Announcement

**flipbook** · 01-05-2018, 06:20 AM

Thought I'd give it ago.. I think one of your CPU's costs as much as my entire system :P

Anyway this was writing the cache to a an Evo SSD ... had about the same speeds writing to the M.2. Drive.

**Paul Oblomov** · 01-05-2018, 06:35 AM

Heh. Not bad. I wonder, where all the AMD ryzen or epycs ?

**jacksc02** · 14-08-2018, 01:53 AM

I will try it out tomorrow on 2990wx

**Paul Oblomov** · 14-08-2018, 02:09 AM

Btw, would be great to have phx logs also. Interesting what would be the slowest part.

**Svetlin.Nikolov** · 14-08-2018, 02:35 PM

Originally posted by jacksc02 View Post

I will try it out tomorrow on 2990wx

Threadrippers are unfortunately NUMA inside, so they won't perform well. You can check here for more info on the recommended hardware: https://docs.chaosgroup.com/display/...sterSimulating

Cheers!

**Raven** · 15-08-2018, 06:54 AM

@Svetlin

So I know there are many factors involved as per the recommendation document.
What machines do you guys prefer to use?
I'm looking at a new machine soon, and I can get the i9-7900 or a Threadripper 1950x for the same price.
Now I don't do tons of fluid work, but enough that I want to consider it in my setup.
If I understand the NUMA issue correctly, the threadripper would effectively be an 8 core machine vs the i9s 10 cores.

Thanks

**Paul Oblomov** · 15-08-2018, 07:15 AM

Let us see TR results first

**Svetlin.Nikolov** · 15-08-2018, 11:08 AM

georgi.zhekov should be able to say more specifically what model of TR we tested Phoenix on - it was a 16 core, 32 thread CPU and produced roughly the same results as an i7-5820K, except for some (not all) FLIP scenes where the TR is slower. We need to check the bus speed on both configurations though, because it is just as important as the CPU.

Using a NUMA CPU won't automatically exclude half the cores - you would need to do it manually from the Simulation rollout. However, with Threadrippers and some Xeons that are NUMA inside but don't register as such in Windows, this can't be done because as much as Phoenix could gather from the OS is that it's a plain old CPU. You could try to work around this by setting the thread number to half the threads and hope that Windows orders the first NUMA node cores first and the rest second (for a 2-NUMA config).

Otherwise, using all cores on a NUMA won't automatically halve the speed. The sim could work a bit slower or even more than twice slower, depending on the CPU.

**georgi.zhekov** · 16-08-2018, 01:47 AM

The TR we tested was the 1950x vs the i7-5820K and as Svetlin said it was pretty much the same result, although it has quite a lot more cores. The bus speed is around the same for both CPUs.

**Raven** · 16-08-2018, 02:40 AM

Thanks for the info guys, So I guess for now the i9 will perform better in the sim department.
Is the NUMA issue something that is solvable in the near future, or would it require a full re-look at the core, as I guess that processors like the TR2 will seam like they under perform.
Do you know if NUMA affects other simulators ie Houdini or realflow?

**nikola.bozhinov** · 20-08-2018, 03:36 AM

Hi,
The problems with the NUMA configurations (and the distributed simulations) is inherent to some of the basic algorithms of the fluid simulations, so solving them is not just implementation problem. This doesn't mean that no optimizations are possible, but it is unclear how much things can be improved with the current algorithms.
In general fluid dynamics simulations are more memory intensive than rendering and that is why multi-core and especially NUMA systems seem to under perform - the bus speed is often the limiting factor, as noted in the docs. And in the NUMA processors the different cores need different time to access the memory, so the problem becomes even greater there. Have in mind that some processors with high core count have NUMA architecture even if that is not said explicitly.

**Paul Oblomov** · 21-08-2018, 11:59 PM

I wonder, why still there's no TR benches here ? What's happened ?!

**Paul Oblomov** · 03-10-2018, 06:03 PM

I saw vlado got his hands on tr 2990. Plz can you test itat full powe and at maybe one memblock out of four (32/4) cores?

**jacksc02** · 04-10-2018, 03:08 AM

2990wx
Total elapsed time - 14m 31s

Announcement

hardware bench

hardware bench

Comment

Comment

Comment

Comment

Comment

Comment

Comment

Comment

Comment

Comment

Comment

Comment

Comment

Comment