Btw, read the measurements in this thread with one thing in mind - they can be much faster or much slower depending on the bus speed of the motherboard. Fluids and simulations in general need more factors than just the CPU alone...
Announcement
Collapse
No announcement yet.
hardware bench
Collapse
X
-
Also, there are not apples to apples comparisons except for the ones posted at the same time by the same person. People keep changing versions (build) of Phoenix.
So I would not draw any conclusions from this thread. It is only confusing people
- Likes 1
Comment
-
Of course, both of you are right, but, as we don't have a real tool like the Vray Benchmark, it's still interesting to see how some computers perform.
Just speaking about rendering, you even have some big differences with "similar builds". But as soon as you change the mainboard, the RAM, the bios setup, everything can change.
To be honest, I still didn't really catch everything about the bus speed and how to find the right hardware if you specifically think about building a rig for simulation.
- Likes 1
Comment
-
Downloaded 4.30 (main branch not nightly) and re-tested the 5950x here are my new results:
Code:CPU AMD RYZEN 5950x 4.875Ghz [32 threads] Total elapsed 8m 41s Performance 4.33 vox/s
Last edited by TekVisual; 08-12-2020, 12:51 PM.
- Likes 1
Comment
-
Originally posted by georgi.zhekov View PostTested with the Car tire burnout scene for 3ds Max and Maya with the Preview disabled during simulation.
Link to the 3ds Max scene - https://docs.chaosgroup.com/display/...Cartireburnout
Link to the Maya scene - https://docs.chaosgroup.com/display/...CarTireBurnout
Code:AMD 3990x 128 threads @ 4.0 GHz Memory 3400 mhz SSD Samsung 970 EVO Plus 3ds Max - 4.30.01 Nightly, Build ID: 2020112730466 [B]Total Elapsed 10 min 22 s[/B] Maya - 4.30.01 Nightly, Build ID: 2020112730466 [B]Total Elapsed 22 min 10 s[/B]
Code:AMD 3970x 64 threads @ 4.23 GHz Memory 3200 mhz 3ds Max - 4.30.01 Nightly, Build ID: 2020112730466 [B]Total Elapsed 9 min 36 s[/B] Maya - 4.30.01 Nightly, Build ID: 2020112730466 [B]Total Elapsed 19 min 46 s[/B]
Code:AMD 2950x 32 threads @ 3.90 GHz Memory 3000 mhz Maya - 4.30.00 [B]Total Elapsed 31 min 04 s[/B]
Comment
-
Originally posted by Paul Oblomov View PostI already h8u!!!111
Interesting, that I have moar vox per sec, but you have less time
Comment
-
Hi, Im new here.
I just want to talk about the Phoenix performance.
First the CarTireBurnout
By default 9m 03s
Compression disabled 8m 36s https://vimeo.com/503901221/be27066608
I like the Droplet Splash so much. https://docs.chaosgroup.com/display/...emplate+Scenes
17min. by default settings. https://vimeo.com/503900534/fd26acce8c
I'm not a programmer, but it seems like Phoenix use an old engine.
I found this on reddit.
https://www.reddit.com/r/matlab/comm...depath_on_amd/
I made a little benchmark to understand what the poor performance really is.
The simple liquid scene runs at ~360M vox/s (DDR3800 FCLK1900, CL15/14/12/22) AMD 5950X 4500MHz https://vimeo.com/503902127/db00187f4c
The same liquid scene with slower Ram at 330M vox/s (DDR3200 FCLK1600, CL15/14/12/22)
Too bad the AMD 5950X supports only dual channel memory.
The next test, Steps Per Frame.
100 Steps at 360M vox/s https://vimeo.com/503902127/db00187f4c
1 Step at 50M vox/s
That's a huge difference. It take a long time to write down the Frame in 3dsmax.
6sec for display 1 Frame in 3dsmax is a very long time.
The raw calculating power seems fast but writing back the data to 3dsmax is very slow, you can notice that with the very small Grid size.
The CPU cores are bored to death.
It seems there is a big latency between the cores.
Please don't blame me, but isn't it better and faster to slice the 3D Grid in the 2D Map ?
For example
Last edited by Aino; 24-01-2021, 03:52 AM.
Comment
-
Hey, these are great tests!
I'm afraid your conclusions are not in the right direction though. It would be great if there was one thing we could just tick and make everything magically 10 times faster, but reality doesn't work like that. First off, the grid-based fire/smoke simulations and the FLIP liquid simulations are handled by completely different solvers. One of these has about 15 simulation stages, the other has about 20. They all do different things with memory and cores, and scale differently too. Different scenes are often bound by the speed of different stages. So it's about improving each of these separately, and there are very few things that affect the whole simulation pipeline.
Since Phoenix 3, we've sped up the simulations on average with about 50%, and of course some setups benefit much more than others, so we have scenes running many times faster than before. The performance gains with systems with many cores are also higher than the average - Phoenix 3 did not scale well at all, and there is still much we can do, so it's a matter of time and balance between this, and also providing features and bugfixes, so everyone is happySo more speedup is definitely possible. I am especially looking at low-res simulations. The CPU loading there can be improved a lot. Such sims have not really been our focus for a long time, since not many clients care about low resolutions, but it's still good to have everything running at full speed regardless.
Cheers!Svetlin Nikolov, Ex Phoenix team lead
- Likes 1
Comment
Comment