Tuning RAM question

[lance36]

CPU:5950x
GPU:6800
RAM:32GB of F4-3600C14-8GTESA
MOBO: Asus x570 tuf gaming pro wifi
Seasonic 850w power supply.
2TB western digital SN850 m.2

Question is about my ram timings. I get no errors with the timings in the screen shot but if I go over 3600 I get errors, and that's with 1:1 IF. If I wanted to go for higher bandwidth where would you start loosening timings. I get to 3733 with the current voltages but if I want to boot @ 3800 my IOD and CCD needs to be 1.05. Now I know I can push my SOC to 1.3 before I should start to worry but what about the other voltages like CCD,IOD, and CLDO?

Thanks for the help!

 

[tabascosauz]

CPU:5950x
GPU:6800
RAM:32GB of F4-3600C14-8GTESA
MOBO: Asus x570 tuf gaming pro wifi
Seasonic 850w power supply.
2TB western digital SN850 m.2

Question is about my ram timings. I get no errors with the timings in the screen shot but if I go over 3600 I get errors, and that's with 1:1 IF. If I wanted to go for higher bandwidth where would you start loosening timings. I get to 3733 with the current voltages but if I want to boot @ 3800 my IOD and CCD needs to be 1.05. Now I know I can push my SOC to 1.3 before I should start to worry but what about the other voltages like CCD,IOD, and CLDO?

Thanks for the help!

You may need around 1.10-1.15V SOC for 3800. No need ever to go above 1.2V for daily settings, 1.3V is insane on Vermeer outside of benching. More VSOC or more IOD doesn't always help anything.

IOD can sit about 50mV below SOC, 1.05 should be fine for 3800. CCD can be lower, I usually have mine 0.95-1.0.

VDDP should be fine 0.9-0.95.

If you get WHEAs even at 1.15V though, I'd just leave it alone where it stays WHEA-free. If you can run 4000 with occasional WHEAs then you can consider suppressing error reporting and running it, if you're below 4000 and WHEAs pop even with the right CLDOs, then just respect the CPU's wishes.

 

[ir_cow]

1.3 is way to high

 

[lance36]

With the current timings @3800 I can only boot with the SOC @1.125, and the CCD and IOD at 1.05 or 1.1 I don't quite remember which one of the two but it definitely won't boot without it. Oh and the ram needs to be at 1.5v for 3800. Haven't even tried 4000. As long as those are fine I'll most likely try it just to see how I like it compared to 3600. Now with 3800 I just wanted to find where it would boot. @ 3733 unit I get before 3600 I get errors in memory tests but no WHEA errors. I've been running this for a while.

 

[Mussels]

1.3v SoC is too high

I don't even need a full 1.15v to run at 3866, i can't run higher without WHEAs

Ryzen likes even timings, so i wouldnt be running all those odd numbers. GDM is going to make it behave differnt to the values you're seeing so you arent getting the full gains from that

 

[tabascosauz]

With the current timings @3800 I can only boot with the SOC @1.125, and the CCD and IOD at 1.05 or 1.1 I don't quite remember which one of the two but it definitely won't boot without it. Oh and the ram needs to be at 1.5v for 3800. Haven't even tried 4000. As long as those are fine I'll most likely try it just to see how I like it compared to 3600. Now with 3800 I just wanted to find where it would boot. @ 3733 unit I get before 3600 I get errors in memory tests but no WHEA errors. I've been running this for a while.

Bumping PLL may help you at the edge of Fabric stability. Not sure what it's called, different boards call it different things, but default is always 1.8V I think.

Didn't help me on 5800X3D (07 code problem), and I never needed it on 5900X, but it's worth a shot up to 2.0V max iirc

I'd try to get to 3733 at least, but not hitting 3800 is not that big a deal - still have to make sure you're actually getting better performance at higher Fabric anyway. You still have some room to tighten. You can drop tRFC a lot, tRRDL should probably be 4 or 6, tRTP can go to 6, can try getting tRC below 40, etc. Just remember that a lot of these might need you to bump VDIMM.

No WHEAs at 3733 is good. memtesting errors is just because your timings profile still needs work.

1.5V is fine and honestly good for 3800 14-13-13. I run 1.55V for that 3800CL14 profile.

1.3v SoC is too high

I don't even need a full 1.15v to run at 3866, i can't run higher without WHEAs

Ryzen likes even timings, so i wouldnt be running all those odd numbers. GDM is going to make it behave differnt to the values you're seeing so you arent getting the full gains from that

VSOC and IOD requirements different, as is Fabric scaling, lot harder for 2CCD

Unless B2 stepping has significantly changed things (which it really doesn't look like), lots of 5800X/5600X can hit 4000+ in their sleep, while it's still lucky for 2CCD to just boot 4000.

 

[lance36]

Bumping PLL may help you at the edge of Fabric stability. Not sure what it's called, different boards call it different things, but default is always 1.8V I think.

Didn't help me on 5800X3D (07 code problem), and I never needed it on 5900X, but it's worth a shot up to 2.0V max iirc

I'd try to get to 3733 at least, but not hitting 3800 is not that big a deal - still have to make sure you're actually getting better performance at higher Fabric anyway. You still have some room to tighten. You can drop tRFC a lot, tRRDL should probably be 4 or 6, tRTP can go to 6, can try getting tRC below 40, etc. Just remember that a lot of these might need you to bump VDIMM.

No WHEAs at 3733 is good. memtesting errors is just because your timings profile still needs work.

1.5V is fine and honestly good for 3800 14-13-13. I run 1.55V for that 3800CL14 profile.



VSOC and IOD requirements different, as is Fabric scaling, lot harder for 2CCD

Unless B2 stepping has significantly changed things (which it really doesn't look like), lots of 5800X/5600X can hit 4000+ in their sleep, while it's still lucky for 2CCD to just boot 4000.

So tightening my timing might clear the errors at 3733? Cool! I won't be heart broken if I can't get 3800 but 3733 would be nice. Thanks!

 

[Mussels]

Bumping PLL may help you at the edge of Fabric stability. Not sure what it's called, different boards call it different things, but default is always 1.8V I think.

Didn't help me on 5800X3D (07 code problem), and I never needed it on 5900X, but it's worth a shot up to 2.0V max iirc

I'd try to get to 3733 at least, but not hitting 3800 is not that big a deal - still have to make sure you're actually getting better performance at higher Fabric anyway. You still have some room to tighten. You can drop tRFC a lot, tRRDL should probably be 4 or 6, tRTP can go to 6, can try getting tRC below 40, etc. Just remember that a lot of these might need you to bump VDIMM.

No WHEAs at 3733 is good. memtesting errors is just because your timings profile still needs work.

1.5V is fine and honestly good for 3800 14-13-13. I run 1.55V for that 3800CL14 profile.



VSOC and IOD requirements different, as is Fabric scaling, lot harder for 2CCD

Unless B2 stepping has significantly changed things (which it really doesn't look like), lots of 5800X/5600X can hit 4000+ in their sleep, while it's still lucky for 2CCD to just boot 4000.

1:1 at 4000 is easy to reach on single CCD chips, but not stable and without WHEAs

 

[tabascosauz]

So tightening my timing might clear the errors at 3733? Cool! I won't be heart broken if I can't get 3800 but 3733 would be nice. Thanks!

Probably won't clear your memtesting errors (probably the opposite if you don't adjust VDIMM), but it's much better to have memtest errors (memory) than WHEA errors (Fabric) that you have little control over.

B-die is fine up to 1.6V daily when cooled properly. I can't hit 3800 on my 5800X3D so I'm running 3733 at 1.53V.

Bdie can do whatever you want, you just gotta give it the beans (volts)

 

[lance36]

Might have gone too far on my TRFCs? She gets hot now, if I understand right 50c is too hot. This was about 20mins in to a test. no errors no whea.

 

[3x0]

Don't worry too much about temps if there are no errors in TestMem5, HCI Memtest and other stability tests

 

[lance36]

Don't worry too much about temps if there are no errors in TestMem5, HCI Memtest and other stability tests

If I don't see those temps while working/gaming it doesn't matter for sure.

 

[3x0]

I generally see ~5°C less in games compared to some stress tests, but it will depend on voltages, timings and CAD_BUS/RTT settings

 

[tabascosauz]

@lance36 what's VDIMM now? If TM5 doesn't start destabilizing due to heat you're still fine, but best to try getting temps down through more airflow, not reducing VDIMM. It depends on setup obviously but I have had gaming temps come rather close to TM5 temps before.

Kinda strange that you were previously at sub-40 and now at 50 from just changing tRFC and not much else, what changed in terms of airflow?

 

[lance36]

@lance36 what's VDIMM now? If TM5 doesn't start destabilizing due to heat you're still fine, but best to try getting temps down through more airflow, not reducing VDIMM. It depends on setup obviously but I have had gaming temps come rather close to TM5 temps before.

Kinda strange that you were previously at sub-40 and now at 50 from just changing tRFC and not much else, what changed in terms of airflow?

When you said VDIMM, is that the same as DRAM voltage? 1.48v I have to loosen the trfc's to get it to boot at 3733. But the bench marks show a trade off of less than 1k of bandwidth for 1 point of Latency. Seems like a bad trade. No changes in air flow I have a lian li 360 aio 4 case fans and a fan curve that sounds like a helicopter taken off. What do you mean by TM5? Oh she only gets to about 45c gaming in Rust. What should my other Trfc's look like?

 

[tabascosauz]

When you sad VDIMM, is that the same as DRAM voltage? 1.48v I have to loosen the trfc's to get it to boot at 3733. But the bench marks show a trade off of less than 1k of bandwidth for 1 point of Latency. Seems like a bad trade.

Yes VDIMM = dram voltage

AIDA is not a very scientific measurement of DRAM performance when you're adjusting individual things. There are a lot of timings that it simply ignores. For alternatives can try membench (the second tab in DRAM Calculator), Linpack, Shadow of the Tomb Raider on lower res/settings, probably some others I'm missing

What's your RAM cooling look like now?

TM5: forgive me, I was thinking of another thread. You need to be memtesting with something like this Memory Testing with TestMem5 TM5 with custom configs | Overclock.net

Cooling: most people serious about B-die will find a way to get direct airflow over the DIMMs, whether by a downdraft cooler, or a fan directly on the DIMMs (eg. stood up on the back of the video card). There is no alternative, unless you like running 3000rpm case fans all day.

 

[lance36]

Yes VDIMM = dram voltage

You're still pretty lax on voltage. It's not a problem but if you want to push as tight a profile as you posted, you will need to bump VDIMM for stuff like tRFC. Like a lot of other things you can't just tighten and expect volts to stay the same. tRFC is pretty important but AIDA also disproportionately favours tRFC for its latency measurement. I wouldn't leave it at 160ns for an expensive B-die kit, but it's also not the end of the world if you decide to stay at 140-145. In any case, big jumps in tRFC make a big difference (e.g. -100, -200) but smaller improvements not so much

AIDA is not a very scientific measurement of DRAM performance when you're adjusting individual things. There are a lot of timings that it simply ignores. For alternatives can try membench (the second tab in DRAM Calculator), Linpack, Shadow of the Tomb Raider on lower res/settings, probably some others I'm missing

What's your RAM cooling look like now?

I can bump it up to 1.5 but I'm not running anything in the background. Hmmmm Tfrc is at 238, I'll bump the VDIMM to 1.5 and see if I can't go down.

 

[tabascosauz]

I can bump it up to 1.5 but I'm not running anything in the background. Hmmmm Tfrc is at 238, I'll bump the VDIMM to 1.5 and see if I can't go down.

If you have a spare fan just stand it up on the back of the GPU and point it at the RAM. No sense compensating for poor cooling with less VDIMM.

I have no doubts your case airflow is great. But when pushing Bdie hard (especially 4DIMMs, always a challenge), you're going to need to do better than just cranking up your case fans. Point a fan at it and you'd be surprised how irrelevant case airflow is (then you can wind em back down to cut back on noise)

Already hitting 50C and not even in a stress test like HCI/TM5 or Linpack is not a great start, you're not even at 1.5V yet. Run TM5 with anta777 or 1usmusv3 configs and see what temps you reach, whether you're stable, and whether you're destabilizing as a result of temps.

 

[lance36]

If you have a spare fan just stand it up on the back of the GPU and point it at the RAM.

Already hitting 50C and not even in a stress test like HCI/TM5 or Linpack is not a great start, you're not even at 1.5V yet. Run TM5 with anta777 or 1usmusv3 configs and see what temps you reach, whether you're stable, and whether you're destabilizing as a result of temps.

The 50c in my screen shot was from about 20 minutes into the memtest that come with the Dram calculator. I can try others tho

 

[tabascosauz]

The 50c in my screen shot was from about 20 minutes into the memtest that come with the Dram calculator. I can try others tho

iirc that's just a longer version of membench no? Decent benchmark but not a good stress test. Just use TM5 to start.

On 2 DIMMs (albeit on ITX boards) I don't think I even reach equilibrium within 20 minutes. Do you have an extra fan you can point at it?

 

[lance36]

iirc that's just a longer version of membench no? Decent benchmark but not a good stress test. Just use TM5 to start.

On 2 DIMMs (albeit on ITX boards) I don't think I even reach equilibrium within 20 minutes. Do you have an extra fan you can point at it?

I'll have to get one. I'll seem how fast it shots up tomorrow with TM5.

 

[tabascosauz]

I'll have to get one. I'll seem how fast it shots up tomorrow with TM5.

If there is absolutely no other option, loosening tRFC should theoretically decrease Bdie's temp sensitivity (ie. raising the threshold for instability from 45C to 50C, for example). But you might have to loosen quite a bit to notice a difference, and slower than about 170-180ns would just defeat the point of paying for Bdie. Best to try addressing the cooling aspect directly first.

 

[Mussels]

TRFC behaves so differently between RAM kits of different density
Mines 704 at 3800 and 712 for 3866, on 32GB sticks


the performance loss from loosening it isn't massive, so definitely loosen it if it helps stability

 

[3x0]

From my experience, tRC is also sensitive to high temps. Had to increase from tRC 36 to 42 to be stable at ~60°C (tRC 36 was stable at ~53°C)

 

[lance36]

If there is absolutely no other option, loosening tRFC should theoretically decrease Bdie's temp sensitivity (ie. raising the threshold for instability from 45C to 50C, for example). But you might have to loosen quite a bit to notice a difference, and slower than about 170-180ns would just defeat the point of paying for Bdie. Best to try addressing the cooling aspect directly first.

It only ran for 6mins, I'm loading the 1usmus config now, I only used the default. What config should I use?