Help with overclocking 192Gb RAM (stock 7950x) would be appreciated!

[MoravianLion]

I’ve been struggling for a while with my 192Gb DDR5 RAM overclocking. I managed to get it fully stable at 5600Mt/s, but only with “default” option of 38,38,38,96,144 (what MSI BIOS offers for manual OC), which is far from ideal. I have had y cruncher running for 12 hours, having my PC to idle for long period of time and was also doing my regular work and gaming. No stability problems.

I’m confident my x870 MSI Carbon can handle better than that, especially after recent BIOS updates. But this is my work related PC and I’m willing to push it only so far without a fear of blowing my RAM.

I noticed there are similar cases on this forum, succesfully done by much more knowledgable people than me, but their configuration vary and I don’t think it’s wise to follow someone else’s approach step by step, even if the hardware might seem similar.

So, if anyone would feel like giving me some advice on how to properly approach memory overclocking for high density DDR5 RAM, what to start with, what to look out for etc., I’d very much appreciate it! There isn’t much (up to date) information regarding this specific combination.

Why 192Gb RAM? I use a lot of RAM for procedural Blender modeling. And I decided to OC it, one for the sake of having such expensive RAM already and also because I want to focus more on assisted game dev programming later. I read high frequency RAM with low latency helps, if GPU can’t fully take over in some cases.

Bonus question: Can RAM stability be influnced after additional undervolting of my CPU?

Here are my specs:

MSI MPG x870e Carbon BIOS 1.A60
7950x (default PBO, VSoC currently locked at 1.3v)
Arctic Liquid Freezer 420 in push pull configuration
Corsair Vengence 4x48Gb DDR5-5200 CL38 PC5-41600 DDR5 SDRAM UDIMM, SK Hynix, M-Die, 2R CMK192GX5M4B5200C38

• currently overclocked to stable 5600Mt/s, but with bad timings and still high voltage of 1.4 (besides start to lowering down voltages, I don’t know what else to do here)
• I also discovered my RAM is version with slightly less stable SK Hynix memory than the original ones from Micron
• RAM cooling is currently being handled by 3x 6k RPM server fans attached together (40x40x28)
• I measure the temperature through HWinfo app, BUT those RAM sensors supposedly measure only PCB temperature and are also unreliable, as sometimes those temperature values in the app freeze (ironic) or entire RAM section disappears. Everything else in HWinfo works well and is fully responsive. But sensors on RAM sticks are acting up every now and then. Don’t know why.
  
  

[Janos]

That’s a 192GB DDR5-5200, CL38-38-38-84 Kit, there won’t be much room for improvement.
Maybe try tREFi 65535 and for tRFC look here

But depends on the ICs that are on your KIT

[GigaBusterEXE]

you have 192GB working at 5600MT/s that is like almost double most people get

consider your self lucky its not running 3600

[Patata]

What’s the read bandwidth with 4 dimms at 5600?

[igormp]

+1 to what others said about those speeds being already great.

I got plenty happy already once I managed 4x64GB@5200MHz 42-42-42, I wouldn’t even dare to try 5600MHz 38-38-38 since it sounds way too ambitious IMO (and will require lots of time finicking around)

[Exard3k]

@Patata

What’s the read bandwidth with 4 dimms at 5600?

Same as with 2 DIMMs
5600 MHz / PC5-44800, so 44.8GB/s per channel → 89.6GB/s theoretical max

[Janos]

@MoravianLion

VSoC currently locked at 1.3v)

Have you tried VSOC at 1.25V?
I have strange problems with USB and HDMI audio when VSOC is too high, and sometimes the first boot fails.

[MoravianLion]

[Janos]

try tREFi 65535 and for tRFC 600, then you get about 75ns

[MoravianLion]

My next step (for now) is to start reducing voltage, to see “how low” I can take it while still stable. I understand it that keeping voltage high, even if within reason, will eventually shorten life span. That goes for both RAM and IMC.

I also read from several sources that sometimes even high voltage can cause instability, to distort memory signal. Or something along those lines. But unless someone will explain to me how exactly I could try to go about making 6000Mt/s stable, I’ll just focus on preserving what works now.

[MoravianLion]

Is there a chance I can destroy my RAM when adjusting timings manually? If it would be just about resetting CMOS to get it back in case it won’t boot, that’s fine. I just don’t want to break it.

[Janos]

@MoravianLion

will eventually shorten life span

I think 1.4 VDD/VDDQ and MC with fan is still reasonable, I would prefer a lower VSOC

[Janos]

@MoravianLion

Is there a chance I can destroy my RAM when adjusting timings manually

no, only your stability.
tREFi and tRFC delivers more performance than all other subtimings combined

[Patata]

Same as with 2 DIMMs
5600 MHz / PC5-44800, so 44.8GB/s per channel → 89.6GB/s theoretical max

In my ddr4 system i lost bandwidth going to 4 dimms with the same frequency and timings(and more volts…).
It went from 53GB/s 61ns to 46GB/s 68ns. Both kits are the same.

Now with 192-256 working in mainstream systems i am thinking of upgrading

[MoravianLion]

Trying that right now, we’ll see! Should I properly test lowering vSoC or timings first? Since RAM being potentially finicky due to high voltage anyway, I should probably test both independently, just to be sure, right? I guess I’ll write again in a few days regarding this.

Also, can RAM stability be influnced after additional undervolting of my 7950x CPU? I was running CoreCycler for testing earlier, but then I found out my RAM was still unstable (I was running it at 6000Mt/s CL32, but didn’t test it properly enough back then. But 5600Mt/s CL38,38,38,96 (BIOS recommendation) went well during 12 hours of non stop testing. And during normal working hours too, so I think it’s safe to declare this stable now.

So, once I’m properly finished with RAM this time, I’d like to try undervolting CPU again.

[MoravianLion]

If you have the need, go for it! 3600Mt/s was working well on my earlier GB B650 DS3H for some $130 even (over 2 years ago). But that board didn’t allow any tinckering, afaik.

Recently, I read quite a lot about all major manufacturers and their x870 chipsets declaring you can get 192Gb or even 256Gb stable at 6000Mt/s (or even slightly more). Decided to buy one, also for potential CPU undervolting and additional NVMe slots.

So far so good, but I think the main “problem” here is quality of my RAM set.

Not that I’m complaining. It serves me well, gaming’s good too and gone are the days when I was running out of 32Gb on my DDR4 rig. Yey!

[Janos]

@MoravianLion

Should I properly test lowering vSoC or timings first?

Always make only one change at a time and then test stability, when the desired timings are reached, slowly reduce the voltage.
Although I don’t see any advantage with VSOC at 1.3V and 5600MT, your system will be stable with 1.25V, maybe you do that first

@MoravianLion

Also, can RAM stability be influnced after additional undervolting of my 7950x CPU?

Always separate things.
Is it better to find the CPU limit first and then tune the RAM? I would think so, but I haven’t really been interested in hard core OC since Sandybridge.
You’re more likely to find the right experts here:
Overclock.net

To gain experience with your system, you might want to proceed as follows:
Run the RAM with safe settings, find the CPU limit, and note the settings.
Second step: Run the CPU with safe settings to find the RAM limit.
Third step: Test whether both settings are stable together.

[MoravianLion]

I decided to find limits for RAM first, as that was the biggest compute bottleneck, imo. I managed to get my 4x48Gb sticks to run rock stable at values below.

I’m confident it could be done better, maybe even fully stable at 6000, but I refuse to put any more effort into this. Thanks again for these tips! And those tREFi and tRFC alone really knocked down the latency by around over 20ns (previously some ~96)!

I also dropped vSoC back to 1.25v as you said. Less was not stable. And RAM voltage down to 1.2v. Less was not stable.

Y cruncher, long desktop idlying, gaming, AI, editing, everything’s rock solid. I’ll take it!

CPU undervolting

About a week later, I started to undervolt my 7950x. Goal was to get lesser consumption, better temperatures and therefore also higher clocks. I ended up at offset of negative 27 for all cores. This is also rock solid. I’m already undervolting with overclocked RAM as described above.

Every evening before going to bed, I restart PC with offset of 1 extra point. Then run CoreCycler for next 10 hours, that goes over 3 iterations for each core. Then the next day, I run a few quick all core heavy tests (UE5 editor, with heavy scene - that proved to be also a good memory stress test in the past), some quick gaming and desktop idlying for about 20-30 minutes. Then for the rest of the day, I use my PC as usual (procedural, CPU heavy 3D modelling, working from home).

Up until offset of 28, no errors were found. Then core 13 showed following error.

I figured that I’ll simply set this core back to previously stable offset of 27 and will keep all other core at 28, which was marked as non problematic with CoreCycler over 10 hour long testing. I booted up Assetto Corsa later in the evening and PC crashed. Then I tried again with the same configuration. Crashed again. Not at all what I expected.

I thought the unstable core was set to previously rock solid value (rock solid from previous day) and rest should also work fine, as no errors were found AND PC worked fine up until evening hours. But it didn’t. The moment I set ALL cores back to 27, together with previously the unstable core 13, everything’s rock solid again.

Think I’m being funny, but it almost seems like my CPU hates having various undervolting values for each core. Is that possible? Am I missing something?

[Janos]

Great job, you can be happy with that, decent latency and relatively low voltage.
And -27 undervolting is really good, I only have -10.

@MoravianLion

but it almost seems like my CPU hates having various undervolting values for each core. Is that possible?

Yes, but I’ve never done it, that’s for people who only run Benchmarks all day.
If your result has motivated you, and you want to find a bit more performance, try to get your memory latency down to around 65ns.
This should also be doable with SK Hynix M-Die.
With 2DPC, you usually pay with bandwidth, latency is less affected, 65ns will be difficult but should be possible.
But your result with a Ryzen 7950x is already very good, It is not a mistake to keep some distance from the cliff

[MoravianLion]

Best CPU Cooler for Ryzen 9 9950x3d?

That’s intentional, the 7950X always goes up to 95C. With R9 AMD changed that again because many users don’t want to understand that the temperature limit isn’t the same for all CPUs and there have been endless discussions about it. Overclocking works essentially via cooling and indirectly through undervolting.

(hijacking an interested post from other thread, because this is more about me than the other guy and also because it was aimed at me, personally anyway)

I’m aware, my ultimate goal was/is to undervolt it to the point my AIO will be able to cool it enough so it can get to as high frequency as PBO allows.

AMD site says it can boost up to 5.7Ghz, so that might be the technical limit I might be able to achieve once thermals are not the ceiling anymore (with undervolting and direct die cooling). But correct me, if I said something wrong here.

One more thing. Let’s say thermals won’t be a problem anymore. I have only a single CPU power cable plugged. There are 2 sockets for this on my x870. Will single CPU power cable be enough for undervolted 7950x (-26 all cores) reaching close to 5.7Ghz? I’m unsure how the voltage/power balance shifts in such scenario. If the thermal limit is not a problem anymore, the consumption will rise, because CPU will overclock itself (desired behavior). So, the power draw will also increase. Currently during full load, it sits at 5.2-ish and some 215w. Used to be 4.9Ghz and some 230w when stock PBO was used (thermal limit was hitting harder, due to no undervolting).

Will I need second CPU power cable plugged in? I’d say yes, but would love hear someone else’s perspective/correction.

Can “unlimited” overclock make CPU unstable, if undervolted?

Or will I just have to “fuck around and find out”?