NUMA Nodes

[w0D1n]

Hi,

 

I am wondering about NUMA Nodes and their influence with gaming.

I build a pc with a AMD Threadripper 3660X and ASUS Zenith II Extreme and i used the all on Auto settings exept for memory witch is 3600Mhz in combination with 1800 infinity fabric speed (1:1).

In auto mode the system wont use NUMA nodes and i thought that is good since i read somewhere that it would hurt gaming performance.

But after testing things for optimal settings i found out some differences as mentioned below:

(This is testes with AIDA64 Extreme and Cinabench 20. And a recent FPS Shooter on a 2080 Super. Motherboard BIOS version: 0807)

 

Without NUMA Nodes (this was even worse on earlier bios versions)

Memory latency = ~80ns

Memory copying speed = ~100 GB/s

Call of Duty Modern Warfare FPS = ~80-100 FPS

Cinabench 20 Score: ~13200

 

With NUMA nodes set to 2 nodes per socket

Memory latency = ~74ns

Memory copying speed = ~110 GB/s

Call of Duty Modern Warfare FPS = ~100-120 FPS

Cinabench 20 Score: ~13500

 

So i have set it to 2 nodes per socket obviously. I understand a little how NUMA nodes work and a extra layer should indeed not do anything good for speed..

But can someone explain to me what the NUMA nodes do in this case and why it actually increases speed/performance?

[mariushm]

I'm probably wrong, as I didn't think this through.... but my first thought would be that by enabling numa, the game's threads are locked to a single chiplet or a couple of chiplets, maybe raising the average boost frequency of the cores and using just 1-2 memory controllers. Probably getting less cache wasted (as some cache is used for swapping data between chiplets so if all threads are on same chiplet then more cache is available)

[w0D1n]

Marius, thank you for your reply.

 

I somewhat figured that indeed. Still wondering why they turn Numa nodes off by default since it hurts overall performance it seems. And with this impact i would think more people would have figured this out and talked about it. But can't seem to find out exactly. This video helped me explain it a little: