goes for everything above 1900, not just 2000. It's like there is no logic to how the system reacts to my fiddling.

 

Is there like a step by step guide that will help me edit the right settings? I get the feeling that I lack a sense of how the settings are connected... e.g. when does what voltage need adjusting and so forth...

 

I'm also very close to cutting a hole in the front of the case for a 140mm fan. These temps just aren't fun.

[Chris Pratt]

The temps are fine. 85C under load is nothing to be sweating over. I also think what you're struggling with here is different aspects of performance. Perf is contextual. Running async with higher memory clocks, for example, might get you a better score in a particular benchmark that favors higher clocks, but there's other workloads where the async penalty would be felt.

 

The goal of (non-extreme) overclocking is to optimize performance for the work *you* do. For example, when I was optimizing my 5900X, I found settings that increased my multicore performance but actually cut some off my single core. However, since I'm a developer that's chewing through builds of 50+ projects, that's a tradeoff that's easily worth it. Multicore matters far more for the work I do.

 

I think you need to step back and evaluate the types of workloads that you have and then focus on optimizing for those. Chasing some arbitrary score doesn't actually mean anything, especially since you aren't beating anyone without exotic cooling anyways.

 

[krakek]

1 minute ago, Chris Pratt said:

The temps are fine. 85C under load is nothing to be sweating over. I also think what you're struggling with here is different aspects of performance. Perf is contextual. Running async with higher memory clocks, for example, might get you a better score in a particular benchmark that favors higher clocks, but there's other workloads where the async penalty would be felt.

 

The goal of (non-extreme) overclocking is to optimize performance for the work *you* do. For example, when I was optimizing my 5900X, I found settings that increased my multicore performance but actually cut some off my single core. However, since I'm a developer that's chewing through builds of 50+ projects, that's a tradeoff that's easily worth it. Multicore matters far more for the work I do.

 

I think you need to step back and evaluate the types of workloads that you have and then focus on optimizing for those. Chasing some arbitrary score doesn't actually mean anything, especially since you aren't beating anyone without exotic cooling anyways.

 

In my limited use I found the combination of unlocked PBO for single core and a modest (to 4600) all core static overclock to be a good combination.

[Marco2G]

5 minutes ago, Chris Pratt said:

The temps are fine. 85C under load is nothing to be sweating over. I also think what you're struggling with here is different aspects of performance. Perf is contextual. Running async with higher memory clocks, for example, might get you a better score in a particular benchmark that favors higher clocks, but there's other workloads where the async penalty would be felt.

 

The goal of (non-extreme) overclocking is to optimize performance for the work *you* do. For example, when I was optimizing my 5900X, I found settings that increased my multicore performance but actually cut some off my single core. However, since I'm a developer that's chewing through builds of 50+ projects, that's a tradeoff that's easily worth it. Multicore matters far more for the work I do.

 

I think you need to step back and evaluate the types of workloads that you have and then focus on optimizing for those. Chasing some arbitrary score doesn't actually mean anything, especially since you aren't beating anyone without exotic cooling anyways.

 

I fear you are right... I fear this because as an autist, that's about the worst challenge you could have presented me with :D.

[Marco2G]

3 minutes ago, krakek said:

In my limited use I found the combination of unlocked PBO for single core and a modest (to 4600) all core static overclock to be a good combination.

From my experience, and I probably did something wrong, overclocking the CPU by screwing with the ratio resulted in it permanently being stuck on max Frequency... Wouldn't that mean a lot of additional power usage?

[Chris Pratt]

1 minute ago, Marco2G said:

I fear you are right... I fear this because as an autist, that's about the worst challenge you could have presented me with :D.

It's not all that bad. You just need to find the performance profile for the apps you use, i.e. is it primarily CPU bound or memory bound. If CPU, is single core or multicore more important? If memory, is bandwidth or latency more important? Then, you start there. The decisions you make in one area can be limiting for how far you can push other things. For example, if you have an app that is memory bound and particularly sensitive to latency, you start with the memory timings, and try to get those as tight as possible, even if you have to sacrifice some clockspeed. Then, once you get that locked down, you work on optimizing the next most important thing. 

[krakek]

3 minutes ago, Marco2G said:

From my experience, and I probably did something wrong, overclocking the CPU by screwing with the ratio resulted in it permanently being stuck on max Frequency... Wouldn't that mean a lot of additional power usage?

My motherboard has a function that allows it to switch between precision boost and a static overclock depending on power use and thermals.

[Marco2G]

Frankly I think I need to just give this a rest... I'm going crazy with this and my frustration level is rising. Probably not worth it at all...

[jwwagner25]

24 minutes ago, Marco2G said:

Frankly I think I need to just give this a rest... I'm going crazy with this and my frustration level is rising. Probably not worth it at all...

yep not worth it man. just turn on PBO, no limits, and let the CPU do what it's meant to do on it's own. manually setting a frequency will lock the CPU to it, and it won't be able to ramp up and down the way it wants to. Also regarding your temp concerns, the 5800x is the hottest running of all zen3 CPU's. As far as maximizing performance you could also make sure your ram is running at it's best possible spec with timings and mt/s. Maybe you already did so. take care. 

[Guest]

Certainly don't give up. 

Concentrate on one thing at a time.

 

5800X is a single core complex and as a result runs hot from the transistor density. Anything short of a full custom loop, this chip will just be hot like that. Basically this means it's boxed at its fastest possible frequency already, you are only going to squeeze basically nothing from it OCing. Turn on PBO or just raise EDC and PPT limits and depending on temps should boost better.

 

Memory OC. 

To tighten the timings on your B-Die kit, more memory voltage. If your IMC is 1.08v with XMP on, you can leave it there. More won't be really in necessary, you're already at 4000mt/s. If stability is an issue, then perhaps address this.

 

But really, in my opinion, this rig is fast as fudge with just xmp enabled. You will only eek out .5% here and there. 

 

Need any specific pointers, ask. But one thing at a time is your best approach. 

 

Good Luck!

[Marco2G]

I want to thank everyone for the help. It's giving me warm and fuzzies considering how many questions I have asked in the last weeks on both Reddit and Linux forums with not a SINGLE usable reply. I was about to give up on the internet community but it seems Linus attracts some good people.