Ryzen 7000 / Zen 4 Owner's Club

[ENTERPRISE]

5 hours ago, J7SC_Orion said:

 

...my 670E Aorus Master has most of these

 

21 minutes ago, neurotix said:

Asus Strix X670E-E Gaming Wifi has them both.

 

Weird that you don't have gear down mode. That's pretty standard (and necessary) for RAM OC on Ryzen.

 

Also @J7SC_Orion you said you set VDDG voltages to 1.0v for 8000mhz. Mine were in Auto when I was trying. Think it might help stabilize me? Also what VDDP are you using?

 

Thanks chaps, I have given MSI feedback.

[neurotix]

Alright trying 8000mhz with the looser timing set that Orion gave me, 1.0v VDDG, and 1.1v VDDP. Doing memtestCL now. Wish me luck, but I do fully expect it to fail.

 

I think my problem with 8000mhz is that it might be cooling-related; I think you guys (kaliz and Orion) have fans pointed at your RAM, yes? In my case, with full RGB and matching RGB on the memory, I don't want a fan sitting on my GPU pointed at the RAM.

 

The interesting thing is that 8000mhz for me is benching stable: I ran Cinebench, y-cruncher and a few other demanding benches for my hwbot yesterday with the RAM at 8000, and it was fine. No Bsods, crashes in the programs or anything. I imagine it would probably be gaming stable too. However, since my rigs primary use case is a home server/streaming server, I cannot afford to have bad memory corrupt files over time, especially on obscure 80s anime I wouldn't be able to download again.

[neurotix]

Sorry for like triple posting.

 

8000mhz with 1.1v VDDP and 1.0 VDDG and the looser set of timings was a no-go. Within 15 minutes memtestCL produced like 15 error windows.

[ENTERPRISE]

2 hours ago, neurotix said:

Sorry for like triple posting.

 

8000mhz with 1.1v VDDP and 1.0 VDDG and the looser set of timings was a no-go. Within 15 minutes memtestCL produced like 15 error windows.

 

You are not alone. I thought my OC was stable, but when playing COD I get weird auto issues and lag spikes and system eventually resets...booo.

[neurotix]

19 minutes ago, ENTERPRISE said:

 

You are not alone. I thought my OC was stable, but when playing COD I get weird auto issues and lag spikes and system eventually resets...booo.

At 2133MHz fclk? RAM still at 6600MHz?

[kaliz]

make sure to set memory context restore to disabled , with this 1.0.0.7 b and c Agesa you will get training issues! AMD is on it to fix it with 1.0.0.8.

Edited October 5, 2023 by kaliz

[J7SC_Orion]

6 hours ago, ENTERPRISE said:

 

 

Thanks chaps, I have given MSI feedback.

 

VDDP...1.05 V (shows as 1.0478 V in Zen timings)...it was one of the voltages I lowered a bit per earlier post; before it was at 1.06 V.

[kaliz]

7 hours ago, neurotix said:

Alright trying 8000mhz with the looser timing set that Orion gave me, 1.0v VDDG, and 1.1v VDDP. Doing memtestCL now. Wish me luck, but I do fully expect it to fail.

 

I think my problem with 8000mhz is that it might be cooling-related; I think you guys (kaliz and Orion) have fans pointed at your RAM, yes? In my case, with full RGB and matching RGB on the memory, I don't want a fan sitting on my GPU pointed at the RAM.

 

The interesting thing is that 8000mhz for me is benching stable: I ran Cinebench, y-cruncher and a few other demanding benches for my hwbot yesterday with the RAM at 8000, and it was fine. No Bsods, crashes in the programs or anything. I imagine it would probably be gaming stable too. However, since my rigs primary use case is a home server/streaming server, I cannot afford to have bad memory corrupt files over time, especially on obscure 80s anime I wouldn't be able to download again.

 

most issues you have with AM5 platform is RAM training. One day you find a good stable setting, it plays all the games benchmarks and stesstests, and the next boot up you get lagspikes or errors. Really crazy.

 

I now have set Powerdown enabled, this seem to help, and memory context restore disabled, so that every time you warm reboot, the ram retrains. If that doesnt work, loading a profile, reboot, shut down system, turn off PSU completely, wait a bit and restart PC so it cold boots with your previously stable settings. Really annoying but if i wanna play BF2042 i do it like this. Then my 8000C32-44 profile with 1,70v is stable, or a 8000C34-45 profile with 1,62v. I always use a fan to DDR5 it gets really warm even on lower voltage. Specially the G.Skill modules, they dont have a thermal pad on PMIC and the heatsink arent great. If i dont use a ram fan i have seen 80C+ with TM5, and also with gameing when GPU heats up. Errors occurs at 50C+ sometimes even 45C+ and thats pretty easy to reach

 

 

Edited October 6, 2023 by kaliz

[J7SC_Orion]

40 minutes ago, kaliz said:

 

most issues you have with AM5 platform is RAM training. One day you find a good stable setting, it plays all the games benchmarks and stesstests, and the next boot up you get lagspikes or errors. Really crazy.

 

I now have set Powerdown enabled, this seem to help, and memory context restore disabled, so that every time you warm reboot, the ram retrains. If that doesnt work, loading a profile, reboot, shut down system, turn off PSU completely, wait a bit and restart PC so it cold boots with your previously stable settings. Really annoying but if i wanna play BF2042 i do it like this. Then my 8000C32-44 profile with 1,70v is stable, or a 8000C34-45 profile with 1,62v. I always use a fan to DDR5 it gets really warm even on lower voltage. Specially the G.Skill modules, they dont have a thermal pad on PMIC and the heatsink arent great. If i dont use a ram fan i have seen 80C+ with TM5, and also with gameing when GPU heats up. Errors occurs at 50C+ sometimes even 45C+ and thats pretty easy to reach

 

 

 

I also leave PowerDown enabled because the RAM is always busy doing something anyways with 2,300++ threads . I have GearDown disabled though. 

 

@ENTERPRISE @neurotix if you get to the point that RAM stress tests that loop over and over are fine for the first 10 times but fail at the eleventh or so, it is worthwhile checking the various DDR5 temps > that type of RAM really dislikes higher temps and starts throwing errors (I use 45 C as the cut-off). Another point is that my final 'fine-grain' DDR5 8000 tuning came down to CPU_VDDIO_Mem. At 1.33 V it works fine (same for DDR_VDDQ) and  it has not given me a single WHEA ever, or other errors since I finished tuning the RAM. I got there by observing the number of memory errors it was showing in a given test loop early on, then made some small voltage adjustments...if the number of errors declined, I knew I was on the right track...rinse and repeat until no errors even over full-length RAM stress tests, y-cruncher et al.

 

All in all, it was rather painless for me, not least as Buildzoid had tested very similar 2x24GB RAM on an Aorus mobo (650E Tachyon) just recently. My Aorus MAster 670E could take those values as a good starting position, and I actually ended up with (mostly) slightly lower voltages and tighter timings at 8000 than he showed. The real tricky bit was to run FCLK at 2167/75 and 8000 while staying within my voltage boundaries.

[ENTERPRISE]

15 hours ago, kaliz said:

make sure to set memory context restore to disabled , with this 1.0.0.7 b and c Agesa you will get training issues! AMD is on it to fix it with 1.0.0.8.

 

11 hours ago, kaliz said:

 

most issues you have with AM5 platform is RAM training. One day you find a good stable setting, it plays all the games benchmarks and stesstests, and the next boot up you get lagspikes or errors. Really crazy.

 

I now have set Powerdown enabled, this seem to help, and memory context restore disabled, so that every time you warm reboot, the ram retrains. If that doesnt work, loading a profile, reboot, shut down system, turn off PSU completely, wait a bit and restart PC so it cold boots with your previously stable settings. Really annoying but if i wanna play BF2042 i do it like this. Then my 8000C32-44 profile with 1,70v is stable, or a 8000C34-45 profile with 1,62v. I always use a fan to DDR5 it gets really warm even on lower voltage. Specially the G.Skill modules, they dont have a thermal pad on PMIC and the heatsink arent great. If i dont use a ram fan i have seen 80C+ with TM5, and also with gameing when GPU heats up. Errors occurs at 50C+ sometimes even 45C+ and thats pretty easy to reach

 

 

 

10 hours ago, J7SC_Orion said:

 

I also leave PowerDown enabled because the RAM is always busy doing something anyways with 2,300++ threads . I have GearDown disabled though. 

 

@ENTERPRISE @neurotix if you get to the point that RAM stress tests that loop over and over are fine for the first 10 times but fail at the eleventh or so, it is worthwhile checking the various DDR5 temps > that type of RAM really dislikes higher temps and starts throwing errors (I use 45 C as the cut-off). Another point is that my final 'fine-grain' DDR5 8000 tuning came down to CPU_VDDIO_Mem. At 1.33 V it works fine (same for DDR_VDDQ) and  it has not given me a single WHEA ever, or other errors since I finished tuning the RAM. I got there by observing the number of memory errors it was showing in a given test loop early on, then made some small voltage adjustments...if the number of errors declined, I knew I was on the right track...rinse and repeat until no errors even over full-length RAM stress tests, y-cruncher et al.

 

All in all, it was rather painless for me, not least as Buildzoid had tested very similar 2x24GB RAM on an Aorus mobo (650E Tachyon) just recently. My Aorus MAster 670E could take those values as a good starting position, and I actually ended up with (mostly) slightly lower voltages and tighter timings at 8000 than he showed. The real tricky bit was to run FCLK at 2167/75 and 8000 while staying within my voltage boundaries.

 

Awesome info guys. 

 

I did not think about Memory Context, I have disabled that now.

 

I also decided to take a step back and try for memory settings that @kaliz suggested, as below:

 

 

6400MT/s with these timings actually is not too far away from the raw speeds I was achieving with 6600MT/s. The only difference is I have a latency of 64ns with the above settings but I feel I can possibly tighten timings up further/try for the 2133 FCLK to improve on that.  I do have my VDDG CCD and VDDG IO at 1.1v which I know is on the higher side but I am hoping this will help in getting to 2133+ FCLK.

 

I will also see if at 6400 I can drop my VDDQ and VDDIO as anything below 1.5v running 6600, I would fail to post. I may try for 1.35 for both VDDQ and VDDIO.

 

I think for the moment tuning at 6400MT/s will be a more achievable goal over trying to push for 6600. I think even though 6600 was "Stable" it was not stable in all scenarios/on the edge. As I said, as soon as I fired up Call Of Duty, I was getting odd audio distortion and frame drops.

[neurotix]

5 hours ago, ENTERPRISE said:

 

 

6400MT/s with these timings actually is not too far away from the raw speeds I was achieving with 6600MT/s. The only difference is I have a latency of 64ns with the above settings but I feel I can possibly tighten timings up further/try for the 2133 FCLK to improve on that.  I do have my VDDG CCD and VDDG IO at 1.1v which I know is on the higher side but I am hoping this will help in getting to 2133+ FCLK.

 

 

2133 fclk will probably have no impact on your latency. That is all timing related. Try primaries of 36-36-36-36 if you can as opposed to 40. Then you may see lower latency.

[J7SC_Orion]

1 minute ago, neurotix said:

 

2133 fclk will probably have no impact on your latency. That is all timing related. Try primaries of 36-36-36-36 if you can as opposed to 40. Then you may see lower latency.

 

...I think higher fclk does make a difference on both bandwidth and latency, at least according to my tests and my setup. Lower primaries, especially CL, have a larger impact on latency, though.  In any event, the key is to recognize when fclk is on the edge because it too can throw various errors before crashing outright. I know this because my system will run with fclk at 2200+, but errors sneak in, so I set it to 2167 and with the Aorus bonus boost that comes to 2175. 

[neurotix]

30 minutes ago, J7SC_Orion said:

 

...I think higher fclk does make a difference on both bandwidth and latency, at least according to my tests and my setup. Lower primaries, especially CL, have a larger impact on latency, though.  In any event, the key is to recognize when fclk is on the edge because it too can throw various errors before crashing outright. I know this because my system will run with fclk at 2200+, but errors sneak in, so I set it to 2167 and with the Aorus bonus boost that comes to 2175. 

Fclk helps a little, yeah, but negligible conpared to CL timings as you already stated. If he could do 30-36-36-36 he'd probably see latency of 62~ns, possibly a little lower. Those were the timings I ran with my B-Die that got me really low latency. I think we still don't know what dies @ENTERPRISEhas. It would be good to see a Thaiphoon Burner report.

 

In my case, with 1.425 mem VDD, 1.35v mem VDDQ, 1.1v VDDP, and 1v VDDG (both) I was not stable at 8000mhz. Should I increase mem VDDQ and try 1.050v VDDG or possibly higher? If I do should VDDP be raised too?

[ENTERPRISE]

3 minutes ago, neurotix said:

Fclk helps a little, yeah, but negligible conpared to CL timings as you already stated. If he could do 30-36-36-36 he'd probably see latency of 62~ns, possibly a little lower. Those were the timings I ran with my B-Die that got me really low latency. I think we still don't know what dies @ENTERPRISEhas. It would be good to see a Thaiphoon Burner report.

 

In my case, with 1.425 mem VDD, 1.35v mem VDDQ, 1.1v VDDP, and 1v VDDG (both) I was not stable at 8000mhz. Should I increase mem VDDQ and try 1.050v VDDG or possibly higher? If I do should VDDP be raised too?

Hynix A-Die is what I have.

[neurotix]

Just now, ENTERPRISE said:

Hynix A-Die is what I have.

Gotcha. @J7SC_Orioncan probably help you with that more than I can, as the settings I'm using for my new RAM (M-Die) came from him

 

Personally I would suggest going with kaliz' timings and 2100 fclk, possibly adding some more RAM voltage to maybe lower timings, especially CAS, as much as you can if you are after lower latency. Buying a different kit is always an option too, mind linking me your kit on Amazon?

[J7SC_Orion]

2 hours ago, ENTERPRISE said:

Hynix A-Die is what I have.

 

...I have exactly one DDR5 kit, that's the M-die. But here are some goodies for Hynix A-die on Ryzen 7000 from Buildzoid. Also check if he has done a newer vid on that combo (this came to mind first). This is in addition to the excellent advice @kaliz has been giving...

 

 

[ENTERPRISE]

12 hours ago, J7SC_Orion said:

 

...I have exactly one DDR5 kit, that's the M-die. But here are some goodies for Hynix A-die on Ryzen 7000 from Buildzoid. Also check if he has done a newer vid on that combo (this came to mind first). This is in addition to the excellent advice @kaliz has been giving...

 

 

Thanks, I think I saw that one but will double check. 

 

So the current news is I have dropped my FCLK to 2000. Even after OCCT and Karhu @11000 % was "Stable" , it wasn't true. 

 

So I am check if the Infinity Fabric is responsible. Now running Y-Cruncher VST as well as the Third Bench to see if it picks anything up. Then will run some games. 

 

If it proves to be the Infinity Fabric I will leave it at 2000 @6400MT/s. 

 

Then I may see if I can go for 6600MT/s with the reduced FCLK while retaining the same timings on my 6400MT/s OC but that is more for E-Peen as 6400MT/s performs almost as good as 6600, but with 6400 I can likely improve my voltages.

 

What is considered a long enough run on Y-Cruncher VST ?

 

[The Pook]

1 hour ago, ENTERPRISE said:

Even after OCCT and Karhu @11000 % was "Stable" , it wasn't true. 

 

On 24/09/2023 at 18:48, The Pook said:

I used to use RAM Test exclusively but I switched to GSAT when I had a intermittent issues that I couldn't track down. RAM Test would run overnight without issues but GSAT would catch it within a couple hours. 

 

 

sudo apt-get update

sudo apt-get install stressapptest

stressapptest -s X -M Y --pause delay Z -W --max_errors 1

 

x = length in seconds 

y = RAM to test in MB 

z = length of time in seconds between pauses

 

you can play with pause delay if you're suspecting you're having an issue during power spikes between idle/load but I haven't found it useful. 

 

4 hours, 30GB, no pause delay: 

stressapptest -s 14400 -M 30000 --pause_delay 99999999 -W --max_errors 0

 

 

 

 

[neurotix]

I would highly, highly suggest getting GSAT and doing what @The Pookdescribed above. This is basically the best memory/memory bus test you can run. It is what Google uses to test and validate stability on their cloud servers. Google Stressful Application Test.

 

Being on Linux myself, I ran it there and my 7600MHz 36-45-45-45 oc was stable in it for 3 hours, which is the minimum I would recommend running to check memory stability.

 

GL E.

Edited October 7, 2023 by neurotix

[ENTERPRISE]

3 hours ago, The Pook said:

 

 

 

sudo apt-get update

sudo apt-get install stressapptest

stressapptest -s X -M Y --pause delay Z -W --max_errors 1

 

x = length in seconds 

y = RAM to test in MB 

z = length of time in seconds between pauses

 

you can play with pause delay if you're suspecting you're having an issue during power spikes between idle/load but I haven't found it useful. 

 

4 hours, 30GB, no pause delay: 

stressapptest -s 14400 -M 30000 --pause_delay 99999999 -W --max_errors 0

 

 

 

 

 

22 minutes ago, neurotix said:

I would highly, highly suggest getting GSAT and doing what @The Pookdescribed above. This is basically the best memory/memory bus test you can run. It is what Google uses to test and validate stability on their cloud server. Google Stressful Application Test.

 

Being on Linux myself, I ran it there and my 7600MHz 36-45-45-45 oc was stable in it for 3 hours, which is the minimum I would recommend running to check memory stability.

 

GL E.

Can I run this in a Linux VM or does it have to be native ? But its a fair point perhaps I need to take advantage of GSAT. 

 

I think I have a Live OS of linux on my diagnostic tools I could use. Any decent GSAT setup guides around?

 

Thanks very much guys.

[neurotix]

9 minutes ago, ENTERPRISE said:

 

Can I run this in a Linux VM or does it have to be native ? But its a fair point perhaps I need to take advantage of GSAT. 

 

I think I have a Live OS of linux on my diagnostic tools I could use. Any decent GSAT setup guides around?

 

Thanks very much guys.

For the first question, no you don't need a Linux VM anymore because of Windows Subsystem for Linux which runs in PowerShell. I've never used it myself so I don't know if it is active by default or if you have to enable it under Windows features, @The Pookwould know more.

 

Second question: no there is not a setup guide, what the Pook gave you IS the setup guide. The "sudo apt update" and "sudo apt install stressapptest" are Terminal commands that install and set it up for you. Then the other command and parameters runs it. I'd suggest trying these commands in Admin PowerShell and see if they work.

Edited October 7, 2023 by neurotix

[J7SC_Orion]

23 minutes ago, ENTERPRISE said:

 

Can I run this in a Linux VM or does it have to be native ? But its a fair point perhaps I need to take advantage of GSAT. 

 

I think I have a Live OS of linux on my diagnostic tools I could use. Any decent GSAT setup guides around?

 

Thanks very much guys.

 

My primary RAM test tools are TechPowerUp Memtest64 (lower left) OCCT, including the Linpack, and Y-Cruncher, with some SuperPi thrown in. Each can be varied re. loops / length all the way up to infinite (not that I would, but I could...).

My version of TM5 has a severe security issue, probably because it was downloaded from a questionable site (did check first, though) --- it attacked Win 11 tpm 

 

 

[The Pook]

36 minutes ago, ENTERPRISE said:

 

Can I run this in a Linux VM or does it have to be native ? But its a fair point perhaps I need to take advantage of GSAT. 

 

I think I have a Live OS of linux on my diagnostic tools I could use. Any decent GSAT setup guides around?

 

Thanks very much guys.

 

if you install Linux then the Linux users win

 

It might be marginally faster at finding errors on a bare metal install but I haven't tested. I just use WSL. You could do it in a VM or live USB instead if you want. 

 

There's no really much to setup for GSAT. Enable WSL, download Ubuntu in Windows Store, start it, install stressapptest (sudo apt-get install stressaptest) and then start it with the parameters you want (for 4 hours with 60GB total: stressapptest -s 14400 -M 60000 --pause_delay 99999999 -W --max_errors 0). Then just let it finish. 

 

you can increase the 60000 to all your RAM if you aren't using your machine while it tests but I leave a few GB free so I can at least ask @iamjanco for my 4090 while the test runs

Edited October 7, 2023 by The Pook

[neurotix]

GSAT is what professionals use.

[neurotix]

Alternatively, you could download the latest Linux Mint Cinnamon, (go down to mirrors and download from one of them) burn it to a flash stick using rufus and as GPT/UEFI. Boot off it and run the same commands in Terminal (should be next to Firefox in thd bottom left of the screen):

 

sudo apt update

sudo apt install stressapptest

stressapptest -s 14400 -M 32000 --pause_delay 99999999 -W --max_errors 1

 

It's really that easy, the biggest hassle will be making the flash stick and waiting for it to write.

Edited October 7, 2023 by neurotix