Issues getting 3800mhz B-Die temperature stable on Ryzen 3000

[hotrob]

Hi All,


I've been having some fun tuning my memory, but I seem to have hit a wall. My second kit of ram came in, and I've got it tuned to 3800mhz 16-16-16-32. I'm happy with the performance here, and I can run memtest 1200% stable no problem.... Until I stress my GPU at the same time. After about 4-5 minutes of stressing my GPU and getting heat into the system, it starts spitting out errors. I spent all day yesterday tweaking the secondaries and tertiaries, only to say duck it and loosen the primaries to the moon (18-22-22- 22-40) and it still did the same thing.



It's not the fclk, because if I lower the mclk below 1600 and keep fclk at 1900, it's fine. I'm running the dram voltage at 1.375; any lower than 1.37 and it's not stable at all, any higher than 1.39 and it is also not stable (due to temperature I assume). I tried raising the procODT and that helped a little, but not enough to make a difference with the temperature stability.

Anyone have any tips? Is there some secondary or tertiary that's particularly temperature sensitive on bdie? Or do I hang my head in shame and drop down to 3600 and tighten the timings some more until I save up to get a watercooling loop just for the ram?


Many thanks!

https://ibb.co/GMsvTZx
(Screenshot of full timings and "easy" memtest validation)

(Specs in signature)
**Edit - Attached screenshot with timings from DRAM Calc

 

[AlienALX]

It's the FCLK.

You need to understand that just because it may be stable at 1900 it doesn't mean the total overall speed is going to be happy.

I run 4133 RAM. At 3666. That was as fast as I could get it to go and be stable. However, whilst that doesn't sound so impressive my timings are 15, and I am yet to push onto 14.

The art with 3000 Ryzen is not the overall brute speed but how tight you can get the timings. In benchmarks it makes a difference, right down to 15. Then I noticed not much was happening (from the 16 I used before) so I left it be. What did make me happy was that when I added a second kit the same I was able to maintain both the clocks and timings.

Water cooling the RAM is a total waste of time dude. Always has been, always will be.

Start at 3666 with your timings at 19, then run AIDA 64 memory bench. Then lower the timings and test again.

Also remember that the higher you run the FCLK the more you run the risk of damaging the CPU.

 

[hotrob]

Oh that's a very interesting tip - thank you very much! I didn't think that memory speed/rank could destabilize the fclk. This gives me a new angle to think from tinker with - I'll let you know how it goes


Also, I didn't know that high fclk could degrade the cpu on its own, I thought it was just iod voltage that would do that.

 

[AlienALX]

From what I have seen on media 3733 was the fastest I have ever seen any one get the RAM speed on R 3000. And usually it just isn't worth it, given that it's the timings that really matter.

Like, some guy on Reddit spent weeks and weeks testing and from what he could see there are two perfect situations on R3000. 3200 14, 3600 16 and or lower.

It's not like Intel where the RAM speed can go higher. It really doesn't matter that much tbh. It's still "only" dual channel, so you will find it getting spanked by quad channel any way.

 

[Bartacus]

I'm daily driving 3800 at 1:1 with my 2x16GB G.Skill kit (3600CL16), but can't do that with 4 sticks, only 2. I also have a RoyalZ 4x16GB kit at 3600CL16, but that kit will not do 3800, it tops out at 3733. I'm running my 5950x at 3800 with the 2x16 kit without messing with the CL16-16-16-36 timings and it seems happy as a clam.

But with all that said, I'm not seeing ANY appreciable difference at 3800 versus 3600, so I'd do what Alien said and focus on tightening the timings at 3600. A lot of people on OCN are chasing those higher IFs, but without much performance benefit, other than maybe very specific memory benchmark programs, which do not reflect everyday performance.

 

[hotrob]

Did some more testing, and 3666 and 3733 were both less stable than 3800, so for my setup, it looks like it'll have to be 3600 or 3800.



For spits and giggles, I tested 1600/1800 fclk and 1900mclk, and still the same issue; 1900 fclk and 1600/1800mclk did not give me any problems, so I don't think it's an issue with my infinity fabric.


If I have to drop down to 3600, then that's the way it's gotta be... Sometimes you just gotta know when you're defeated. (That, or learn to stop worrying and love the irql_not_less_or_equal BSOD).



I know the performance gains aren't really worth mentioning (particularly with my dog of a r5 3600 that can barely boost past 4 ghz sc), but what really kills me, is I can run memtest all day at 3800 cl16, without any problem. It only gives errors when I add the heat from a GPU stress test. It's so close to stable I can taste it....