is there a way to see how much memory a sim will use? either system or gpu. Turbulence for example just shows you, with this voxel resolution you'll use max this much ram, which makes setting your division size easy to max out GPU ram.

also turbulence has a setting where once it runs out of GPU ram it switches to CPU, Is this possible. It seems like currently once you run out, the sim just errors out.

thanks
Chris

Open a Textport window and type the word ‘memory’. This will display how much memory Houdini is currently using. Because simulations change over time you may want to move to a frame where the simulation is in full swing and type ‘memory’.

thanks for the response, I'm more interested in how much vram it will use. I have a tendency to have gpu-z up all the time because gpu rendering, so I can see there how much vram a sim is using. I'd just like to have it so I can see total that will be used if the whole volume is used, so I can set the division size not to go over.

or have it jump to CPU when GPU ram is done.

with Turbulence, there are lots of things to change, to save ram, “don't use certain channels, or go slower but use less ram” which lets you get the biggest sim possible out of your system. With Houdini it kinda feels like click OpenCL and it will be faster until it runs out at which point you get problems. Really just guess work as far as I can tell, which seems not the houdini way

As far as I can tell OpenCL is still kind of a pipe dream. I don't see any significant benefit at this time of H15.5. If you are putting all your eggs in the OpenCL basket you will be disappointed. At best I have been able to measure a 30% usage of my GPU but when you run out of vRAM, as you often do in larger sims, Houdini will just crash and you will have lost the time you invested up to that point.

So I always stick to CPU and ignore all those OpenCL checkboxes that SideFX has so gingerly sprinkled throughout the system. But go ahead and give them a try, I did. Spend a day turning them on and off and conducting tests and you may come to the same conclusion. You may be better off spending time overclocking your CPU. A faster Ghz does improve sim times.

But I would love to see an example HIP file where OpenCL significantly decreases sim time in a complex simulation.

@chris - can you log an feature request for the fail-safe fallback to CPU OpenCL - it's a good idea. Under the Support menu - Rfe. Also log one to calculate the memory needed. It just needs to be coded up. It's not really guess work but requires you to calculate the amount of data in the fields. That could be automated.

You can currently write out Checkpoint or Sim files, so when the GPU eventually runs out of Ram and crashes, you can then keep simulating with the CPU openCL.

It's best to test OpenCL yourself - Pyro just ran 50% quicker in a test I just did. Sounds great to me!

OpenCL is faster if it can fit in GPU memory for me, I have a titanX so 12gigs and it's fast, so it seems worth it to use it, about %50 faster on the pyro sim I've been working on, However, that being said, Turbulence is about 10x faster with cuda on the same card vs cpu same system, so It seems like there is some room for improvement

I'll submit feature requests, and see what they say,

thanks

Got to ask, why bother with Houdini if it's 10x faster in Turbulence?

@Enivob: OpenCL…pipe dream?….I take it you havent heard of what's coming in H16 yet huh?

I have not heard or read a word. I hope it's great!

Artye
Got to ask, why bother with Houdini if it's 10x faster in Turbulence?

Oh no question there are shots that are better to do in Turbulence, however, and this is a big however, Turbulence gives you very little control, and is an island inside lightwave, can't save out vdb, and you can't direct it nearly as well as you can in Houdini. it's very basic by comparison with Houdini, Not saying it's not great and hasn't got the job done for years, But there are things I can do in Houdini that just aren't possible in Turbulence, Big feeling sims, are easier to accomplish in Houdini,

If I could take the solver speed of Turbulence, and have it in Houdini with Houdini's control absolutely that would be great, also Turbulence 2 looks amazing, you can use multiple GPU's and they all take chunks adaptability allowing for huge sims. It's not out yet, so it's not a fair comparison, but it looks amazing for big size of sims.

http://on-demand.gputechconf.com/gtc/2015/video/S5611.html [on-demand.gputechconf.com]

check out 10 mins in. So that all being said, I have used TFD for years, And it does have some features I would like to have in Houdini, but If I had to spend the rest of my life with one, I would go Houdini,

Nice! Cool to see adaptive domains too.