Data for AMD Threadripper 16 core 2950x CPU for inxi

Manjaro Development QA
1

I’ve hit an issue but the person did not provide the required data. For the AMD 16 core threadripper 2950x which is I believe Zen+ generation (right before Zen2), there’s a problem with how the data is working, a 16 core 2950x, which had 2 cpu dies, is showing up as an 8 core with 2 cpu dies, which is wrong.

I came up with a simple 1 liner that will crudely supply the data I need to debug this issue:

for i in $(find /sys/devices/system/cpu/ -type f); do echo ${i}::$(cat $i 2>/dev/null);done | sort > cpu-data-sys.txt; cp -f /proc/cpuinfo cpu-data-cpuinfo.txt

This will generate 2 files; cpu-data-sys.txt and cpu-data-cpuinfo.txt

which you can upload to a pastebin or whatever, then I can use those to emulate the cpu to see what’s going on there, but I can’t guess, and I can’t resolve the issue without that data.

You can also see easily on your system if inxi shows your core count as half the real count, like if you have a 16 core and inxi says 1x cores: 8 with -Cxx.

Note this is not about multithreading, which of course will show 2x the threads since 2 threads per physical core.

All it takes is one person with this cpu, or one enough like to generate this error in output, which is easy to test with inxi -Cxx.

I’m hoping to nail down this issue before next inxi since I know it’s real on at least one user system, and I also know I cannot guess what the data is doing to make iget half the total cores but the right dies.

Here’s your chance to help make inxi better!! Normally I don’t have to post for this type of data, but this is so specific that there’s no way I can guess. I had a dataset from threadripper 3990x, which is zen 2, but it works fine, I tested that data and it shows the expected 64 cores, 128 threads, no issues. So the issue may only have existed briefly in early Ryzen cpus with 2 dies. It does not exist on Epyc as far as I can tell.

2 Likes
3

@linux-aarhus has a Threadripper. Maybe he can provide feedback.

4

Re hardware forums, odds drop a lot because they must be running linux kernel, which is where the data comes from, and have /proc/cpuinfo.

I try to avoid non linux areas due to the very low quality signal to noise ratio for what I need info on. I use them for raw data etc, like techpowerup, who are great, but maybe I’ll try.

This is annoying because the actual issue reporter could take the roughly 15 secoinds it takes to run those commands, upload the files, and off we go. No idea why he won’t do it.

1 Like
5

That was either a spam bot or user copying & pasting AI garbage. The related post has been flagged as spam and the user account is in review.

1 Like
6

Oh, you’re right, I thought that read a bit weird, it has that generic ChatGPT structure, nicely abstract and basically just copying what general matching topics have as content.

Forum modding is going to get a lot harder sigh… I feel for you guys.

7

I can do that - I am not sure how helpful it is as the requested cpu is 2950x and mine is 5945wx - but if it can be of any help - hardware info follows

cpu-data-cpuinto.txt → https://termbin.com/49h8
cpu-data-sys.txt → https://termbin.com/c699

1 Like
8

lets see if Twitter can help …

9

linux-aarhus, very much appreciated, thank you.

I’ll post back tomorrow and let you know if the data let me fix the thing, though it may be a hard fix requiring a somewhat core refactor of the deep cpu logic, but I’ll see.

thanks for taking the time to provide this, God Jul

10

I ran this, the issue is not solved, though the data sample is appreciated anyway since it helps further pin down where the issue lies.

12 core AMD Ryzen Threadripper PRO 5945WX s in the above data, Zen 3 era.

So this issue is really narrow, it’s apparently maybe only found on the Zen+ series with 2 or more dies. This data shows that there are not > 1 die listed, so that appears to be something that only happened quite early in the zen series, which is maybe why it simply did not show up in testing.

I’m going to have to think about this one, because trying to do this without the real data is difficult, since there are some real guesses involved, which however require some core refactors, which are not easy to do, so it’s safest to not guess and wait for the real data.

I’m hoping I have at least one dataset that I didn’t really notice had only half the cores listed.

Note that if anyone has a rzyen cpu that inxi shows only half the real core counts for, you probably have this issue, and that is the data I need.

But these samples are good too because I think now I’ve seen clearly that zen 2 and zen 3 do not have this issue.

11

I recall that for the first and second generation Ryzen, each die was assigned to a NUMA node for management, which is also one of the reasons for the lower performance compared to the newer Ryzen processors, right?