Hey all, I recently upgraded my 3970x to a 3990x without reinstalling windows. (Running windows 10 education)
Unfortunately, I’m seeing 0% scaling (for my work related software) after upgrading the CPU. (My workload can only use 64 threads, however- I can open 2 windows at once). I also tried disabling SMT, no change.
Do you guys recommend reinstalling windows 10? I’m thinking since this windows was installed on a 32core cpu it may be creating issues for my work software. And if so, which version of windows 10? I’ve been told Windows 10 for workstation or enterprise may yield better results.
*for reference, I do see good (but not perfect) scaling in cinbench r20, getting up to 28500 score with pbo enabled. Not quite the 30k that some reviewers got though.
I was told by the developer of the software I run that the workload should have scaled by opening a second tab of the program. I guess now I’ll have to try a VM
Yes, it does. Running 1 window of the software will load up around 60% of my cpu. Running 2 windows of said software loads my cpu to 100% on all 128 threads; however both solvers will run about 2x slower, resulting in real gain in terms of speed.
The load doesnt bounce around according to task manager. 1 window will load up the first core group, the second window loads up the other core group. It may be something to do with the software itself.
It seems the software you use doesn’t support windows processor groups and is thus not group-aware.
For example I found that all software using the Java Programming Language is not group-aware and can thus not use more than 64 threads on Windows. Similar for some Python based tools. Software needs to be adapted to make it group-aware.
Rendering tools, like Cinebench R20 and Blender are group-aware and will use all 128 threads. Can you verify this?
I also discovered that Windows has a System Environment variable called NUMBER_OF_PROCESSORS, can you check its value?
Thank you I will ask the developer about this. Where can I check the windows variable?
Also, is there any workaround I can do from my end?
I’m happy if I can run 2x windows of the software and have them access the core groups independently of each other calculating 2 separate workloads. Just not sure how to do it. Vm didn’t seem to work, because vbox can only have 32 virtual processors assigned to it.