What system resources do fluid/liquid sims need?

   7775   9   1
User Avatar
Member
210 posts
Joined: Jan. 2014
Offline
The title is actually pretty clear.
I was playing around with this stuff and espacially with liquid sims my machine quickly is overchallanged. I'd say that somewhere around 10 million particles is my machines max, beyond that even canceling the sim can take 30 minutes or longer. All over the www thare videos of artists showcasing their sims with, I don't know, 15, 20 even up to 30 million particles.
What kind of machines do they have?
What hardware is important for a powerful machine, concerning liquid sims in particular? CPU, RAM, graphics card, all?
…and how dense should a highres sim be? What is highres and what lowres? I mean when your working at a big studio, where you can simulate your scene on a farm, 10 - 20 million particles might be lowres but for the avarage human, what resolution seems to be appropriate?
User Avatar
Member
1755 posts
Joined: March 2014
Offline
RAM RAM baby…

How much RAM do you have?
A powerful CPU is also needed, obviously.
User Avatar
Member
210 posts
Joined: Jan. 2014
Offline
Oh sorry, of course:

Win 7 x64
AMD FX-8320 (8 core, 3.5 GHz)
12 GB RAM
Nvidia GeForce GT 240
User Avatar
Member
637 posts
Joined: June 2006
Offline
the 12gb is low at least 32gb for ig sims. what can help is to run it on linux because the ram usage is lower (i think up to 30%).
User Avatar
Member
210 posts
Joined: Jan. 2014
Offline
32 GB? AT LEAST?! Where am I supposed to get that from. I think my mainboard supports only up to 16 GB.
User Avatar
Staff
4177 posts
Joined: Sept. 2007
Offline
While more RAM is always great, that doesn't always magically make everything faster. Your CPU speed, Hard Drive, and Graphics Card have a huge impact as well. Have you confirmed that you're running out of memory in the Task Manager? If you can get an SSD, that should help too. And Houdini under Linux is much more efficient. Make sure your graphics drivers are up to date.

As far as density of particles, it totally depends on the shot; Igor posted some tests on Vimeo a year or so ago (H13), this gives you an idea of scale: https://vimeo.com/86802065 [vimeo.com]. And his workstation specs in the comments were:

On every video that i posted peopls ask for workstation spec, its 4-5yearts old
I7 950 3.05Ghz
24GB 1333Mhz
8800GT 512MB ram
Ubuntu 12.04

So he was simming 16million particles in ~7 hours on that machine. Also remember, your collision geometry can have a huge impact as well (if it's deforming uncessarily or if its way more detailed than needed).

Hope that helps inform expectations a bit!
I'm o.d.d.
User Avatar
Member
210 posts
Joined: Jan. 2014
Offline
goldleaf
Your CPU speed, Hard Drive, and Graphics Card have a huge impact as well.
Really, the graphics card has an impact on particle sims?

goldleaf
Have you confirmed that you're running out of memory in the Task Manager?
Yes, the peak level was somewhere around 11.3 - 11.5, so pretty much out of memory, yes.

goldleaf
And Houdini under Linux is much more efficient.
Indeed I was already considering having a Linux setup parallel to my Windows (main) system but so far I havn' found a way to conviniently share files between them. I was actually surprised to see what it takes to share files that are on the same hard disc. It's bizarre that with dropbox it's easier to share files between seperate machines than on one and the same machine.

goldleaf
So he was simming 16million particles in ~7 hours on that machine.
I can do that too, on about 5 frames

goldleaf
Hope that helps inform expectations a bit!
Yes, at least it brings clarity. That means for now my liquid experiments are over and I will focus on stuff my machine can handle.
Thanks
User Avatar
Member
2551 posts
Joined: June 2008
Offline
What you really need, above all else, is free disk space. No matter how much memory you have you will eventually have to write either .sim files or .bgeo files to the disk and this is when you will need free space.

You can configure Houdini to not use RAM at all for fluid sims and have it spool directly to the disk. This is one solution to simming on a lower spec machines. This means your sim will run a bit slower (the speed of your disk drive) but you may be able to sim above your computer's typical ability.

OpenCL graphics card acceleration is turned OFF by default so to leverage that you need to enable it in the solver. In my experience it is about a 30% increase in simming speed when enabled.

You can also employ basic tricks to sim lower and render higher. For instance sim with less particles, skin the fluid and then subdivide afterwards. Also you can use displacement materials to make your fluid look more complex than it really is. Also let's not forget our most useful tool, motion blur. After all that fluid is moving so blur the flank out of it.

Having said that, it always comes back to solving for the shot, not just solving for some science experiment to test a system. We can all invent a scene that our computer can not solve. Look at the shot requirements and the camera placement and scale needed and begin there.
Using Houdini Indie 20.0
Windows 11 64GB Ryzen 16 core.
nVidia 3050RTX 8BG RAM.
User Avatar
Member
240 posts
Joined: April 2014
Offline
When you guys run out of RAM on a sim, what happens? It seems that for many of you the system starts using the disc - which is strange because even with a 64 GB system, Houdini will crash on my Windows 8 system when it hits the RAM limit.

I've also found that a simulation with a low particle count but a large velocity field takes up a ton more RAM than a high particle count / small velocity field sim.
User Avatar
Staff
4177 posts
Joined: Sept. 2007
Offline
Graphics card does influence sim speed in a gui session, independent of OpenCL in solvers. After cooking, Houdini has to draw stuff in the viewport. You can see this by running a sim in hbatch vs Houdini, and either timing it yourself by hand or using the Perfromance Monitor.

Also, viewing collision geometry can also slow things down in many cases, though this gets better with each release; still not recommended to always view collision get though, just when debugging/verifying collisions. Obviously your system will determine how much this impacts, but with detailed colliders, I could see a difference even on a Quadro K5000 at a place I used to work.
I'm o.d.d.
  • Quick Links