Better core scaling?

Started by Kadri, June 15, 2020, 07:14:20 am

Previous topic - Next topic

Kadri

Hi Matt. If you compare Blenders core scaling to Terragen's i wanted to ask if there is anything new you can say.
With so much new and relatively affordable AMD CPU's it is more important then it was i think.
Can we expect better core scaling in the near future?

https://opendata.blender.org/

WAS

Is Cycles now NUMA aware? It wasn't last year in 2019. A EPYC only saw half cores being used, while it was still tiering the top of the list.

I do think our benchmarks are interesting and I wonder if there is bias in how TG handles the CPUs compared to Blender, or if people's systems benchmarking TG aren't well optimized, like lower end MBs etc.

Kadri

June 20, 2020, 12:51:35 pm #2 Last Edit: June 20, 2020, 12:54:52 pm by Kadri
I don't know Jordan.

With both SMT ON i will do a basic test with Lightwave one with 16 and 32 cores.
It is an older (v.11.63) version but it had multi core support since many years.
Just curious.

Kadri


I tried with Lightwave with 4,8,16 and 32 cores restricted within Lightwave and with task manager affinity setting:

4_cores_367_seconds
8_cores_187_seconds
16_cores_97_seconds
32_cores_53_seconds

I tried the same with a basic scene in terragen with the same method:

4_cores_263_seconds
8_cores_142_seconds
16_cores_80_seconds
32_cores_48_seconds

I expected worse actually with Terragen. It isn't so bad as i thought it would be. Not sure what to say. A longer benchmark (with the benchmark scene) might be better maybe. After 16 cores it looks like it is getting a little worse but i can live with that.

WAS

There is some added time with both situations but it seems TG is a little worse.

Kadri

Yes. Curious how much worse it is with more cores.