Ryzen 1950x is very slow in TYFLOW

[Heerockss]

Hi,
first of all thank's for your great work.

I tried Tyflow on my laptop with 7700 HQ cpu and speed was great but when i tried same scene on my computer ryzen 1950x 32gb ram corsair 3600 mhz speed was 4time slower than my 7700hq.

I turned of auto multi thread and selected 16thread of my cpu
result was faster than when I simulate with all of cpu threads.

does anyone have the same problem?

That sounds like a similar problem some people do have with PhoenixFD and Ryzen CPU`s. They have some problems with NUMA that slows down the calculation considerably. 

I run into a similar problem with my dual xeon workstation. When calculating lower resolution simulations, a single numa node (less threads) are faster than with both nodes. I think its because the communication between the two nodes takes longer than the actual calculation process.

Thank's Heerockss,
I compared a simple sim in phoenixFD
7700 HQ : 9m 40 s
1950x : 4m 28s
in this comparision 1950 is 2x faster than 7700HQ.
but in tyflow the result is opposite 7700 is 2x faster than 1950.
hope Tyson gives us information about this.

I did also some testing with TyFlow on one of my dual Xeon machines: CPU: 2xXeon e5-2670v3 (24C/48T) @2.7ghz  GPU: Maxwell TitanX

2x 2670v3 (opencl off) 02:09 Min  <--- Both Numa Nodes at 100% usage but really slow


1x 2670v3 (opencl off) 01:11 Min  <--- Only one Numa Node!

1x 2670v3 (opencl on) 00:38 Min  <--- A lot faster with OpenCL and one Numa Node

2x 2670v3 (opencl on) 1:12 Min  <--- And again lot slower with both Numa Nodes


I think there is indeed a slowdown for multiple numa nodes. Why your AMD Chip performs well in Phoenxi but bad in TyFlow i cant answer but dont forget that not all Ryzen or Threadripped behaved like this in Phoenix, there are some exceptions.

Yes our king and savior Tyson Ibele might enlighten us