What are your experiences with NEMIX RAM?

[erock]

Can you share your experiences with NEMIX RAM? I am interesting in learning about your actual experiences not speculation. Please describe your experiences in terms of product quality, performance, shipping quality and customer service.

Here is my experience with NEMIX.

I have built several non-critical machines for scientific computing using Supermicro server motherboards (H11SSL-I and H11DSi) with Epyc Naples and Rome processors (7551, 7302 and 7f52). To save money on a couple of these machines I purchased NEMIX RAM described as “Supermicro certified” with model numbers that are similar to tested RAM on the Supermicro mobo websites. So far I have purchased around 32 16GB RAM sticks and have run calculations that take multiple days running at 75% CPU load and 80% RAM load off and on for about 6 months without encountering any issues.

I purchased the RAM on Newegg with delivery occurring on time and taking around 3-5 days. Two shipments out of three involved high quality plastic RAM trays that were wrapped in bubble wrap and placed in a thick plastic bags. However, one shipment involved a large plastic RAM tray with lots of empty slots and the tray was placed in a plastic envelope without bubble wrap. The edges of this large plastic RAM tray were deformed presumably due to rough treatment during shipping and some of the deformation was getting close to where the RAM sticks were located.

Other than the one shipment issue I am am pleased with the product and have not had to exercise the lifetime limited warranty (I can’t find a lot of info on this warranty).

 

[TType85]

Ordered 4 sticks of 32gb DDR4-3200 ECC UDIMM via Amazon. Ended up getting 3 sticks of 32GB and one 8GB stick. I returned them with no issue. Ended up getting a different set from another vendor.

 

[unwind-protect]

50% of sticks we received at $OLDJOB had memtest errors.

 

[erock]

50% of sticks we received at $OLDJOB had memtest errors.

What memory tester did you use? I will try to replicate in my system.

 

[erock]

Ordered 4 sticks of 32gb DDR4-3200 ECC UDIMM via Amazon. Ended up getting 3 sticks of 32GB and one 8GB stick. I returned them with no issue. Ended up getting a different set from another vendor.

Thank you for sharing your experience. Can you share the vendor that you finally went with?

 

[TType85]

Thank you for sharing your experience. Can you share the vendor that you finally went with?

OWC since I was able to get it next day. https://www.amazon.com/OWC-3200MHz-PC4-25600-Unbuffered-Workstation/dp/B0CBT183N2

 

[erock]

50% of sticks we received at $OLDJOB had memtest errors.

Were you able to return the bad sticks? I am curious how the return process worked for you. I have seen mostly positive reviews on this mixed with some negative. Also, what kind of RAM did you test (ECC? RDIMM? Etc…)?

 

[RolloZ170]

50% of sticks we received at $OLDJOB had memtest errors.

had same with samsung. 2133Mt RAM tested @2666

 

[erock]

had same with samsung. 2133Mt RAM tested @2666

Interesting, so the “good” Samsung RAM showed similar results. What kind of ram was this? Was this ecc, ddr4 etc?

I plan on running memtest86+ per your suggestion from the other post tonight or tomorrow. I am running some quick tests with memtester in the OS now. I will report the results I get on this large batch of Supermicro compatible NEMIX RAM.

 

[RolloZ170]

Interesting, so the “good” Samsung RAM showed similar results. What kind of ram was this? Was this ecc, ddr4 etc?

M393xxx PC4-2133P

 

[unwind-protect]

What memory tester did you use? I will try to replicate in my system.

Whichever memtest variant is included with Debian.

 

[unwind-protect]

Were you able to return the bad sticks? I am curious how the return process worked for you. I have seen mostly positive reviews on this mixed with some negative. Also, what kind of RAM did you test (ECC? RDIMM? Etc…)?

Yes, but we actually exhausted the available supply of 32 GB ECC sticks at the time and also ran out of time. The machines in question have 64 GB each, not 128 as we intended.

We also tried slower speed (via memory multiplier), but it was really dead cells.

 

[Stephan]

Do yourself a favor and use Passmark Memtest86. It will show all kinds of errors which a normal memtest from Windows boot loader, memtest86+ open source version etc. will not report. Believe me I tried them all, Passmark is the only thorough one. The reason is that Passmark's knows CPU internals and can read and display memory controller internal error registers. Before some ECC correction comes in. Let it run for 24 hours.

Another method is to use stress-ng with a little zRAM like so: https://forums.servethehome.com/index.php?threads/ram-error.36932/#post-341921 Here rasdaemon on Linux will report similarly to what Passmark reports.

A third method is linux kernel compilation in RAM in a loop, e.g. makepkg on Arch from /run/user/1000/build. This will really stress out the VRMs and show if the overall system is sufficiently cooled. Makes CPU pull its full TDP, plus what RAM needs. Really close to worst case. Check sensors and, again, rasdaemon.

If all three come out as working without issues, you're good.

Had problems with Samsung in the past, prefer Micron or SK Hynix for DDR4 RDIMM. Don't buy without return policy.

 

[erock]

Do yourself a favor and use Passmark Memtest86. It will show all kinds of errors which a normal memtest from Windows boot loader, memtest86+ open source version etc. will not report. Believe me I tried all of them all, Passmark is the only thorough one. The reason is that Passmark's knows CPU internals and can read and display memory controller internal error registers. Before some ECC correction comes in. Let it run for 24 hours.

Another method is to use stress-ng with a little zRAM like so: https://forums.servethehome.com/index.php?threads/ram-error.36932/#post-341921 Here rasdaemon on Linux will report what Passmark reports.

A third method is linux kernel compilation in RAM in a loop, e.g. makepkg on Arch from /run/user/1000/build. This will really stress out the VRMs and show if the overall system is sufficiently cooled. Makes CPU pull its full TDP, plus what RAM needs. Really close to worst case. Check sensors and, again, rasdaemon.

If all three come out as working without issues, you're good.

Had problems with Samsung in the past, prefer Micron or SK Hynix for DDR4 RDIMM. Don't buy without return policy.

Thank you for the tips. I will give these tests a try and share the results.

 

[erock]

Do yourself a favor and use Passmark Memtest86. It will show all kinds of errors which a normal memtest from Windows boot loader, memtest86+ open source version etc. will not report. Believe me I tried all of them all, Passmark is the only thorough one. The reason is that Passmark's knows CPU internals and can read and display memory controller internal error registers. Before some ECC correction comes in. Let it run for 24 hours.

Another method is to use stress-ng with a little zRAM like so: https://forums.servethehome.com/index.php?threads/ram-error.36932/#post-341921 Here rasdaemon on Linux will report what Passmark reports.

A third method is linux kernel compilation in RAM in a loop, e.g. makepkg on Arch from /run/user/1000/build. This will really stress out the VRMs and show if the overall system is sufficiently cooled. Makes CPU pull its full TDP, plus what RAM needs. Really close to worst case. Check sensors and, again, rasdaemon.

If all three come out as working without issues, you're good.

Had problems with Samsung in the past, prefer Micron or SK Hynix for DDR4 RDIMM. Don't buy without return policy.

OK, I started running Passmark Memtest86 on a machine with H11DSi + 2x7f52 (BIOS 2.4) and a machine with H11DSi + 2x7302 (BIOS 2.1). Memtest86 using the defaults including the parallel option showed cpu access errors on the 7f52 machine and froze on the 7302 machine @ 1.5hr. I set the test config to 1 processor and Memtest86 is moving forward on the 7f52 machine (currently. @ 2.75 hrs) but froze again on the 7302 machine @ around 1.5h again. These are well documented problems especially with Supermicro boards and may be due to BIOS problems:

MemTest86 - Freezing and Lockups

One quick thing I am thinking about trying is to turn off SMT but if I can’t find a solution what would be the second best software for testing memory errors? Memory errors are a pain for the types of scientific calculations I do and typically cause solvers to fail. I can write checks in my code to rerun a solver if an issue occurs but it would be nice to mitigate these by identifying and replacing problem RAM sticks.

 

[Stephan]

Like I said, stress-ng. If you do not have a little zRAM configured or can't use it, you need to hike the percentage up from say 50% in steps up to 86% (my limit) until the kernel starts killing the worker processes of stress-ng (out of memory killer). Decrease by two and that's your limit. With zRAM 1/32 of RAM and basically empty Arch Linux booted from stick I can go up to 86%. Of course you want to exercise as much memory as possible.

nice stress-ng --vm $(nproc) --vm-bytes 50% --vm-keep --vm-populate --vm-madvise willneed --verify -v -t 24h --tz --perf

Start rasdaemon and tail system log looking for errors. Read https://forums.servethehome.com/index.php?threads/ram-error.36932/#post-341921 again.

 

[alex_stief]

Waitaminute... during all this time of trying to diagnose cryptic and unexplainable errors and resets, we have been dealing with Nemix RAM?
Because f them. If you still can, return that memory.

 

[Stephan]

Waitaminute...

Löl... there is a back story?

I recently saw on ebay a boat load of Cisco-labelled 2666 DDR4 RDIMM 32 GB which was Micron. Bought a few, no issues even after the nasty tests I described above. Really cheap, seller obviously wanted them just gone.

 

[alex_stief]

The backstory here is that erock has had a few problems with trying to get the 7F52 to run stable. Corrupting disks, inexplicable resets, dfferent behaviour with different CPUs. There are a few threads here.
The other part of that backstory is Nemix RAM itself. I have read from too many people who had problems with it, especially on AMD Epyc systems. I would not trust it. And I am yet to see Nemix on any memory QVL.

 

[Stephan]

There is a German saying: If you buy cheap, you buy twice.

Looks like erock needs to sail way closer to QVL.