GPU for simulation

   10100   4   2
User Avatar
1 posts
Joined: Jan. 2015
Hi, everyone!
My question for you - What GPU to buy for simulation in houdini?
I can take Tesla k20 (very cheap ~ 900 $) and i don't know worth it or not. For that price now sold more cards (Titan, Geforce) with similar specifications CUDA number of processors, memory speed and other parameters. The price about ~ $ 1200. But they are already on the new chipsets. Is the difference in performance between these cards? And if buy Quadro k4000 or k5000 for Tesla to make “Maximus”, will accelerate the simulation? I'm talking about opencl. Simulation smoke, fire (pyro) and flip.

Today I'm testing my Geforce 670 on smoke sphere. With the grid 300x300x300, Houdini gives an error “OpenCL Exception: clEnqueueNDRangeKernel (-4)”. Latest drivers does not helped. There is a sense in this idea simulation with the GPU? If configure a complex scene with fire and smoke in grid 1000x1000x1000 or more, it be correct considered with Tesla or Titan without errors like OpenCL Exception: clEnqueueNDRangeKernel (-4)?

p.s. sorry for my english :roll:
User Avatar
2406 posts
Joined: June 2008
Pyro seems to benefit the best from OpenCL acceleration. Fluid can use OpenCL as well but I only see about a 30% usage out of my GTX660 with 2GB from Houdini, on Windows.

There is nothing wrong with adding new parts to your computer but I wouldn't drop $1200 on a graphics card and expect Houdini to magically work better. I have a Quadro K4000 at work and it does not really do anything other than add a lot of heat to the box. Which is nice in the winter time, it keeps my feet warm.
Using Houdini Indie 19.5.
Ubuntu 64GB Ryzen 16 core.
nVidia 3050RTX 8BG RAM.
User Avatar
260 posts
Joined: Nov. 2014
no real benefit of using GPU for simulation, on serious jobs you will be limited by GPU memory and on very fast CPU the difference is negligible (on slow CPU you will see speed increase, but you will still be limited by GPU memory). Buy more RAM or faster CPU.
User Avatar
5064 posts
Joined: July 2005
For large simulations, you need lots of VRAM - like the Titan X with 12GB. The newest Nvidia drivers support 64b OpenCL addressing, so it will be able to access all 12GB (previously it was limited to ~4GB). That'll work well with your 300x300x300 sim. I believe we've tested the 12GB cards up to 500^3.

If you can find a Nvidia GEForce 700 or 900 series card with 6GB or 8GB, those will work reasonably well with larger sims as well. The K20 with 5GB might be borderline for large volume sim work. Quadros with lots of VRAM are simply too expensive to recommend.
568 posts
Joined: Nov. 2005
what driver did You use, i only get 3 gig ram for use in opencl on a titan x. that would be too cool to use 12 gig ram

tried it on windows and centos

thanks in advance
  • Quick Links