Ryzen AM5 ; 7000

[RasmusP]

Thanks a lot!!
I tested my ram settings with 10 minutes of Y-Cruncher, then Linpack and then 2 hours of Testmem 5 with the Ryzen 3D preset.
At first it seemed stable but all 3 got errors.
Then only Testmem5, but within two minutes!

Then got it all stable for two hours of Testmem5, but the latest ACC Patch crashed seemingly at random for a LOT of people.

I'm gonna buy Karhu and let it run overnight!
Especially since it doesn't fry my CPU...
I mean it's a joke anyway with my 7800X3D that probably doesn't know how to run into thermal throttling, but still not great with 0.38€/kWh in Germany to run it at 100% for 24 hours...


About performance: I really need to make a real comparison and not just some quick tests here and there, but it was quite a lot, even with the 3D!
Timings are a lot more effective for ACC than Transferrate though.
So 5400 with tuned subtimings is better than 6200 with sluggish timings.

 

[Neilski]

Heck, some PCs can't even run the Linpack X and Y-Cruncher tests at all because of the CPU heat regardless how stable their RAM is.

Huh? Surely the CPU should just throttle if the temperature spikes up...
What's the symptom you mean, in this instance - a crash?

 

[Durge Driven]

Huh? Surely the CPU should just throttle if the temperature spikes up...
What's the symptom you mean, in this instance - a crash?

I don't know what people are doing.
I manually set thermals to 80C for mine ?
Actually according to what I see 700 in single thread is very good and multi barely drops
like a few % and temps are great stressing CPU less 24/7.
With 360 cooler it sits at 40C ( has settled @ 39C now I notice )

What you loose in sims is inconsequential if you have already set lower FPS
It never ramps up in rF2 / LMU ( sits around 65C)
I limit rF2 to 144 and LMU to 120, CPU does not work as hard.
I never get thermals dropping FPS or anything else it is absolutely solid barely drops from max.

I understand low resolution tests but for gaming I have never seen the point.
If you run those same tests at 3440 the gap will shrink hugely.

 

[Spinelli]

Huh? Surely the CPU should just throttle if the temperature spikes up...
What's the symptom you mean, in this instance - a crash?

Depends if CPU temp limits are enabled, at which temp/s they're enabled, as well as how much the CPU is overclocked (if it's overclocked at all).

Running such a stressful program like Linpack Xtreme, Y-Cruncher, etc. hammers the CPU with very high power. Heat contributes to instability too. A CPU may crash/error a test at, say, 88 degrees but may pass it all day with the exact same overclock, voltages, etc. at, say, 73 degrees. The lower the CPU temp, the less voltage is needed to stabilize a given CPU frequency.

I like to reduce my clockspeeds when running programs like LinX and Y-C when I'm doing pure RAM stability testing. I don't want a possible CPU instability/crash to interfere with RAM testing. If there's an error or crash, I want to be 100% sure it was the RAM and not possibly the CPU.

I don't know what people are doing.
I manually set thermals to 80C for mine ?
Actually according to what I see 700 in single thread is very good and multi barely drops
like a few % and temps are great stressing CPU less 24/7.
With 360 cooler it sits at 40C ( has settled @ 39C now I notice )

View attachment 782956

What you loose in sims is inconsequential if you have already set lower FPS
It never ramps up in rF2 / LMU ( sits around 65C)
I limit rF2 to 144 and LMU to 120, CPU does not work as hard.
I never get thermals dropping FPS or anything else it is absolutely solid barely drops from max.

I understand low resolution tests but for gaming I have never seen the point.
If you run those same tests at 3440 the gap will shrink hugely.

7800X3D's default throttling temp is 89 degrees and perfectly safe. By lowering the throttle temp from 89 to 80 degrees, you're not really accomplishing anything besides decreasing your system's potential performance.

 

[Neilski]

A CPU may crash/error a test at, say, 88 degrees but may pass it all day with the exact same overclock, voltages, etc. at, say, 73 degrees.

Fair enough, though I've never experienced that.
If I were tuning up a machine and it failed at any temperature, I'd treat that like any other fail and detune it until it passed. (Unless of course I could guarantee that it could never again reach the temperature at which it failed.)

 

[Spinelli]

Fair enough, though I've never experienced that.
If I were tuning up a machine and it failed at any temperature, I'd treat that like any other fail and detune it until it passed. (Unless of course I could guarantee that it could never again reach the temperature at which it failed.)

I understand what you're saying and agree for the most part but for some people, programs like LinX, Y-C, Prime95 (AVX2/AVX512 + small FFTs only) etc. are the exception because they place such an ungodly & unrealistic load, power, and heat on the CPU. Even stock CPUs will often fail these tests if they don't have their auto-throttling settings enabled.

Cooling can play a big part in it all too though.

Unless of course I could guarantee that it could never again reach the temperature at which it failed.

Yes, very good point, well temp & type-of-load combo.

 

[RasmusP]

I understand low resolution tests but for gaming I have never seen the point.
If you run those same tests at 3440 the gap will shrink hugely.

Yeah the tests were pretty useless, since my 3080 can only output 110-140 fps with my optimised settings and I'm using a limiter at 97 fps with my 100 Hz gsync monitor
But I've got some microstutters here and there and even some bigger frame time spikes, so I was just curious about the raw CPU performance in ACC between Win 10, Win 11 22H2 and 23H2 with the new CPU scheduler update.
It clearly shows, why I'd like to have a 240 Hz monitor so I wouldn't need any syncing to have a smooth and tearing free image..
My 0.1% low average are all above 100 fps, so in theory, I should never have any stutters at all.
But the weird thing is: If I set the limiter to 80 fps, I'll have drops into the 70's. If I set the limiter to 120 fps, I'll have drops into the 110's.
Basically I'll always have drops, no matter where the limiter is.

Win 10 in fullscreen mode is by far the smoothest though. FPS are basically rock solid.
With my main Windows, 22H2 in borderless windowed (multitasking during practice sessions with my team), I always have some stutters without any clue. It's like ACC simply decides to drop a frame, randomly.

 

[RasmusP]

@Spinelli just ran Karhu for 4 hours with 5400 and my tuned timings. 0 errors.
Then put it to 6000 and eh.. well...:
View attachment 783287
Edit: attached screenshot got broken.. It was 198 errors in 3 minutes

TestMem5 also got a few errors within 10 minutes, but Karhu got them WAY quicker!
Nice recommendation, gonna compare fps and decide in which direction to tune

 

[RasmusP]

Update:
To make the test a bit less "scientific but quite useless" and more about my personal scenario, I've kept my graphics settings and didn't reduce the resolution. GPU riding it's limit at 95-99% load.
Loaded up a replay with me alone on track and recorded 300 seconds with CapFrameX to get a nice and steady test.

Results:
6000 MHz:

5400 MHz:

Looking at the graph, you can clearly see the real issue for me (and probably borderless mode in ACC):

So I'm gonna do the 1024x768 CPU benchmark on my Win 10 in Fullscreen mode. And also do 6000 EXPO vs. 5400 custom-tuned Timings

 

[RasmusP]

5400 EXPO Timings:

5400 tuned:

6000 tuned: (slightly raised voltages, no Karhu/TM5 errors within 10 minutes)

6000 EXPO:

I'm a bit surprised tbh... My last test was with my Win 11 22H2 which has significantly lower fps overall nowadays, but had identical fps to my old Win 10 from 2020.
I guess not reinstalling Win 10 in November 2022, when upgrading from i5 10600k to Ryzen 7600X was a mistake

Anyway, the results don't really make a lot of sense, lol.
Guess I'll use EXPO timings, although all of them are slower?


Here's Zentimings with EXPO vs my tuned timings:
EXPO:

Tuned:

 

[RasmusP]

Should be stable enough now

 

[Spinelli]

[delete]

 

[Spinelli]

Update:
To make the test a bit less "scientific but quite useless" and more about my personal scenario, I've kept my graphics settings and didn't reduce the resolution. GPU riding it's limit at 95-99% load.
Loaded up a replay with me alone on track and recorded 300 seconds with CapFrameX to get a nice and steady test.

Results:
6000 MHz:
View attachment 783299

5400 MHz:
View attachment 783300



Looking at the graph, you can clearly see the real issue for me (and probably borderless mode in ACC):
View attachment 783301

So I'm gonna do the 1024x768 CPU benchmark on my Win 10 in Fullscreen mode. And also do 6000 EXPO vs. 5400 custom-tuned Timings

If a 6000 MHz tune/setup isn't stable (ie. throwing errors) it can result in no better performance, or even worse performance, than a 5400 MHz tune so make sure it's stable before seriously comparing benchmark numbers.

Karhu
I suggest enabling cache for Karhu (in Karhu's options), that'll speed the test up even more (faster error detection) and improves IMC error detection better.

Aim for at least 6400% coverage. Some people suggest 10,000% instead. Most people agree that 20,000 is overkill. PC component manufacturer, Sabrent, recommends 12,800%. They also recommend having 1 instance of Karhu open per CPU thread but that seems wrong as 1 instance of Karhu should take care of all threads. The free demo of HCI MemTest (used to be incredibly popular during DDR3 and early DDR4 days) should have a separate instance opened for the amount of CPU threads you have but I've never, ever heard of Karhu or TM5 requiring that (or even the paid version of HCI MemTest for that matter) therefore that may be incorrect info from Sabrent.

TM5
Download it here, this is the most up to date version: https://github.com/CoolCmd/TestMem5
Translate the page using your browser or a website (eg. google translate).

This also includes 8 different tests:

You no longer have to manually move files around or have a separate install/folder of TM5 for each test like with the old version. You can select the desired test in the settings -> testing section of TM5.

Program should still be run as admin (I can't remember why but it was absolutely crucial with the old/original version).

TM5 Tests
The most common tests are generally 1usmus v3 and Absolut. Absolut was apparently made to replace Extreme but I've done tests where Extreme found errors quicker than Absolut so I also use Extreme.

I have no idea when/if the two DDR5 tests (Intel and Ryzen3D) should be run as opposed to 1usmus v3 and Absolut/Extreme. I have a theory though: aside from Karhu, a lot of RAM testers were/are not detecting DDR5 errors properly (or missing them alltogether). So maybe that's what the 2 DDR5 tests are. I could be 1000% wrong about that though so don't take my word for it.

1usmus v3 has a cheat sheet online. The cheat sheet helps in determining why the error is occurring. For example, TM5 w/ 1usmus v3 gives you an error code of "6". Look up "6" on the cheat sheet to find the likely reason for the error (too little VCCSA voltage, too low of a specific subtiming, etc.). I'm not sure if the cheat sheet is only for DDR4 though and/or only Intel.

The new version (linked above) of TM5 does the test in terms of time rather than cycles like the old version. The general consensus, in terms of cycles, for stability seems to vary. Some say 8 cycles, some say 21 to be really sure. I would forget about the cycles and just let 1usmus v3 and Absolut/Extreme run for at least 6 hours each.

RAM Temps
Keep HWinfo open to check for RAM temps (if your RAM has temp sensors). Your RAM may be fully stable and only erroring purely due to heat-induced instability. Different RAM sticks have different tolerances to heat (and some subtimings less tolerant to heat too). So research what sticks you have (eg. Hynix A-Die, etc.).

Other Programs
Close down as many programs as you can. Web browsers, sim racing background programs (wheel/pedals software, etc.), etc. You want as much unused/free RAM as possible for the RAM tests to use. The exception is a monitoring program like HWinfo (to track things like temperatures).

 

[Spinelli]

Should be stable enough now
View attachment 783436

I'm 99% sure you're stable.

Here you go: https://www.overclock.net/threads/amd-ddr5-oc-and-24-7-daily-memory-stability-thread.1800926/
Tons of help & info there.

It seems, with DDR5, most people are suggesting around 40,000%. I guess the 6.4-10 K numbers are for DDR4 only.

Karhu seems to have removed the % from their own site's FAQ section and now list time only:

Either way, it's almost certain your RAM and CPU's IMC are fully stable at those speeds/settings

 

[Neilski]

6000 tuned: (slightly raised voltages, no Karhu/TM5 errors within 10 minutes)

This one wasn't the highest average, but the pic below it seemed to show (to the naked eye) the least spiky behaviour... But then maybe that comes and goes?

 

[RasmusP]

Holy moly, that was a massive and nicely sorted write-up! Thanks a lot!

If a 6000 MHz tune/setup isn't stable (ie. throwing errors) it can result in no better performance, or even worse performance, than a 5400 MHz tune so make sure it's stable before seriously comparing benchmark numbers.

Yeah I figured the same, but it's weird that the 5400 with EXPO timings is better than 5400 with my tuned timings. Doesn't make any sense at all. I guess it's just within the margin of error between runs..

Karhu
I suggest enabling cache for Karhu (in Karhu's options), that'll speed the test up even more (faster error detection) and improves IMC error detection better.

Yep. I've read the manual/FAQ and did that

Karhu should take care of all threads.

I think so too. It even has a threads setting, so probably old info from Sabrent in that regard..

TM5

Did everything you wrote there, thanks anyway!

I have no idea when/if the two DDR5 tests (Intel and Ryzen3D) should be run as opposed to 1usmus v3 and Absolut/Extreme. I have a theory though: aside from Karhu, a lot of RAM testers were/are not detecting DDR5 errors properly (or missing them alltogether). So maybe that's what the 2 DDR5 tests are. I could be 1000% wrong about that though so don't take my word for it.

The DDR5 Ryzen preset showed errors way faster than the others, so I used that one. Ran it last night for 8 hours, no errors
But I'm currently running the 1usmus preset, just to be sure..

1usmus v3 has a cheat sheet online. The cheat sheet helps in determining why the error is occurring. For example, TM5 w/ 1usmus v3 gives you an error code of "6". Look up "6" on the cheat sheet to find the likely reason for the error (too little VCCSA voltage, too low of a specific subtiming, etc.). I'm not sure if the cheat sheet is only for DDR4 though and/or only Intel.

That's nice! If I ever feel the need to tweak further, I'll use that!

RAM Temps
Keep HWinfo open to check for RAM temps (if your RAM has temp sensors). Your RAM may be fully stable and only erroring purely due to heat-induced instability. Different RAM sticks have different tolerances to heat (and some subtimings less tolerant to heat too). So research what sticks you have (eg. Hynix A-Die, etc.).

Apparently no sensor on the sticks, but the mobo has a dd5 socket sensor, which showed 50°C maximum when running TM5 over night. And 48°C max when racing.
I have Kingston 5600 CL36 Hynix M-Die sticks. 2x 16GB.
That's why I used the "easy DDR5 timings" from Buildz0id. He's the one recommending to buy Kingston for AM5, since all sticks from 5600 CL36 and above are 100% Hynix. 6400+ seems to be 100% A-Die, but not sure about that.
A bit of tweaking and they are now 99,999% stable
I didn't find info about temperature stability, but I'm happy for now.

Other Programs
Close down as many programs as you can. Web browsers, sim racing background programs (wheel/pedals software, etc.), etc. You want as much unused/free RAM as possible for the RAM tests to use. The exception is a monitoring program like HWinfo (to track things like temperatures).

Yep, did that!

Here you go: https://www.overclock.net/threads/amd-ddr5-oc-and-24-7-daily-memory-stability-thread.1800926/
Tons of help & info there.

I've read about 30 pages in there since buying the 7600X in November 2022, but at some point I was like "eh, too much time being 'wasted', I'm just gonna use Buildz0id's timings and be done with it"

Either way, it's almost certain your RAM and CPU's IMC are fully stable at those speeds/settings

Yep, me too

 

[RasmusP]

This one wasn't the highest average, but the pic below it seemed to show (to the naked eye) the least spiky behaviour... But then maybe that comes and goes?

Yeah I have absolutely no idea... I did the tests as scientifically as I could (stopping all services, pausing windows updates, pausing defender etc. etc.), but the fluctuations are pretty wild...
I don't really feel the itch to repeat these tests on my Win 11 22H2, but I know the fps difference with different RAM timings was definitely there, when using a full AI grid at Spa. I remember about 10% for the average fps.

But then my Win 10 shows more than 10%+ fps than the Win 11 22H2, so maybe then the RAM doesn't make such a big difference.

It's quite weird in general. The summary I take from this all:
Always install two different Windows and compare the fps - your install might be screwed, lol

 

[Neilski]

I've read about 30 pages in there since buying the 7600X in November 2022, but at some point I was like "eh, too much time being 'wasted', I'm just gonna use Buildz0id's timings and be done with it"

Yeah, it's so complex with so many moving parts that it's hard to disagree with the "life's too short" approach... especially when Buildzoid himself appears to think nothing of having an 11-hour live stream dedicated exclusively to his painstaking efforts to tune a single RAM kit (I'm thinking of a very recent stream, which IIRC finished up with him being pretty disappointed with the outcome ).

 

[Neilski]

But then my Win 10 shows more than 10%+ fps than the Win 11 22H2, so maybe then the RAM doesn't make such a big difference.

It's quite weird in general. The summary I take from this all:
Always install two different Windows and compare the fps - your install might be screwed, lol

Yeah that 10% variation is surprisingly big. But I guess I'm considerably further into the "life's too short" camp than you (hey, I'm older ), cos nope, not gonna have two Windows installations, even to catch issues like this ;-P

 

[Durge Driven]

If I was still running 1080 all of this would be extremely relevant.
Anyway I could gain performance ( sheer fps ) I would just give it back to headroom so no point here.
I run 3440 on rig 4K and windowed mode to 4K on Neo QLED both at 144Hz
rF2 limited to 144fps LMU 120fps in json and 10% + to spare.
Towers run smoother faster cooler quieter.
So pump extra frames from them not needed.

I more into OS setup which I unmercifully strip to absolute bare essentials adding the best security and online speed.

Also got to take into account I care for nothing except rF2 and LMU so basically all tests are redundant, used to spend a lot of time on this but no more. Even 3DMark I have not installed for years even to look at new tests.

You do your tried and true perfect as possible OS, dot the i's, cross the t's ........then just leave it alone