Multiple 9th Gen Builds at Stock Failing OCCT Linpack

[jerubedo]

Hi everyone. I've come across quite an oddity that I'm hoping someone has any knowledge on, or that at least someone else can confirm my findings. It seems as if there's a bug involving OCCT 5.0 and up when using the Linpack 2019 test and with the 9700K and 9900K (possible other 9th gen CPUs as well). Here's what happened: I built a new rig for a friend using the 9900K. I overclocked it for him and did some stress testing. I passed 8 hours of Prime95 small FFTs, 12 hours of Prime95 blend, and 12 hours of OCCT 4.5.1 Linpack (2012 version used in 4.5.1). Then I decided to check out the latest version of OCCT just for consistency and to see how the new UI looks. So I did so and to my surprise I got several errors. So then I put EVERYTHING back to stock and ran 5.1.1 again and it failed again. So then I thought I might have faulty RAM, so I changed it out and ran again. Again, failed with several errors. Then I thought maybe it was a motherboard issue, like maybe I bent a pin by accident? Although I doubted it because I'm pretty careful. So I grabbed another motherboard and ran again. Again failure. Then I decided to try to run 5.1.1 on my own rig with a 9700K and a known stable OC. It failed there too! Then I put everything at stock and ran again. Failure again. Then  I decided to call up a buddy who also has a 9700K and no OC. He's never OCed it and just runs it at stock. He also got errors when running. Then I tried version 5.1.0 and 5.0.0 and both of them had the same behavior as well. I even tried the latest beta as well, and the same thing happens. Finally I tried 5.1.1 on my 6700K build and it ran JUST FINE for 12 hours and no errors, so it's not like it's just the application. It's the combination of version 5.0 and up and the 9700K or 9900K or other 9th gen maybe. This is driving me nuts. These systems all pass 24 hours of Prime95 blend and OCCT 4.5.1 Linpack (2012). Any ideas?

 

Here's the system specs for all 3 systems along with screen shots for 2 of them:

 

The 9900K build:

9900K

Z390 Aorus Ultra

16 GB DDR4 RAM 3000MHz C16

RTX 2080

Corsair RMx 650w

Windows 10 1903

 

The 9700K build:

9700K

Z390 Aorus Pro

16 GB DDR4 RAM 3200MHz C14

GTX 1080 Ti

EVGA G2 750w

Windows 10 1903

 

My friend's 9700K build:

9700K

Z390 Aorus Pro

16GB DDR4 RAM 3000MHz C16

RTX 2070

Corsair TXM 550w

Windows 10 1809

 

Screen shots for 9900K build and 9700K build:

 

 

 

Also sorry for accidentally covering up the max temps in that second screen shot, by the time I realized the window positions I had already closed everything out. I can tell you, though, that temps were all in the mid-70s.

 

So, is there some inherent problem with the test itself or with the CPUs? Or just some fundamental incompatibility?

[jerubedo]

Update: I also found this as well:

 

 It looks like someone else had the same issue with a stock 9900K, but they didn't post which version of OCCT they used. I'd wager it was 5.0 or up based on the posting date, though.

 

[jerubedo]

A modest bump. Does anyone have any feedback at all? 

[tincanalley]

16 minutes ago, jerubedo said:

A modest bump. Does anyone have any feedback at all? 

I have the same issue with my 9900K and OCCT 5.X.X.  I can do all the other test but linpack and small sets.  Both of those produce a large number of errors.  I have run Prime95 small sets for hours without issue, but there must be something in the latest versions of OCCT that isn't handled the same as the older versions.  I have learned to skip those tests and rely on others and daily use to feel confident my OC is stable.  So far I am at 5.1GHz, no AVX offset and at 1.295V.  I'm going to be pushing for 5.2 soon.

[jerubedo]

8 hours ago, tincanalley said:

I have the same issue with my 9900K and OCCT 5.X.X.  I can do all the other test but linpack and small sets.  Both of those produce a large number of errors.  I have run Prime95 small sets for hours without issue, but there must be something in the latest versions of OCCT that isn't handled the same as the older versions.  I have learned to skip those tests and rely on others and daily use to feel confident my OC is stable.  So far I am at 5.1GHz, no AVX offset and at 1.295V.  I'm going to be pushing for 5.2 soon.

Thanks for the reply. I just spent the last 8 hours running the small data set in OCCT, and both stock systems do pass that test. They ONLY fail in 2019 Linpack. So if you are seeing errors on the small dataset then there likely is instability, even if it seemingly doesn't affect your 24/7 usage.

[Falkentyne]

39 minutes ago, jerubedo said:

Thanks for the reply. I just spent the last 8 hours running the small data set in OCCT, and both stock systems do pass that test. They ONLY fail in 2019 Linpack. So if you are seeing errors on the small dataset then there likely is instability, even if it seemingly doesn't affect your 24/7 usage.

Can you please download prime95 29.8 build 3 from their website (this is the most current version and more relevant to modern CPUs), and do a small FFT (FMA3) stress test?

This should test the 24k-96k FFT range or something around there.

 

Tell me if you can pass this.  Also have HWinfo64 running in the background so you can look for WHEA correctable logged errors as well.

[jerubedo]

1 minute ago, Falkentyne said:

Can you please download prime95 29.8 build 3 from their website (this is the most current version and more relevant to modern CPUs), and do a small FFT (FMA3) stress test?

This should test the 24k-96k FFT range or something around there.

 

Tell me if you can pass this.  Also have HWinfo64 running in the background so you can look for WHEA correctable logged errors as well.

Yes, 29.8 is what I used for the small FFT tests and I was indeed running HWinfo64. No errors on reported for an 8 hour run on any of the systems, and no WHEA errors logged in HWinfo64.

[Falkentyne]

Just now, jerubedo said:

Yes, 29.8 is what I used for the small FFT tests and I was indeed running HWinfo64. No errors on reported for an 8 hour run on any of the systems, and no WHEA errors logged in HWinfo64.

Was FMA3 enabled or disabled?

[jerubedo]

Just now, Falkentyne said:

Was FMA3 enabled or disabled?

enabled

[Falkentyne]

Interesting.  What memory size did you test in OCCT?

Perhaps it's the IMC at fault here and you need to increase VCCIO and VCCSA?

Prime95 small FFT FMA3 does not hit main memory at all below 112K-128K+.  But OCCT tests main memory.  

Years ago when the linpack binaries were first released, it was said Intel themselves used them for stress testing their CPU's.  This was...my god...must have been around 2006 ?

 

*Edit*

Try this program and see if you can pass it.

This should heat up that CPU even more than FMA3 small FFT or OCCT (it uses linpack binaries too).

I forgot what sample size/memory size to use though!

 

LinX-v0.9.1-English.zip

[jerubedo]

1 minute ago, Falkentyne said:

Interesting.  What memory size did you test in OCCT?

Perhaps it's the IMC at fault here and you need to increase VCCIO and VCCSA?

Prime95 small FFT FMA3 does not hit main memory at all below 112K-128K+.  But OCCT tests main memory.  

Years ago when the linpack binaries were first released, it was said Intel themselves used them for stress testing their CPU's.  This was...my god...must have been around 2006 ?

I used 90% which wound up using somewhere in the ballpark of 15GB of RAM. Do you think it would be the VCCIO/VCCSA on all three stock systems, though? They are indeed all Gigabyte boards from the same lineup (Aorus). So is your assumption that all three of the boards simply aren't providing enough voltage at stock settings to the IMC?

[Falkentyne]

43 minutes ago, jerubedo said:

I used 90% which wound up using somewhere in the ballpark of 15GB of RAM. Do you think it would be the VCCIO/VCCSA on all three stock systems, though? They are indeed all Gigabyte boards from the same lineup (Aorus). So is your assumption that all three of the boards simply aren't providing enough voltage at stock settings to the IMC?

I honestly don't know.

system shouldn't crash at stock regardless, *ESPECIALLY* if RAM is set to 2133 mhz and not XMP.  I do know there are 'issues' with microcode bugs in 7700K-9900K which caused Apex Legends to crash (this was verified to be a bug in the processor) which was worked around in the semi recent patch, but I have no idea about OCCT.  Did you try the program I posted?  It puts a bigger amps load than OCCT, and also uses linpack binaries.

[jerubedo]

1 minute ago, Falkentyne said:

I honestly don't know.

system shouldn't crash at stock regardless, *ESPECIALLY* if RAM is set to 2133 mhz and not XMP.  I do know there are 'issues' with microcode bugs in 7700K-9900K which caused Apex Legends to crash (this was verified to be a bug in the processor) which was worked around in the semi recent patch, but I have no idea about OCCT.  Did you try the program I posted?  It puts a bigger amps load than OCCT, and also uses linpack binaries.

Yes, to be clear, it's the Linpack test I'm using in OCCT. The LinX program you posted uses Linpack 2012, which passes no problem for 24 hours, but Linpack 2019 reports errors.

[jerubedo]

If anyone has any further ideas, I'd love to hear them. Right now it's looking like the test is throwing false positives perhaps.

[DizzyLizzard]

Thank goodness for this thread!!!! I just built a 9700K computer and it was my first time and I thought I messed it up bad! It fails at stock quickly on the Linpack 2019 test. So this is normal for a stock 9th gen to fail the 2019 Linpack test? I actually RMA'ed my RAM already and my motherboard, and I was about to RMA the CPU next. Am I good? All other tests pass. I am using the latest BIOS on my Aorus Elite board. Thank you for the tips fellas

[jerubedo]

On ‎7‎/‎1‎/‎2019 at 11:47 PM, DizzyLizzard said:

Thank goodness for this thread!!!! I just built a 9700K computer and it was my first time and I thought I messed it up bad! It fails at stock quickly on the Linpack 2019 test. So this is normal for a stock 9th gen to fail the 2019 Linpack test? I actually RMA'ed my RAM already and my motherboard, and I was about to RMA the CPU next. Am I good? All other tests pass. I am using the latest BIOS on my Aorus Elite board. Thank you for the tips fellas

I wouldn't say it's "normal" per se. Nothing should fail at stock. But it is looking like some fundamental bug in either OCCT's implementation of the 2019 Linpack test or a bug in the 2019 Linpack test itself when dealing with 9th gen Intel CPUs (maybe 8th gen too, it would be great if we could get a confirmation on that from anyone on here).

[SteveOrlando]

I'm having the same exact issue here on my brand new 9900K at stock that I just installed today. MOBO is the Aorus Ultra. It fails the Linpack test (2019) every....single....time. It doesn't fail the older Linpack tests or any other test. I can't really tell if this thread determined if this was normal or not? Is this normal? Do 9900K's simply fail this test at stock? I'm using the latest version of OCCT. Windows 10, all updates installed.

[jerubedo]

On 10/18/2019 at 1:36 AM, SteveOrlando said:

I'm having the same exact issue here on my brand new 9900K at stock that I just installed today. MOBO is the Aorus Ultra. It fails the Linpack test (2019) every....single....time. It doesn't fail the older Linpack tests or any other test. I can't really tell if this thread determined if this was normal or not? Is this normal? Do 9900K's simply fail this test at stock? I'm using the latest version of OCCT. Windows 10, all updates installed.

Yes, I've determined over testing dozens of systems now that the 9900K and the 9700K both fail the 2019 Linpack test at stock settings and with a multitude of motherboards (not just Gigabyte). I still can't say why it fails (is it a software defect specifics to these CPUs?), but it does. For stability testing use 2012 Linpack instead along with Prime95. 

[Falkentyne]

5 hours ago, jerubedo said:

Yes, I've determined over testing dozens of systems now that the 9900K and the 9700K both fail the 2019 Linpack test at stock settings and with a multitude of motherboards (not just Gigabyte). I still can't say why it fails (is it a software defect specifics to these CPUs?), but it does. For stability testing use 2012 Linpack instead along with Prime95. 

Do they fail LinX 0.9.5 or 0.9.6 (35000 sample size?)

These are even more stressful than OCCT's linpack.

0.9.6 uses the most recent linpack binaries.  0.9.5 also uses recent binaries.

Give one of those two a try, with a 10-20 run set.  And i know why OCCT is failing now, and it is simply because the CPU is unstable in that test.  It's not a flaw in OCCT.  It's because the boards are undervolting too much (lying to the processor about what AC Loadline it should be using at stock).

 

You can pass prime95 and fail LinX 0.9.5.  Even the brutally hard FMA3 112K fixed FFT test (this is the hardest FFT to pass for FMA3; note that 36K is the hardest to pass for AVX1).

 

Also OCCT doesn't tell you what "errors detected" even means.  LinX 0.9.6 shows you what failed.

[jerubedo]

3 hours ago, Falkentyne said:

Do they fail LinX 0.9.5 or 0.9.6 (35000 sample size?)

These are even more stressful than OCCT's linpack.

0.9.6 uses the most recent linpack binaries.  0.9.5 also uses recent binaries.

Give one of those two a try, with a 10-20 run set.  And i know why OCCT is failing now, and it is simply because the CPU is unstable in that test.  It's not a flaw in OCCT.  It's because the boards are undervolting too much (lying to the processor about what AC Loadline it should be using at stock).

 

You can pass prime95 and fail LinX 0.9.5.  Even the brutally hard FMA3 112K fixed FFT test (this is the hardest FFT to pass for FMA3; note that 36K is the hardest to pass for AVX1).

 

Also OCCT doesn't tell you what "errors detected" even means.  LinX 0.9.6 shows you what failed.

I grabbed a copy of version 0.9.5 (couldn't find 0.9.6 anywhere) and so far its been running for 3 hours (over 20 runs completed) without any errors (OCCT LinPack 2019 would have usually failed by now). I'll continue running it for a few hours more for good measure. I used the 35000 sample size as you suggested, all threads, and all available RAM.

 

How did you determine that the issue is the under-volting involving LLC?

 

I can also use versions 1.0.0 or 1.1.1 if that might help as well, let me know.

[Falkentyne]

 

1 hour ago, jerubedo said:

I grabbed a copy of version 0.9.5 (couldn't find 0.9.6 anywhere) and so far its been running for 3 hours (over 20 runs completed) without any errors (OCCT LinPack 2019 would have usually failed by now). I'll continue running it for a few hours more for good measure. I used the 35000 sample size as you suggested, all threads, and all available RAM.

 

How did you determine that the issue is the under-volting involving LLC?

 

I can also use versions 1.0.0 or 1.1.1 if that might help as well, let me know.

https://hwtips.tistory.com/1611

 

Grabbed a copy of the latest buggy Gigabyte Z390 Master beta bios on tweaktown beta thread (GUI overhaul though is gud)

 

(tREFI doesn't work, 9998-65534 sets it to 65534, a few other options missing, CPU internal VR settings still do nothing except AC/DC loadline values)

Disabled Thermal Velocity Boost voltage optimizations, set CPU to 4.7 ghz, and AC Loadline to 95 (0.95 mOhms).

This stopped the Gigabyte stock undervolt (TvB could not be disabled in previous bioses as it wasn't exposed).  This gave me a higher load VR VOUT (about 1.125-1.135v at 165 amps) and current draw with intel default loadline calibration (Standard/Normal) with auto voltage, but it passed everything.  Note: stock does NOT mean MCE is enabled (that's auto overclocking) and it implies VRM Loadline (Loadline cal) is left at intel specified mOhms (1.6 mOhms for 8 cores).  Note that the maximum AC Loadline is equal to the default VRM loadline.  AC loadline can be lowered to make a CPU undervolt to save power.

 

Note, I can get the same VR VOUT with TvB voltage optimizations enabled (drops CPU VID 1.5mv every 1C as the CPU is cooler, starting at 100C; TvB disabled uses the base VID as if the CPU were operating at TJMax), by raisng AC Loadline to 110 (1.1 mOhms).

LinX 0.9.5, OCCT small data set with AVX2 (this uses exactly the same current and wattage as small FFT 15K FMA3 prime95 29.8 v6), OCCT Linpack 2019.

Nothing failed.

At a lower load VR VOUT (like 1.15v, by lowering the AC Loadline setting), Prime95 FMA3 would pass everything, but LinX 0.9.5 would occasionally give a wrong residual.  Like maybe one or two residuals in 20 loops.  Of course that is beyond diamond level stability you're aiming for to pass that, but still...

When OCCT says errors, it simply means a residual was wrong.  It just doesn't show you, unlike LinX.

 

Some stuff about boards lying to the processor about its desired AC Loadline to try to make it undervolt:

 

https://rog.asus.com/forum/showthread.php?106375-MCE-explanations-and-others&highlight=explanations

 

[jerubedo]

24 minutes ago, Falkentyne said:

At a lower load VR VOUT (like 1.15v, by lowering the AC Loadline setting), Prime95 FMA3 would pass everything, but LinX 0.9.5 would occasionally give a wrong residual.  Like maybe one or two residuals in 20 loops.  Of course that is beyond diamond level stability you're aiming for to pass that, but still...

What's odd in my case, though, is that 40 or so loops produced no wrong residuals for me at stock (and yes, MCE is disabled) in LinX 0.9.5. I'm trying 0.9.6 now from the link you provided (although I can't read anything in it lol). I'll see if anything is different with this version. For now it remains that the only place I've seen an error is with the LinPack 2019 test. I'll update with my findings after a few hours of running 0.9.6. I might also try 1.1.1 as well just for good measure.

[SteveOrlando]

4 hours ago, Falkentyne said:

Do they fail LinX 0.9.5 or 0.9.6 (35000 sample size?)

These are even more stressful than OCCT's linpack.

0.9.6 uses the most recent linpack binaries.  0.9.5 also uses recent binaries.

Give one of those two a try, with a 10-20 run set.  And i know why OCCT is failing now, and it is simply because the CPU is unstable in that test.  It's not a flaw in OCCT.  It's because the boards are undervolting too much (lying to the processor about what AC Loadline it should be using at stock).

 

You can pass prime95 and fail LinX 0.9.5.  Even the brutally hard FMA3 112K fixed FFT test (this is the hardest FFT to pass for FMA3; note that 36K is the hardest to pass for AVX1).

 

Also OCCT doesn't tell you what "errors detected" even means.  LinX 0.9.6 shows you what failed.

I just finished 20 runs in 0.9.6. I don't think there were any errors reported. Linpack (2019) still shows errors pretty quickly though. 

[Falkentyne]

11 minutes ago, SteveOrlando said:

I just finished 20 runs in 0.9.6. I don't think there were any errors reported. Linpack (2019) still shows errors pretty quickly though. 

Did you use the newest OCCT?  I ran the test for about an hour and stopped it.  I couldn't deal with the fans repeatedly ramping up and down every 10 seconds.  At least LinX 35000 runs full speed for about a minute.  I'm using Noctua 3x 3000 RPM industrial fans so yeah.  And I'm sick of the OCCT nag screen.  With how powerful prime95 29.8 build 6 and how much of a power virus LinX 0.9.6 35000 is, there's not even a need for OCCT.  Plus seeing scientific exponent residuals being different is a bit more interesting than "errors detected."

 

Was I supposed to set a specific option in OCCT linpack?  I just pressed go with whatever it gave me.  I guess I used a too small memory size?

 

LinX can do that crap with the fans on a tiny sample size too.

I'll try again later if you guys tell me to git gud.  Right now I'm torturing my CPU with FMA3 112K in place FFT again.

[jerubedo]

49 minutes ago, Falkentyne said:

Did you use the newest OCCT?

I use 5.1.0, which it looks like isn't the newest one anymore, but it's only a few months old. I can try the new 5.4.0 version, too.

 

50 minutes ago, Falkentyne said:

Was I supposed to set a specific option in OCCT linpack?  I just pressed go with whatever it gave me.  I guess I used a too small memory size?

I use the default settings myself, under the linpack tab, which is 90% RAM usage and the default size.

 

Also, 20 runs of 0.9.6 passed just fine, as seen here. I think I'll try 1.1.1 next and I'll keep you posted: