[gupsterg]

Sample 1 9600X ~240MB/s with 2133MHz FCLK

Spoiler: Sample 1 9600X

Sample 2 9600X ~245MB/s with 2200MHz FCLK. Needs more SOC & VDDP for 6000MT/s, didn't work at 1.05 0.925 like other, not yet fined tuned to see where it lands. Same timings between both, but at the time sample 1 set to looser tREFI (which should be a gain).

Spoiler: Sample 2 9600X

Sample 1 vs sample 2 stock VID per core, sample 1 seems better, once do CO on sample 2, will have better idea.

Spoiler

 

[cefurkan]

Any recommendations?

Prime95 RAM test can't pass more than few hours how to fix? AMD Ryzen 9 9950X and Corsair 96GB(2x48) Vengeance 6400Mhz CL32 DDR5

 

[anamolydetected😉]

I do. This exact config. 9700X + Karhu RAM Test + KGuiX interface.

At defaults I'd get ~160-173 MB/sec with 6000 MT/s.

When I set CPU Cache to Enabled, I'd get ~214-227 MB/sec.

Thx friend! I'm testing 6200 since it's finally stable on the latest bios... just enabled cache and seeing 237-241. Good to know my 2x32 kit is performing well

 

[anamolydetected😉]

Sample 1 9600X ~240MB/s with 2133MHz FCLK

Spoiler: Sample 1 9600X

View attachment 2692472

Sample 2 9600X ~245MB/s with 2200MHz FCLK. Needs more SOC & VDDP for 6000MT/s, didn't work at 1.05 0.925 like other, not yet fined tuned to see where it lands. Same timings between both, but at the time sample 1 set to looser tREFI (which should be a gain).

Spoiler: Sample 2 9600X

View attachment 2692471

Sample 1 vs sample 2 stock VID per core, sample 1 seems better, once do CO on sample 2, will have better idea.

Spoiler

View attachment 2692474

Nice, thx! I'm doing 6200 with 2200.. enabled cache now and very similar test speed to yours, although you've got a tiny bit higher speed actually, I should have a little room yet to tighten a few secondary timings. After that I'll try to drop my voltages a bit.. currently 1.3 vsoc and 1.51 VDD with 1.45 VDDIO/VDDQ. Passed OCCT 2hrs, TM5, and Karhu overnight.. pretty stoked with the latest bios making 6200 possible for this 2x32 kit.

 

[gupsterg]

When I used SMU Debug Tool on sample 1 9600X to do CO profile I was thinking why the heck are 2 cores greyed out just after core 2 and not say down at bottom.

Spoiler: Sample 1 9600X SMU Debug Tool

When I tried sample 2 9600X two differing cores were greyed out, so it looks like SMU Debug Tool shows which of the cores are disabled.

Spoiler: Sample 2 9600X SMU Debug Tool

Sample 2 9600X is defo being sold off.

Sample 1 9600X needs SOC 1.2V VDDP 1.05V for 6400MT/s C28 1:1 GDM On. Sample 2 9600X I can't even stabilise 6200MT/s C28 1:1 GDM on with SOC 1.2V VDDP 1.05V. All the same HW in use, except CPU.

On core clocks at stock sample 2 looked slightly worse then sample 1 in same testing. Now what I call level 1 CO tune where I make all cores same VID as lowest at stock, it still didn't raise clocks ro 5450MHz. You can see that in above screenshots and below is screenie side by side of VID profile record I'm creating from testing.

Spoiler

Sample 1 9600X on left, sample 2 on right.

Only thing sample 2 has going for it is FCLK 2200 (not tried higher yet), other than that poorer IMC, poorer core clocks. Don't think it will be as good for FMAX +200MHz looking at how it's behaving for stock FMAX.

 

[anamolydetected😉]

When I used SMU Debug Tool on sample 1 9600X to do CO profile I was thinking why the heck are 2 cores greyed out just after core 2 and not say down at bottom.

Spoiler: Sample 1 9600X SMU Debug Tool

View attachment 2692588

When I tried sample 2 9600X two differing cores were greyed out, so it looks like SMU Debug Tool shows which of the cores are disabled.

Spoiler: Sample 2 9600X SMU Debug Tool

View attachment 2692589

Sample 2 9600X is defo being sold off.

Sample 1 9600X needs SOC 1.2V VDDP 1.05V for 6400MT/s C28 1:1 GDM On. Sample 2 9600X I can't even stabilise 6200MT/s C28 1:1 GDM on with SOC 1.2V VDDP 1.05V. All the same HW in use, except CPU.

On core clocks at stock sample 2 looked slightly worse then sample 1 in same testing. Now what I call level 1 CO tune where I make all cores same VID as lowest at stock, it still didn't raise clocks ro 5450MHz. You can see that in above screenshots and below is screenie side by side of VID profile record I'm creating from testing.

Spoiler

Sample 1 9600X on left, sample 2 on right.

View attachment 2692596

Only thing sample 2 has going for it is FCLK 2200 (not tried higher yet), other than that poorer IMC, poorer core clocks. Don't think it will be as good for FMAX +200MHz looking at how it's behaving for stock FMAX.

By "I make all cores same VID as lowest at stock" you mean per core undervolt?

I thought most 9600x/9700x could do 2200 fclock.. or maybe I was a bit lucky with mine. But anyways, just to say, for 6200 I have errors or system freeze with 1.285 vsoc or less.. 1.3v is fully stable a couple days of testing now Karhu, TM5 and OCCT. And Vdd @6200 needs 1.55v (even 1.53v has errors) -of course this is with dual rank 2x32gb..so he's a little power hungry, haha.

And 6000 errors at 1.2v vsoc, 1.25v is stable (probably a little less works, idk yet).. with 1.4 vdd

Statuscore seems like a pretty useful tool

My 9700x I'm guessing undervolts pretty well..haven't seen others post theirs though, so idk. And i didn't push the absolute max on each core (but 6-10 higher undervolt for all of the cores had errors here and there in longer testing, so I just backed them all off).. in case you or anyone else want to compare:
My per core is:
0 -32
1 -24
2 -24
3 -16
4 -14
5 -20
6 -24
7 -24

With this I can also do +200mhz with no issues... other than it runs pretty hot, haha.

 

[Alexshunter]

By "I make all cores same VID as lowest at stock" you mean per core undervolt?

I thought most 9600x/9700x could do 2200 fclock.. or maybe I was a bit lucky with mine. But anyways, just to say, for 6200 I have errors or system freeze with 1.285 vsoc or less.. 1.3v is fully stable a couple days of testing now Karhu, TM5 and OCCT. And Vdd @6200 needs 1.55v (even 1.53v has errors) -of course this is with dual rank 2x32gb..so he's a little power hungry, haha.

And 6000 errors at 1.2v vsoc, 1.25v is stable (probably a little less works, idk yet).. with 1.4 vdd

Statuscore seems like a pretty useful tool

My 9700x I'm guessing undervolts pretty well..haven't seen others post theirs though, so idk. And i didn't push the absolute max on each core (but 6-10 higher undervolt for all of the cores had errors here and there in longer testing, so I just backed them all off).. in case you or anyone else want to compare:
My per core is:
0 -32
1 -24
2 -24
3 -16
4 -14
5 -20
6 -24
7 -24

With this I can also do +200mhz with no issues... other than it runs pretty hot, haha.

Hello, does it makes sense to give different values for each Cores? I think i read somewhere the weakest core’s voltage will be used for all Cores. Anyway i use -35 all core C0, with +150MHz and 101MHz BCLK overclock. AIDA64 stress errors at arround 15Min, but stable enough everything.

 

[gupsterg]

By "I make all cores same VID as lowest at stock" you mean per core undervolt?

I opt for a method that I call VID harmonization per core CO profiles (see OP here).

Spoiler

First I get stock VID per core, same test load per core, say Statuscore. Then whatever core (sometimes a few cores are lowest/very similar depending on CPU sample) I match every core VID to the lowest VID, that I call Level 1 per core CO profile. A core or two will be on 0 CO magnitude, these will be the better cores on CPU, tend to have lowest VID and higher clocks.

Then usually globally ticking the level 1 CO profile by say -3 across all cores (or whatever magnitude I decide) I create another profile.

Level 1 -10 -3 -7 0 -14 -15 (all cores harmonized ~1.29V)
Level 2 -16 -9 -13 -6 -20 -21 (all cores harmonized ~1.25V)

Each deeper CO profile will usually gain MHz and or lower VID and or lower temperature, again behaviour depends on CPU sample.

When I globally tick the level 1 CO profile to create level 2, 3, 4,etc, the VIDs tend to harmonise again but at lower values. There is a caveat though, say if a poor clocking core hasn't recached FMAX, then what happens is due to the clocks raising the SMU keeps voltage same, then when you tweak the CO again (may have to repeat) there will be a point where FMAX is reached, then VID starts dropping on next step of CO magnitude reduction, again all based on CPU sample.

I keep creating these levels of CO profile until a core (sometimes a few cores as some take same magnitude) reaches maximum negative CO magnitude. This I call MAX CO profile, this profile may pass some stability testing, but always tend to fail on extensive testing. So then I just roll back to another profile.

I thought most 9600x/9700x could do 2200 fclock.. or maybe I was a bit lucky with mine.

From what I have gathered AMD AGESA rules set FCLK 2000MHz for high memory clocks, so I think AMD reckon all CPUs can do 2000MHz. Then from what I'm seeing 2133MHz seems quite normal to get, 2200MHz seems like decent and some CPUs don't do that, and 2233MHz is rare.

But anyways, just to say, for 6200 I have errors or system freeze with 1.285 vsoc or less.. 1.3v is fully stable a couple days of testing now Karhu, TM5 and OCCT. And Vdd @6200 needs 1.55v (even 1.53v has errors) -of course this is with dual rank 2x32gb..so he's a little power hungry, haha.

And 6000 errors at 1.2v vsoc, 1.25v is stable (probably a little less works, idk yet).. with 1.4 vdd

I'd say that's a poorer IMC. My decent 9600X (sample 1) 6000C28 1:1 GDM On SOC 1.05 VDDP 0.925V, 6200C28 1:1 GDM On SOC 1.1V VDDP 0.95V, 6400C28 1:1 GDM On SOC 1.2V VDDP 1.05V. As you can see 6400 is where the big jump of voltage is from 6200, so really pushing it there, doubt it will do 6600 1:1, not tried as been doing other things.

Statuscore seems like a pretty useful tool

It is , I've used it since AM4 1000 series, also on 2000/3000/5000, plus Threadripper 1000/2000, now AM5 9000 (7000 didn't intertest me to get AM5), if I was honest it was how AMD flipped the 3D cache that made me buy AM5, as wanted to try 9800X3D due that and what W1zzard had in his review.

My 9700x I'm guessing undervolts pretty well..haven't seen others post theirs though, so idk. And i didn't push the absolute max on each core (but 6-10 higher undervolt for all of the cores had errors here and there in longer testing, so I just backed them all off).. in case you or anyone else want to compare:

Spoiler: My per core is:

0 -32
1 -24
2 -24
3 -16
4 -14
5 -20
6 -24
7 -24

With this I can also do +200mhz with no issues... other than it runs pretty hot, haha.

I have a 9700X as well . Not yet bin'd it as been busy first seeing if I was keeping which DDR5 kit. Now that's done, started meddling with CPUs. I do have a Crosshair, as that has ECLK, but so far the TUF has been good. What to see what the best I get out of HW on TUF, then see what the Crosshair gives me.

I'd say (especially how I set per core CO), per core CO profiles are unique to a CPU, after having done several AM4 CPUs and now 2x AM5 9000 series.

I have a bit of HW tinkering addiction 😊.

 

[gupsterg]

Hello, does it makes sense to give different values for each Cores?

Yes, as they differ.

I think i read somewhere the weakest core’s voltage will be used for all Cores.

True, as single power plane on CPU.

Just be aware CO value we set is not really just voltage.

The CO value we set tweaks how the SMU of CPU profiles the core for frequency and determines voltage to give for frequency.

Say I have a core that is reaching ~ 5400MHz on a 9600X, lets say voltage is ~1.3V. I set a -5 CO on just that core. Lets say the CPU does ~5425MHz after adjustment, the SMU may decide to give still ~1.3V, as core clock increased.

Then you readjust CO to say -8 on that core only, you may see ~5450MHz, the SMU may decide to still give it ~1.3V, as core clock is reaching FMAX.

Then you readjust CO to say -10 on that core only, you will continue to see ~5450MHz, but now the SMU will start dropping voltage, to say ~1.29V. Now further reduction of CO value will keep yielding voltage drop, whilst sustaining MAX core clock.

Anyway i use -35 all core C0, with +150MHz and 101MHz BCLK overclock. AIDA64 stress errors at arround 15Min, but stable enough everything.

If it's stable for your use case and your happy sweet. If you want to get better stability and perhaps improved performance, then per core CO is the way IMO.

Check out OP here, will be adding more soon to it .

 

[anamolydetected😉]

I opt for a method that I call VID harmonization per core CO profiles (see OP here).

Spoiler

First I get stock VID per core, same test load per core, say Statuscore. Then whatever core (sometimes a few cores are lowest/very similar depending on CPU sample) I match every core VID to the lowest VID, that I call Level 1 per core CO profile. A core or two will be on 0 CO magnitude, these will be the better cores on CPU, tend to have lowest VID and higher clocks.

Then usually globally ticking the level 1 CO profile by say -3 across all cores (or whatever magnitude I decide) I create another profile.

Level 1 -10 -3 -7 0 -14 -15 (all cores harmonized ~1.29V)
Level 2 -16 -9 -13 -6 -20 -21 (all cores harmonized ~1.25V)

Each deeper CO profile will usually gain MHz and or lower VID and or lower temperature, again behaviour depends on CPU sample.

When I globally tick the level 1 CO profile to create level 2, 3, 4,etc, the VIDs tend to harmonise again but at lower values. There is a caveat though, say if a poor clocking core hasn't recached FMAX, then what happens is due to the clocks raising the SMU keeps voltage same, then when you tweak the CO again (may have to repeat) there will be a point where FMAX is reached, then VID starts dropping on next step of CO magnitude reduction, again all based on CPU sample.

I keep creating these levels of CO profile until a core (sometimes a few cores as some take same magnitude) reaches maximum negative CO magnitude. This I call MAX CO profile, this profile may pass some stability testing, but always tend to fail on extensive testing. So then I just roll back to another profile.

From what I have gathered AMD AGESA rules set FCLK 2000MHz for high memory clocks, so I think AMD reckon all CPUs can do 2000MHz. Then from what I'm seeing 2133MHz seems quite normal to get, 2200MHz seems like decent and some CPUs don't do that, and 2233MHz is rare.



I'd say that's a poorer IMC. My decent 9600X (sample 1) 6000C28 1:1 GDM On SOC 1.05 VDDP 0.925V, 6200C28 1:1 GDM On SOC 1.1V VDDP 0.95V, 6400C28 1:1 GDM On SOC 1.2V VDDP 1.05V. As you can see 6400 is where the big jump of voltage is from 6200, so really pushing it there, doubt it will do 6600 1:1, not tried as been doing other things.



It is , I've used it since AM4 1000 series, also on 2000/3000/5000, plus Threadripper 1000/2000, now AM5 9000 (7000 didn't intertest me to get AM5), if I was honest it was how AMD flipped the 3D cache that made me buy AM5, as wanted to try 9800X3D due that and what W1zzard had in his review.



I have a 9700X as well . Not yet bin'd it as been busy first seeing if I was keeping which DDR5 kit. Now that's done, started meddling with CPUs. I do have a Crosshair, as that has ECLK, but so far the TUF has been good. What to see what the best I get out of HW on TUF, then see what the Crosshair gives me.

I'd say (especially how I set per core CO), per core CO profiles are unique to a CPU, after having done several AM4 CPUs and now 2x AM5 9000 series.

I have a bit of HW tinkering addiction 😊.

Thanks for the detailed reply! I'm going to digest this info today

What made immediate sense was your other reply "Say I have a core that is reaching ~ 5400MHz on a 9600X, lets say voltage is ~1.3V. I set a -5 CO on just that core. Lets say the CPU does ~5425MHz after adjustment, the SMU may decide to give still ~1.3V, as core clock increased. "

I definitely just learned something about AM5 per core tuning that I didn't know

I'll go back through my testing and use the statuscore app, because this is pretty interesting with your methodology! I just went for max stable per core (running corecycler app for hours and then overnight as well), found the edge of stability for each core and then backed it all off by 8-10 and called it good because I was more keen to get after memory tuning, haha.

And right, I may have a weaker IMC... but also I have seen ZERO people thus far with a 2x32gb kit and 6400 stable... and 6200 2x32 usually needs 1.28v for most Zentimings and discussions I've run across. I'm pretty tempted to swap motherboard and also at least TRY a 2x24 or 2x16 kit just to see if my IMC is in fact pretty terrible, lol. -Funny though, I have nearly the highest CPU score on 3dmark for a 9700x with any GPU and rank 1-2 in every free benchmark on there with my CPU and GPU combo.

Btw, I majored in Electronic Technologies... this kind of stuff is a fun hobby for me as well.

 

[gupsterg]

@anamolydetected😉

No worries .

I did quite an update to OP today, still plenty more to do to it.

My sample 2 9600X is for sure poorer for CO per core. Clocks have not even reached stock FMAX 5450MHz.

Even though CO profile differs between samples. You'll see level 1 to 2 and 2 to 3, the jump per level is same magnitude I was checking, ie global increase of -6.

Sample 1 left side, sample 2 right side.

Spoiler

Now just gonna start doing FMAX +200MHz CO profiling. I don't have high expectations there either for sample 2.

 

[shaze133]

Just did delid and add Thermal Grizzly Heatspreader on my 9950X last week, always trying to get noise lower in my computers but always want as much speed as possible too!

Spec:
9950X delided + Thermal Grizzly HPHS using Conductonaut
Noctua NH-D15 G2 LBC Single Fan using TG Kryonaut Extreme paste
ASUS X870E ProArt 1003 Bios
2x48G GSkill Trident Royal 6400 (pmic on my TeamGroup T-Create Expert kit blew up and this was the replacement kit I got from the shop), CPU really does not like over 6000MT, 6200 needs at least 1.29vsoc. (I did try the CPU/Ram in my old B650E Taichi and it could do 6200MT @ 1.275vsoc stable, not sure why the X870E is different possible the Taichi sends more voltage?)

I had been running a conservative CO since I got the chip per CCD -20/-25. After delid I decided to find limits of CO for my chip, at first I did the scan with some Curve Shaper values accidentally not cleared and got very conservative results. Then I cleared the CS/CO settings and used Hydra scan then just did +5 to the final values.

Figured this was a good point to start stability testing...

Cpu Platinum stability certificate - 1/25/2025 5:10:17 PM

The holy grail of CPU stability testing ( 12 hours !) : 2h CPU Large Extreme, 2h CPU Large, 2h CPU Medium, 2h CPU Small, 2h Linpack v2021, 2h CPU Large ( 1 thread, cycling )

www.ocbase.com

After the tests I went through and adjusted +1 for values where the core rating didn't really match the final CO value just for extra safety. Ended up with this...

I can't get it to fail on anything so far... both windows and linux are completely stable. Plan to do y-cruncher/p95 on single/random number of cores next I guess. Any suggestions to find instability would be great.

 

[gupsterg]

Just did delid and add Thermal Grizzly Heatspreader on my 9950X last week, always trying to get noise lower in my computers but always want as much speed as possible too!

Spec:
9950X delided + Thermal Grizzly HPHS using Conductonaut
Noctua NH-D15 G2 LBC Single Fan using TG Kryonaut Extreme paste
ASUS X870E ProArt 1003 Bios
2x48G GSkill Trident Royal 6400 (pmic on my TeamGroup T-Create Expert kit blew up and this was the replacement kit I got from the shop), CPU really does not like over 6000MT, 6200 needs at least 1.29vsoc. (I did try the CPU/Ram in my old B650E Taichi and it could do 6200MT @ 1.275vsoc stable, not sure why the X870E is different possible the Taichi sends more voltage?)

Spoiler

View attachment 2692939

I had been running a conservative CO since I got the chip per CCD -20/-25. After delid I decided to find limits of CO for my chip, at first I did the scan with some Curve Shaper values accidentally not cleared and got very conservative results. Then I cleared the CS/CO settings and used Hydra scan then just did +5 to the final values.

Figured this was a good point to start stability testing...

Spoiler

View attachment 2692935

Cpu Platinum stability certificate - 1/25/2025 5:10:17 PM

The holy grail of CPU stability testing ( 12 hours !) : 2h CPU Large Extreme, 2h CPU Large, 2h CPU Medium, 2h CPU Small, 2h Linpack v2021, 2h CPU Large ( 1 thread, cycling )

www.ocbase.com

After the tests I went through and adjusted +1 for values where the core rating didn't really match the final CO value just for extra safety. Ended up with this...

Spoiler

View attachment 2692937

I can't get it to fail on anything so far... both windows and linux are completely stable. Plan to do y-cruncher/p95 on single/random number of cores next I guess. Any suggestions to find instability would be great.

When the PMIC blew on your T-Create kit what were you doing with them? ie MT/s, voltage, temperature.

I got a 9600X which isn't liking above 6000MT/s 1:1, other 9600X is sweet up to 6400MT/s 1:1, some testing 8000MT/s 2:1.

When you tested on B650E Taichi same AGESA version? perhaps some of the ODT/resistances differed?

In my view I'd say your CO profile is quite deep and not conservative. 50 is max magnitude for CO and you have several cores at 45. What scalar value you using?

Try AIDA64 CPU FPU CACHE, that seemed to push a CO profile I was using to breaking point.

Spoiler: My CO testing

I didn't test a single test for lots of hours, as wanted to test multiple loads.

1. CoreCycler P95 PASS 12hrs.
2. CoreCycler Y-Cruncher PASS 6hrs.
3. Standard Y-Cruncher stress test 15min, 80C temp, cTDP 105W.

At this point started using a 6000C28 RAM profile with CO profile.

4. CoreCycler Y-Cruncher, cTDP 105W, PASS 3.5hrs.
5. CoreCycler AIDA64 CPU, CACHE, FPU, cTDP 105W, PASS 10.5hrs.
6. Standard Y-Cruncher stress test, 80C temp, cTDP 105W, PASS 6hrs.
7. RealBench stress mode, 16GB, 80C temp, cTDP 105W, PASS 1hr.
8. Kahru RAM Test, Cache Enabled, 80C temp, cTDP 105W, PASS 5k%.
9. AIDA64 CPU SHA3 80C temp, cTDP 105W, PASS 15min (seen mention of this in threads as a thing to do, significance no idea, had no issues running it).
10. Kahru RAM Test CACHE + FPU, 80C temp, cTDP 105W, PASS 21k%

Realbench was run as it hits CPU/RAM/GPU and I wanted to test more of the system in use.

CO profile for above on 9600X was -38 -27 -35 -24 -47 -47 Scalar: Auto

Used standard AIDA64 CPU FPU CACHE and CO profile failed ~10min. Using up to scalar 10x didn't stabilise profile for that test load.

So rolled back to -25 -24 -32 -21 -44 -44, Scalar: 1x, then passed 3.33hrs. Which looking at my profiling I set cores 0 wrong and should have been 35 , will be re-visiting when use that CPU, at present using another.

 

[shaze133]

When the PMIC blew on your T-Create kit what were you doing with them? ie MT/s, voltage, temperature.

Nothing special, I think 6200MT, 1.4V VDD/Q running Karhu, temp showed 3700C max in monitor I was seeing around 54/58C after an hour before I left it to run overnight with ambient sub 20C, after reboot one of the modules wouldnt boot, it was still "fine" before that reboot. Sent back to shop who sent it to TeamGroup who confirmed faulty PMIC. There was the normal burnt PCB smell from the stick in question.

When you tested on B650E Taichi same AGESA version? perhaps some of the ODT/resistances differed?

Same AGESA, 1.2.0.2 and 1.2.0.2B both same result. I checked on a ASUS X670E-F and same thing it can run 1.275v SOC for 6200 on this CPU/Ram. I dont really understand the ODT/resistance stuff so could be that I guess?

In my view I'd say your CO profile is quite deep and not conservative. 50 is max magnitude for CO and you have several cores at 45. What scalar value you using?

Conservative compared to what Hydra recommended, I ran it with max test setting and it recommended to run -31,-37,-30,-39,-40,-40,-39,-40,-50,-50,-50,-50,-50,-47,-50,-50.

Scalar is x2, LLC @5 I believe, its the Recommended OC value in ASUS BIOS.

Been running y-cruncher BKT cycling single cores for over 48hrs in linux now with no errors (just leaving it running constantly in background). Running normal desktop workloads also with no errors.

Try AIDA64 CPU FPU CACHE, that seemed to push a CO profile I was using to breaking point.

Will try this thanks, never realized AIDA64 has some stress test.

1. CoreCycler P95 PASS 12hrs.
2. CoreCycler Y-Cruncher PASS 6hrs.
3. Standard Y-Cruncher stress test 15min, 80C temp, cTDP 105W.

Will do p95 single thread in background next, will let y-cruncher go another 24 hours. Takes ages to go through 16 cores.

Already did I think 6 hours of standard y-cruncher component test with no issues. I get the feeling the 85C cap is helping, I really doubt this would be stable at max temps.

Initially when I tried individual CO when I got the chip I always had to set Curve shaper with positive offset for high temp values else it would be unstable with much lower values than this. Running it with 85C cap (>=90C is high temp for curve shaper) basically does the same thing but seems to cut out overshoot a lot more.

 

[gupsterg]

@shaze133

That wasn't a high voltage RAM profile on T-Create kit, so probably just a duff PMIC. Cheers for experience share .

AGESA can include IMC FW updates, I don't know how to check, but it was stated by The Stilt in some AM4 threads and I doubt AM5 doesn't also. Plus I think motherboard makers tune certain aspects, as all the latest ASUS UEFIs in changelog have:-

3.Added support for up to 5200MT/s when four 64GB memory modules (total 256GB) are installed. The exclusive AEMP option will appear when compatible models are populated.

ZenTimings/Ryzen Master can show the ODT/resistances to compare between boards, setups.

Yeah LLC5 seems stock on ASUS board from what I've seen on mine and others shares.

On AM4 and AM5 when changing scalar from Auto to manual it always shows as 2x, but I don't think Auto defaults to that, as I recall on AM4 The Stilt guided 1x as stock. So I usually use that.

Not played with Curve Shaper yet, plan to . So far only done RAM OC, CO per core.

Yeah temperature limit will limit CPU frequency, if that is hit before a power limit like PPT/TDC/EDC. As FIT is present on AM5 CPU also, even if power limits extended it must safe guard CPU, see this post by The Stilt on AM4.

 

[Mr.Sunshine]

Just did delid and add Thermal Grizzly Heatspreader on my 9950X last week, always trying to get noise lower in my computers but always want as much speed as possible too!

Spec:
9950X delided + Thermal Grizzly HPHS using Conductonaut
Noctua NH-D15 G2 LBC Single Fan using TG Kryonaut Extreme paste
ASUS X870E ProArt 1003 Bios
2x48G GSkill Trident Royal 6400 (pmic on my TeamGroup T-Create Expert kit blew up and this was the replacement kit I got from the shop), CPU really does not like over 6000MT, 6200 needs at least 1.29vsoc. (I did try the CPU/Ram in my old B650E Taichi and it could do 6200MT @ 1.275vsoc stable, not sure why the X870E is different possible the Taichi sends more voltage?)
View attachment 2692939

I had been running a conservative CO since I got the chip per CCD -20/-25. After delid I decided to find limits of CO for my chip, at first I did the scan with some Curve Shaper values accidentally not cleared and got very conservative results. Then I cleared the CS/CO settings and used Hydra scan then just did +5 to the final values.

Figured this was a good point to start stability testing...
View attachment 2692935

Cpu Platinum stability certificate - 1/25/2025 5:10:17 PM

The holy grail of CPU stability testing ( 12 hours !) : 2h CPU Large Extreme, 2h CPU Large, 2h CPU Medium, 2h CPU Small, 2h Linpack v2021, 2h CPU Large ( 1 thread, cycling )

www.ocbase.com

After the tests I went through and adjusted +1 for values where the core rating didn't really match the final CO value just for extra safety. Ended up with this...
View attachment 2692937

I can't get it to fail on anything so far... both windows and linux are completely stable. Plan to do y-cruncher/p95 on single/random number of cores next I guess. Any suggestions to find instability would be great.

Are you using PBO? Run AIDA64 stress test. I'll be shocked if you pass while running PBO @200mhz

 

[shaze133]

Are you using PBO? Run AIDA64 stress test. I'll be shocked if you pass while running PBO @200mhz

Yes I am using PBO, am not using any Freq Offset.

 

[shaze133]

@shaze133

That wasn't a high voltage RAM profile on T-Create kit, so probably just a duff PMIC. Cheers for experience share .

AGESA can include IMC FW updates, I don't know how to check, but it was stated by The Stilt in some AM4 threads and I doubt AM5 doesn't also. Plus I think motherboard makers tune certain aspects, as all the latest ASUS UEFIs in changelog have:-



ZenTimings/Ryzen Master can show the ODT/resistances to compare between boards, setups.

Yeah LLC5 seems stock on ASUS board from what I've seen on mine and others shares.

On AM4 and AM5 when changing scalar from Auto to manual it always shows as 2x, but I don't think Auto defaults to that, as I recall on AM4 The Stilt guided 1x as stock. So I usually use that.

Not played with Curve Shaper yet, plan to . So far only done RAM OC, CO per core.

Yeah temperature limit will limit CPU frequency, if that is hit before a power limit like PPT/TDC/EDC. As FIT is present on AM5 CPU also, even if power limits extended it must safe guard CPU, see this post by The Stilt on AM4.

Thank you very much for your replies, AIDA64 was what I needed, it fails after ~15minutes reliably. Now the fun of finding which cores are failing

Quickly tried adding +5 too all offsets and it went for over an hour. I will probably run these offsets over night with AIDA64 just to be sure before I start the process of elimination.

ZenTimings/Ryzen Master can show the ODT/resistances to compare between boards, setups.

Zentimings shows completely wrong info for this current board, it shows ProcODT Pu 43.6, but it is 480. I tried setting it manually but it made no difference to the value in Zentimings. I just had it all set to Auto on both boards. Not super fussed as the 6000MT setup seems very close in performance to the 6200MT where I had to dial back things like TRFC/CAS, and the ODT stuff makes sense as for why it could be different. As long as there is nothing wrong with this board I am happy.

Yeah temperature limit will limit CPU frequency, if that is hit before a power limit like PPT/TDC/EDC. As FIT is present on AM5 CPU also, even if power limits extended it must safe guard CPU, see this post by The Stilt on AM4.

Great info thank you! This is much more sensible than the crazy Intel LGA1700 platform I came from (RIP 13900K).

 

[anamolydetected😉]

Thank you very much for your replies, AIDA64 was what I needed, it fails after ~15minutes reliably. Now the fun of finding which cores are failing

Quickly tried adding +5 too all offsets and it went for over an hour. I will probably run these offsets over night with AIDA64 just to be sure before I start the process of elimination.


Zentimings shows completely wrong info for this current board, it shows ProcODT Pu 43.6, but it is 480. I tried setting it manually but it made no difference to the value in Zentimings. I just had it all set to Auto on both boards. Not super fussed as the 6000MT setup seems very close in performance to the 6200MT where I had to dial back things like TRFC/CAS, and the ODT stuff makes sense as for why it could be different. As long as there is nothing wrong with this board I am happy.



Great info thank you! This is much more sensible than the crazy Intel LGA1700 platform I came from (RIP 13900K).

You might try the CoreCycler app too.. I found it really helpful.. just bumped everything up +2 or -2 until I passed/failed and then found the Max point of everything passed one cycle per core. Of course then a few hours running the cycler found a couple more cores with errors and overnight found the rest... backed everything off by 4 from there and have been dead stable since, even after adding +200 fmax.

Running 6200 1:! with GDM disabled is another story, lol.. I try and I fail, and try some more, haha.. and then I threaten my motherboard that I'll return it, lolz

 

[gupsterg]

Thank you very much for your replies, AIDA64 was what I needed, it fails after ~15minutes reliably. Now the fun of finding which cores are failing

No worries , I see forum as place to gain and give from collective experience .

On two samples of R5 9600X my CO per core profiles have been sensitive to AIDA64 CPU FPU CACHE. @Blameless suggested test to me, originally my CO per core had been passing quite a few hours of various test loads, and to me it seemed too easy to pass.

Quickly tried adding +5 too all offsets and it went for over an hour. I will probably run these offsets over night with AIDA64 just to be sure before I start the process of elimination.

I don't know if you are aware. AM4/AM5 CPUs have a single power plane. Simply put each core when loaded on it's own, will use voltage differing to when multiple cores are loaded, when multiple cores are loaded the dominant core determines voltage to all cores, as single power plane. AFAIK cache also plays a role for final voltage.

Cache runs at CPU MHz, so when we raise CPU MHz so does cache MHz. So errors potentially can be from cache running at higher speed, besides a core MHz.

I have done several AM4 CPUs CO per core (multiple 5900X, multiple 5800X3D, single 5700X3D) and have so far on AM5 2x 9600X, 1x 9700X, 1x 9800X3D. Try the VID harmonization method I use, results can be gained quick IMO, in this post is my ~7hrs run on AIDA64 CPU FPU CACHE, that profile I gained in very little time .

Zentimings shows completely wrong info for this current board, it shows ProcODT Pu 43.6, but it is 480. I tried setting it manually but it made no difference to the value in Zentimings. I just had it all set to Auto on both boards. Not super fussed as the 6000MT setup seems very close in performance to the 6200MT where I had to dial back things like TRFC/CAS, and the ODT stuff makes sense as for why it could be different. As long as there is nothing wrong with this board I am happy.

Are you setting it at 480? if not how are you finding out is 480?

I did have a play with ODT/resistances in UEFI on my board TUF X670E (do also have Crosshair X670E Hero not used it yet). I changed some settings with Ai Tweaker > DRAM Timings Control. They didn't reflect correctly in ZenTimings/Ryzen Master.

Then I remembered from screen shots I took at UEFI defaults on AM5 there are states of RAM. See this post.

You may find changing settings with P0 on the end reflect better in ZenTimings/Ryzen Master.

Great info thank you! This is much more sensible than the crazy Intel LGA1700 platform I came from (RIP 13900K).

No problems, welcome to another defector .