crash Prime95 errors + sometimes crashing. Only when two RAM sticks used.

[geost77]

I have new+used parts computer just built. But after few hours of use and of course some benchmarks I found there are too many crashes. This include Cinebench R20 sometimes just popup an error when I click RUN and fail to start benchmark. Sometimes after that popup I close and reopen it and it run, sometimes after I close popup I get blue screen and it have different error message, not always one and same.
Other programs also sometimes just crash with or without some error they show up.

 

I did full MemTest64 and no errors. But when I try Prime95 almost instantly I get error at Worker 20 (sometimes 2-3 sec after test start, sometimes up to a minute). Soon after Worker 19 usually also result an error.

[Nov 21 03:10] Worker starting

[Nov 21 03:10] Beginning a continuous self-test on your computer.

[Nov 21 03:10] Please read stress.txt.  Choose Test/Stop to end this test.

[Nov 21 03:10] Test 1, 31000 Lucas-Lehmer iterations of M9537183 using FMA3 FFT length 480K, Pass1=384, Pass2=1280, clm=1.

[Nov 21 03:10] FATAL ERROR: Rounding was 0.5, expected less than 0.4

[Nov 21 03:11] Hardware failure detected, consult stress.txt file.

[Nov 21 03:11] Torture Test completed 0 tests in 0 minutes - 1 errors, 0 warnings.

[Nov 21 03:11] Worker stopped.

 

After some experiments I found this: If I remove on of two RAM sticks - I don't get and error (at least at 2-3 min test, as it usually show up in first seconds). I did test both sticks one by one at all 4 slots and also in all combination of both sticks at once. Include in single channel mode. No matter how, if two sticks - error, if one - it works at any slot. No matter if at 2144Mhz or at XMP 3200MHz as advertised - Corsair Vengeance RGB Pro 16GB (2x8GB) 3200Mhz - brand new out of the box.

 

My configuration:

CPU: Ryzen 9 3900x (stock, no overclocking) - used, of course under 2 more years warranty

MB: MSI B450 Tomahawk Max - brand new out of the box, BIOS not touched/default. No XMP, no overclocks, nothing, all default. Even reset just to be sure not touched before sent to me.

RAM: Corsair Vengeance RGB Pro 16GB (2x8GB) 3200Mhz
PSU: recommended at top of the list at LTT forum - Bitfenix Whisper Gold 750W - used, but was working fine so far, before moved to this system.

VID: GTX 1060 3GB or ASUS EAH5450 (fanless, no power) - just to eliminate any doubts PSU can't power system with 1060 (while I am sure it is way above power needed for this system).

 

So I need help, any advice what may case this crashes and wrong calculations (as they happen at important scientific software too). I can't locate which part to blame - is it RAM (new, works any stick independently, just not if both at once), MB failing to support both sticks (BIOS, power settings - they should work with default, right, why need to up volts for RAM on brand new MB), or CPU failing at dual channel ... so I have no idea which part to send for fixing or replacement under warranty - they are all from different online stores here in my country. Still - is it possible to be PSU - only part without warranty ... 

 

Problem is it was hard to get all this parts already, I have no way to get more to test - like other RAM sticks, other MB, other CPU of same kind, or even PSU (other one is very old 400W Seasonic - it will have no enough power to run tests also 4 pin MB instead of 6 pin B450 have). None of my friends have anything to help me test, by changing parts and try locate it this way. So I am limited with what I have and any software I can get.

And hoping for my favorite Youtube LTT for help or at least any point/advice I can follow to locate issue. Other than just send all warranty parts to their stores and they find they are OK and I have to pay all deliveries + penalty for repair work/tests of perfect part.

Cheers from Bulgaria!

[porina]

5 minutes ago, geost77 said:

FFT length 480K

This is the interesting part when running Prime95, as it helps understand what was under stress at the time it failed. As a rough guide, multiply the FFT size by 8 to get the effective data size. Multiply that by the number of threads running at the time. I assume it is 24 for your CPU. That's about 90MB, compared to the 64MB total L3 cache I think. As the total data set is bigger than the cache, it will be working your IMC and ram also, so those could also be under suspicion. If you run one stick and not dual channel, you halve the ram bandwidth. This will lower the load on the cores significantly as it does make much use of ram bandwidth.

 

So... that doesn't narrow it down. Further testing with Prime95 is suggested. First, SMT doesn't really help in these workloads and this will slightly simplify testing (although it would also then not test the SMT parts). Run custom test with a number of threads equal to your cores (12). Use "in place" FFTs. If you leave the FFT size at 480k (set to both min and max), running one per core it will only stress the cores and L3 cache, and not so much the IMC/mem. See if that passes or fails. If you want to run with SMT, then that's 24 threads at 240k FFT to remain within cache. I'm a little cautious when the data size is close to the L3 cache limit, since there is some read only data that is used, so maybe also try a slightly lower FFT size to give more headroom. See if that passes or not, as that would point more at the cores if it fails, more at the IMC/ram if it passes.

 

BTW I had a Ryzen CPU that was stable on one mobo, but failed only Prime95 on another mobo (seemed stable for other things). I suspect power delivery on the latter is to blame. I've not bothered to see if I can work around it, maybe with a small positive voltage offset if I were to try.

[geost77]

26 minutes ago, porina said:

This is the interesting part when running Prime95, as it helps understand what was under stress at the time it failed. As a rough guide, multiply the FFT size by 8 to get the effective data size. Multiply that by the number of threads running at the time. I assume it is 24 for your CPU. That's about 90MB, compared to the 64MB total L3 cache I think. As the total data set is bigger than the cache, it will be working your IMC and ram also, so those could also be under suspicion. If you run one stick and not dual channel, you halve the ram bandwidth. This will lower the load on the cores significantly as it does make much use of ram bandwidth.

 

So... that doesn't narrow it down. Further testing with Prime95 is suggested. First, SMT doesn't really help in these workloads and this will slightly simplify testing (although it would also then not test the SMT parts). Run custom test with a number of threads equal to your cores (12). Use "in place" FFTs. If you leave the FFT size at 480k (set to both min and max), running one per core it will only stress the cores and L3 cache, and not so much the IMC/mem. See if that passes or fails. If you want to run with SMT, then that's 24 threads at 240k FFT to remain within cache. I'm a little cautious when the data size is close to the L3 cache limit, since there is some read only data that is used, so maybe also try a slightly lower FFT size to give more headroom. See if that passes or not, as that would point more at the cores if it fails, more at the IMC/ram if it passes.

 

BTW I had a Ryzen CPU that was stable on one mobo, but failed only Prime95 on another mobo (seemed stable for other things). I suspect power delivery on the latter is to blame. I've not bothered to see if I can work around it, maybe with a small positive voltage offset if I were to try.

 

Thanks for advice ... I had SMT off at BIOS. And set 12 cores and 480K at Prime95. Not sure about that "Memory to use in MB"  - screenshot of setup attached. Run with these settings. And same error just 5-6 sec after start at Worker 10:

 

[Nov 21 05:44] Worker starting
[Nov 21 05:44] Beginning a continuous self-test on your computer.
[Nov 21 05:44] Please read stress.txt.  Choose Test/Stop to end this test.
[Nov 21 05:44] Test 1, 31000 Lucas-Lehmer iterations of M9537183 using FMA3 FFT length 480K, Pass1=384, Pass2=1280, clm=1.
[Nov 21 05:44] FATAL ERROR: Rounding was 0.5, expected less than 0.4
[Nov 21 05:44] Hardware failure detected, consult stress.txt file.
[Nov 21 05:44] Torture Test completed 0 tests in 0 minutes - 1 errors, 0 warnings.
[Nov 21 05:44] Worker stopped.
 

 

 

[porina]

For the test I wrote, you need to check the box "Run FFTs in-place". That will run the code more like normal, where the data is kept in the same location. If you don't check the box, the software will move that used data throughout the ram (memory to use option). So hit the "in-place" option and see what that does.

 

BTW it is not necessary to turn off SMT in bios when doing this testing. You can just set the threads to 12 at the top of the options, so one thread per real core.

[Vishera]

1 minute ago, porina said:

For the test I wrote, you need to check the box "Run FFTs in-place". That will run the code more like normal, where the data is kept in the same location. If you don't check the box, the software will move that used data throughout the ram (memory to use option). So hit the "in-place" option and see what that does.

 

BTW it is not necessary to turn off SMT in bios when doing this testing. You can just set the threads to 12 at the top of the options, so one thread per real core.

That's a great way to test if the problem is the CPU or not!,Nice 

[geost77]

5 minutes ago, porina said:

For the test I wrote, you need to check the box "Run FFTs in-place". That will run the code more like normal, where the data is kept in the same location. If you don't check the box, the software will move that used data throughout the ram (memory to use option). So hit the "in-place" option and see what that does.

 

BTW it is not necessary to turn off SMT in bios when doing this testing. You can just set the threads to 12 at the top of the options, so one thread per real core.

Ops, I missed that one, sorry. Had it run now with SMT enabled, but with proper settings - 12 cores, 480K and "Run FFT in-place". After 3 minutes - no error ...


Before that I had it tested with "Memory to use = 32MB" and for 2 min I saw no error too. But that was before you pointed my error at "Run FFTs in-place". 

 

I am newbie ... so what all that mean ... CPU is the one that fails, not RAM, MB ... and he one I have to try send to be fixed under warranty (or replaces, no idea if that fixable).

PS: anything else to do. I will also run Prime95 later for at least 8 hours (night) or more to see if still no error with this settings (may help when I try explain issue at warranty service, as here they will usually try reject it and say it is all OK, if it load Windows ... bad service is common here  Even when all prices are almost twice higher than US.

 

[Vishera]

Just now, geost77 said:

Ops, I missed that one, sorry. Had it run now with SMT enabled, but with proper settings - 12 cores, 480K and "Run FFT in-place". After 3 minutes - no error ...


Before that I had it tested with "Memory to use = 32MB" and for 2 min I saw no error too. But that was before you pointed my error at "Run FFTs in-place". 

 

I am newbie ... so what all that mean ... CPU is the one that fails, not RAM, MB ... and he one I have to try send to be fixed under warranty (or replaces, no idea if that fixable).

PS: anything else to do. I will also run Prime95 later for at least 8 hours (night) or more to see if still no error with this settings (may help when I try explain issue at warranty service, as here they will usually try reject it and say it is all OK, if it load Windows ... bad service is common here  Even when all prices are almost twice higher than US.

 

Try one hour of Small FFTs,it runs on all cores and threads and it checks the "Run FFTs in-place" by default.

[geost77]

18 minutes ago, Vishera said:

Try one hour of Small FFTs,it runs on all cores and threads and it checks the "Run FFTs in-place" by default.

 

OK, running now 12 cores (not 24 with SMT) Small FFTs and will let you know in 1 hour. So far so good ... not usual error after just 5 sec.



I also did Small FFTs with 24 cores just before that and it failed a few Workers in few seconds. So I guess I have to do 12 cores, right ?

[geost77]

Well, not 1 hour ... 1 min into Small FFTs test (yes, longer than usual 2-5 sec) but Worker 10 again error:

 

[Nov 21 06:31] Worker starting
[Nov 21 06:31] Beginning a continuous self-test on your computer.
[Nov 21 06:31] Please read stress.txt.  Choose Test/Stop to end this test.
[Nov 21 06:31] Test 1, 220000 Lucas-Lehmer in-place iterations of M1633941 using FMA3 FFT length 80K, Pass1=320, Pass2=256, clm=2.
[Nov 21 06:32] Test 2, 220000 Lucas-Lehmer in-place iterations of M1611557 using FMA3 FFT length 80K, Pass1=320, Pass2=256, clm=2.
[Nov 21 06:32] FATAL ERROR: Rounding was 1.669145144e+16, expected less than 0.4
[Nov 21 06:32] Hardware failure detected, consult stress.txt file.
[Nov 21 06:32] Torture Test completed 1 tests in 1 minutes - 1 errors, 0 warnings.
[Nov 21 06:32] Worker stopped.
 

Also tried long run with manually set 12 cores, 480K and Run FFTs in-place - same, 1-2 min later worker 10 failed.

[Vishera]

3 minutes ago, geost77 said:

Small TFFs test ... Do you want me to try long run with manually set 12 cores, 480K and Run FFTs in-place?

No need to,we have enough data that points to the CPU causing the problem.

Downclock the CPU by 500MHz,do the Small FFTs again for one hour.

[porina]

1 hour ago, Vishera said:

No need to,we have enough data that points to the CPU causing the problem.

Downclock the CPU by 500MHz,do the Small FFTs again for one hour.

If it is failing on small FFTs like this, that points towards the CPU itself.

 

One thing I was meaning to try on my unstable system was to go a slightly different way. If the bios allows it, try adding a small positive voltage offset. Maybe +0.05v and see if that helps.

[Vishera]

Just now, porina said:

If it is failing on small FFTs like this, that points towards the CPU itself.

 

That's what i said:

1 hour ago, Vishera said:

points to the CPU causing the problem.

 

Just now, porina said:

One thing I was meaning to try on my unstable system was to go a slightly different way. If the bios allows it, try adding a small positive voltage offset. Maybe +0.05v and see if that helps.

Both lowering the clock speed and increasing the voltage should increase stability.

[geost77]

1 hour ago, Vishera said:

No need to,we have enough data that points to the CPU causing the problem.

Downclock the CPU by 500MHz,do the Small FFTs again for one hour.

Base clock at AMD website is 3800Mhz, so I was able to lower it with -500Mhz at BIOS to 3300Mhz. Small FFTs, 12 cores, 1 hour - no errors.

[porina]

12 minutes ago, Vishera said:

That's what i said:

I was agreeing with you  

 

12 minutes ago, Vishera said:

Both lowering the clock speed and increasing the voltage should increase stability.

Yup, different sides of the same coin. A 500 MHz drop is quite a bit of a performance hit (although still good to confirm if that helps), whereas the voltage increase would probably have a smaller impact due to running clocks at power limit. 

 

1 minute ago, geost77 said:

Base clock at AMD website is 3800Mhz, so I was able to lower it with -500Mhz at BIOS to 3300Mhz. Small FFTs, 12 cores, 1 hour - no errors.

Interesting, so next step may be to try returning clocks to normal, and instead see if you can apply +0.05v offset.

 

To confirm, in the original post the errors/crashes were happening with different software, not just Prime95. We're mainly using prime95 for testing and diagnosing.

[geost77]

19 minutes ago, porina said:

If it is failing on small FFTs like this, that points towards the CPU itself.

 

One thing I was meaning to try on my unstable system was to go a slightly different way. If the bios allows it, try adding a small positive voltage offset. Maybe +0.05v and see if that helps.

 

I can try this too, I guess ... Just not sure, a person here mentioned my volts are already too high by default at default BIOS settings:

 

I can change Auto to a few options and from tehre try make it specific volts. But how much volts exactly ? As at Auto they move constantly, sometimes I see at BIOS 1.350v, then 1.475 and above or below these ...

 

But I guess, as mentioned, that will be the same as lowering clock -500Mhz ... which was stable. But at default it is not.

 

[geost77]

2 minutes ago, porina said:

I was agreeing with you  

 

Yup, different sides of the same coin. A 500 MHz drop is quite a bit of a performance hit (although still good to confirm if that helps), whereas the voltage increase would probably have a smaller impact due to running clocks at power limit. 

 

Interesting, so next step may be to try returning clocks to normal, and instead see if you can apply +0.05v offset.

 

To confirm, in the original post the errors/crashes were happening with different software, not just Prime95. We're mainly using prime95 for testing and diagnosing.

Yes, computer crashed at Cinebench R20 "RUN" click instantly too and also at few others. I think 2-3 times at just normal operation, like open Photoshop or some software.
 

About +0.05v - I am not sure hat is my actual volts value, so no idea what to set ... see my post above with screenshots. BIOS keep changing that volts value every second when at Auto.

[porina]

3 minutes ago, geost77 said:

I can try this too, I guess ... Just not sure, a person here mentioned my volts are already too high by default at default BIOS settings:

It is normal for Ryzen CPUs to have high voltage at lower loads, since that is needed to boost single cores to high clock speeds.

 

Quote

I can change Auto to a few options and from tehre try make it specific volts. But how much volts exactly ? As at Auto they move constantly, sometimes I see at BIOS 1.350v, then 1.475 and above or below these ...

I'm thinking of trying +0.05v which would be under offset option I think. It should do what it says, and basically adds +0.05v to whatever voltage it would otherwise use. I'd be hesitant to go higher than that, because like you say, the idle voltages would get rather higher too.

 

Changing the voltage with frequency is both great in terms of efficiency, but also a pain in that is can be harder for you to do what you might want to.

 

Quote

But I guess, as mentioned, that will be the same as lowering clock -500Mhz ... which was stable. But at default it is not.

Even if this works, I guess the question will be it is worth using as a solution, or more likely it is time to call it a bad CPU and try to get a replacement. However note increasing voltage may be seen as overclocking, and that may be a reason to deny warranty.

[geost77]

4 minutes ago, porina said:

It is normal for Ryzen CPUs to have high voltage at lower loads, since that is needed to boost single cores to high clock speeds.

 

I'm thinking of trying +0.05v which would be under offset option I think. It should do what it says, and basically adds +0.05v to whatever voltage it would otherwise use. I'd be hesitant to go higher than that, because like you say, the idle voltages would get rather higher too.

 

Changing the voltage with frequency is both great in terms of efficiency, but also a pain in that is can be harder for you to do what you might want to.

 

Even if this works, I guess the question will be it is worth using as a solution, or more likely it is time to call it a bad CPU and try to get a replacement. However note increasing voltage may be seen as overclocking, and that may be a reason to deny warranty.

I found it and increased:

 

 

but if that will void warranty - I better change back to default and not even test it this way. As if CPU is bad I better try replace it. I did not pay $700 for second hand CPU that just can't work at default settings ...

 

 

[Vishera]

3 hours ago, porina said:

or more likely it is time to call it a bad CPU and try to get a replacement.

 

3 hours ago, geost77 said:

As if CPU is bad I better try replace it. I did not pay $700 for second hand CPU that just can't work at default settings ...

I think you should use the warranty and get a replacement since the CPU is bad.

[geost77]

13 hours ago, Vishera said:

 

I think you should use the warranty and get a replacement since the CPU is bad.

Yes, that is my plan. Will call them Monday. Unfortunately I got this CPU second hand, so I am worried they will accept warranty. As I have no idea how last owner used it and if voided warranty in any way, like over volt it ... But even if all fine - these companies here and warranty here  - they usually say - "it works, we found no issues" and that is all - it have to be complete dead to fix/replace. If it works, even out of specs - they 99% say it's perfectly fine.

So I just hope ... if not - I have no other choice but:
1/ underclock it a lot, -500MHz (tried -200Mhz and -300Mhz it still fails in a few minutes).
2/ push more volts I guess and it is stock cooler and will run even hotter with Prism (noisy one!).
3/ sell it, but I am not that person that can't sell it when I know it is bad ... so best to describe it's issue and lower price a lot and if anyone want it this way and use it with less Mhz or +volts ... it still works fast, as 24 threads so if I give it same price as 2700x or 3700x (second hand) it will be good price even with this issue I guess. And then need to get new 3900x CPU and pay the difference, which I am now fine.

Fingers crossed on warranty ...

[geost77]

Thanks all ! Just letting you know - CPU was replaced by local company. I got new one 3 days ago and it works perfectly fine now.