Page 1 of 1

Is Clarisse NUMA aware?

Unread postPosted: Sat Mar 27, 2021 5:56 pm
by machieb

I have a AMD Epyc Workstation with 2x64Cores 256Threads.
I wanted to know if Clarisse is NUMA aware, using all threads?
I have a few UEFI options to configure NUMA (NPS0, NPS1, NPS2, NPS4)
Do you have any experience?

Greetings Marcus

Re: Is Clarisse NUMA aware?

Unread postPosted: Mon Mar 29, 2021 8:24 am
by sam
Hi there,

Wow that's certainly a beast you have there!
Clarisse should use all the cores but it will ignore the locality of the core it assigns the tasks to.

Re: Is Clarisse NUMA aware?

Unread postPosted: Thu Apr 01, 2021 2:37 am
by nisachar
I want to see some benchmarks with that leviathan!

Re: Is Clarisse NUMA aware?

Unread postPosted: Mon Apr 05, 2021 3:40 pm
by machieb
Hello, I did a few benchmarks but was a little bit disappointed.

The specs are 2x64core AMD EPYC 7742 2,25GHz 512 GB DRR4 3200 ECC RAM.
Allcore Turbo is at 2.6GHz.

Maybe it is a windows 10 threading problem or has something to do with the NUMA nodes /processor groups but the 128cores/256threads don´t scale in performance as it should.
64cores scale good in performance, but with the second 64cores performance have less impact.
For example if I limt Cinebench R23 to 128threads I get a score of 68341points. If I raise the core limit to 192cores I get 73710points and with 256threads I get 76414points.
In Cinebench R20 Multithreaded I get a score of 28555 compared to a Threadripper 3990X which has about 23407. Thats about +20%
In Blender Open Data benchmark I got this results:
bmw27 22,62; classroom 59,36; fishy cat 31,01; koro 58,74; pavillion_barcelona 60,78; victor 147,25
A Threadripper 3970X renders the same scenes in: bmw27 54,18; classroom 143,81; fishy cat 71,01; koro 111,03; pavillion_barcelona 147,22; victor 278,99
In Clarisse I compared with a Threadripper 3970X 32Cores. I rendered an image with the EPYCs in 2min4sec, the Threadripper 3970X needed 5min 16sec for the same image. Thats about 2,54x
So my conclusion is that in the moment, more than 64cores/128threads is a waste of money for non server related tasks.
In Clarisse and Blender I´m around 40% faster than a threadripper 3990X but for a high price (9600€ for the two CPUs).

One good thing is that the system runs cooler and therefore more sillent because of the lower processor frequency. On the other hand because oft he two cpu sockets, it consumes more power.

Re: Is Clarisse NUMA aware?

Unread postPosted: Tue Apr 06, 2021 10:34 am
by kai01w
Would be interesting to see if using linux would make a difference.

Re: Is Clarisse NUMA aware?

Unread postPosted: Tue Apr 06, 2021 3:14 pm
by machieb
Yes that would be interessing indeed! I heard that Linux can handle massive core counts better than windows, but I not familiar with Linux.