MattSimpson
New Member
Hi All,
There is a performance penalty if an application runs on cores that are on e.g. two distant CCDs I reckon they also call them chiplets.
I am looking at AMD Milan 7543 32-core model. I reckon it comprises of 4x 8-core CCDs.
In the BIOS you can make all on a single NUMA0 or 2 or 4 NUMA system but in this exercise we pursue a single NUMA configuration.
I reckon
0-7 - CCD 1
8-15 - CCD 2
16-23 - CCD 3
24-32 - CCD 4
So the application should run in one of those ranges but it would be good to know which next CCD to use if I need more cores.
Do you know how to workout the layout of the AMD processor on Linux?
Best
There is a performance penalty if an application runs on cores that are on e.g. two distant CCDs I reckon they also call them chiplets.
I am looking at AMD Milan 7543 32-core model. I reckon it comprises of 4x 8-core CCDs.
In the BIOS you can make all on a single NUMA0 or 2 or 4 NUMA system but in this exercise we pursue a single NUMA configuration.
I reckon
0-7 - CCD 1
8-15 - CCD 2
16-23 - CCD 3
24-32 - CCD 4
So the application should run in one of those ranges but it would be good to know which next CCD to use if I need more cores.
Do you know how to workout the layout of the AMD processor on Linux?
Best