NucBox K8 Plus 8-core 96GB Ryzen 7 8845HS - Random crashes under load - y-cruncher

Hardware Hub MiniPCs and Specialty PCs
1

Built many Android ROMS on old intel box (32GB RAM), upgraded to AMD NucBox K8 Plus with 96GB (two 48GB DDR5) Ubuntu 24.04, all appears OK, BUT, under “load” (android build) it fails RANDOMLY, literally every time different place, java, rust, clang, ldd, anything, and all kids of corruptions, segv, etc. etc. even with make -j 1!!!

At first reduced O2 to O0, but failures kept coming in other files, and cannot get a single clean build. Reading found many suggestions, and spent ~2 months trying them out, without much luck, it has improved, but not stable enough to build through a ROM!?!

When running on single DIMM (48GB) runs clean, each one tested separately, but, when both are in, it fails. Tried lowering watts using ryzenadj utility, and various other tools, but nothing thus far fixed.

Speed is not an issue, just stability/consistent results.. so that build can run through..

In this journey, found a pretty amazing tool, called y-cruncher:

ht t
p s ://
www. numberworld. org/y-cruncher/

it has (option 2) stress-test mode, which runs continuously until error, and sure enough, it became much simpler to reproduce without any need for Android build environment.

Within minutes or hours, the operation fails similar to:

Iteration: 1  Total Elapsed Time: 1076.350 seconds  ( 17.939 minutes )
Running BKT: Passed  Test Speed:  1.71 * 10^09  bits / sec                     
Running BBP: Passed  Test Speed:  1.54 * 10^08  terms / sec                    
Running SFTv4: Passed  Test Speed:  4.11 * 10^09  bits / sec                   
Running SNT: Passed  Test Speed:  6.18 * 10^08  bits / sec                     
Running SVT: Passed  Test Speed:  1.64 * 10^09  bits / sec                     
Running FFTv4: Passed  Test Speed:  1.06 * 10^09  bits / sec                   
Running N63:                                                                   

Exception Encountered: ConvolutionFailedException

Coefficient is too large.


Error(s) encountered on logical core 0.

Running N63: Failed  Test Speed:  2.62 * 10^08  bits / sec                     
Error encountered.  Total Elapsed Time: 1076.350 seconds  ( 17.939 minutes )

Running latest firmware on latest kernel v6:

Firmware Version: NucBox K8 Plus 1.04
   Firmware Date: Thu 2026-05-14
    Firmware Age: 3month 1w 6d

Linux 6.17.0-1032-oem #32-Ubuntu SMP PREEMPT_DYNAMIC Fri Jul 24 05:46:26 UTC 2026 x86_64 x86_64 x86_64 GNU/Linux

Manufacturer not very responsive, can’t get past the “sale support” :slight_smile:

Tried just about everything under the Sun!!
Like:

  • dusable AVX512 and similar options like “Core Performance Boost” and “global c-state”

  • Lowered to 4800MT/s biggest change to improve, cannot go lower, starts to crash worse..

  • Tweaked kernel with MANY different combinations, no luck there either, tried iommu=pt, etc. etc. etc. current flags are:

    crashkernel=512M amdgpu.gpu_recovery=1 amdgpu.runpm=0 ipv6.disable=1 split_lock_detect=off pcie_aspm=off

  • Changed lots of kernel parameters, no luck (swappiness, etc.)

  • Tried MANY combinations using yzenadj tool, no visible change on results:

yzenadj --fast-limit=35000 --slow-limit=35000

The y-crunher fail in N63 and VT3 so far..

Any tips, welcomed!! esp. which forum /group are best able to solve such an issue :slight_smile:

2

y-cruncher is the most efficient hw (mainly memory) test tool. At this point calculating pi digits is a side show (with >300t digits published).

If it takes about 10mins for y-cruncher to fail, it’s most likely a heat related issue (it takes a while to build up the heat).
You can validate this assumption by running the same test in a row, it should fail much quicker the next time.

Bad ram is exposed in a single run that utilizes most of the memory.

3

Thanks @jode for reply!
After many test, really doubt it’s bad memory, since running each DIMM separately (in same box), all is fine for 24hrs in both ROM/build and y-cruncher. memtest86 prints green PASS in all cases (single and dual dimms).. Glad to have arrived to y-cruncher (using the stress test mode, not pi calc), was unaware of that Amazing program!!

Very possible heat, but then what can one do? will it work only in Winter months? or if we run A/C to 10C? :slight_smile: Hoping to find some config/settings that it will run/smoothly, even slower, much slower, but just not fail randomly..

Tried limiting cpu, e.g (to 3.3Ghz):

 cpupower frequency-info 
analyzing CPU 2:
  driver: amd-pstate-epp
  CPUs which run at the same hardware frequency: 2
  CPUs which need to have their frequency coordinated by software: 2
  energy performance preference: performance
  hardware limits: 417 MHz - 5.10 GHz
  available cpufreq governors: performance powersave
  current policy: frequency should be within 1.09 GHz and 3.30 GHz.
                  The governor "powersave" may decide which speed to use
                  within this range.
  current CPU frequency: 1.70 GHz (asserted by call to kernel)

also lowered:

# dmidecode |grep Configure
        Configured Memory Speed: 4800 MT/s
        Configured Voltage: 1.1 V
        Configured Memory Speed: 4800 MT/s
        Configured Voltage: 1.1 V

seems most stable.. Power supply is original, funny part is right next to it a much older Intel box, running much slower and never fails.. Vendor is not returning useful replies, they do read emails, but front-line support is just a closed loop to tell u to install “latest firmware”, which we did ~3 months ago right when it came out.

If anyone else shares or knows of a solution, or method to arrive, will be glad to know.. Tried FAN on box, but it was far, maybe if the fan is DIRECTLY on the box ? :slight_smile:

Thanks!

4

Your quick tests seem to indicate that heat may be the source of your troubles.

Normally, the first step to mitigate the issue is using a more aggressive fan control option. This can be set in the UEFI/BIOS.

It’s also possible to control fan speed from the OS. I see that you’re using Linux, so maybe the following is a good resource to explore this option: Controlling Fan Speed in Linux: A Comprehensive Guide — linuxvox.com

Cheers,

Jochen

5

Thanks for reading/replying, hopefully will resolve this at one point and will update.

Meanwhile:

  1. the 48GB DDR5 dimms are OK, since running each separately, over 24hrs on y-cruncher yields no errors, only when they are together, in 96gb config, the y-cruncher fails in K8+ amd box.

  2. also both run memtest64 in bios mode, PASS many hours, pretty sure they are OK, just the combination.. also set

ryzenadj --stapm-limit=35000 --fast-limit=35000 --slow-limit=35000

and just about anything can think of, no luck.

  1. now moved same sticks to a K11 box, same BIOS, tests are running will know soon.

Does anyone happen to know where GMKTec EVO T1 BIOS lives?

Just for reference, since they make it so hard to find, here is the BIOS for K8/K11 (which are identical md5sum to each other)

h t t p s : // drive.google.com/drive/folders/1cpri9iS3NG-v_KnGfKKFHMcV-0530vcQ?hl=en
6

Just to close-loop on this, after many weeks/months, the issue appears to have been “Seating” of the memory sticks in the actual machine.. Nothing has changed in BIOS/OS/tests, kept trying testing one, and after validating memtest86 for 4+ cycles, added the second, and in some cases it was failing, other times it held just fine, even days of y-cruncher!

I guess i never realized just how “fragile” this is.. the fact it shows “96gb” or whatever number, doesn’t mean a lot.. if contacts are not good for high-load, such failures can take your time :slight_smile:

here is a screenshot from same exact machine with 96GB running y-cruncher for days!!

After it worked, cleaned all contacts with electrical cleaner and it appears to hold fine, but not taking these for a while now :slight_smile: after replacement , suggest to run memtest or y-cruncher to validate and avoid stray/sporadic errors which can corrupt rom builds, or AI workload.

image
image573×499 156 KB