Hey guys,

thinking to upgrade my old Firepro v7900 GPU on my workstation. One thing is already sure: no AMD GPU ever again. I use Houdini in Windows and Linux and I have to say, that AMD is really doing a horrible job when it comes to drivers and Linux support. That is also the main reason why I want to upgrade.
I mainly do VFX and fluid sims and I was thinking to upgrade to a Quadro m4000 or a GTX 980 TI. I have to say, I love the Firepro drivers under Windows - they just work perfectly fine with no artifacts nor glitches/lags in the viewport. So I'm kinda unsure if a GFX would give me the same experience or if I would be better off with a Quadro. Seen that the price difference between the 2 is not that big, it is not about the money but rather about the performance and experience that I have while working in Houdini and Maya.

Thanks for your input!

Sorry Vivo i can't answer your questions but i have very similar questions.
So i hope it's ok when i attach my questions to yours.

I want to upgrade my workstation too.
But i don't know which specs are the most important for houdini to simulate fast.
“Core Clock”? “Memory Clock” or maybe “Memory Size”?
Does Houdini use SLI or does it work with two different cards with different specs?

What is a good gpu setup for houdini?

Thank you guys for your help!

Vivo3d
I mainly do VFX and fluid sims and I was thinking to upgrade to a Quadro m4000 or a GTX 980 TI.

I would go with the 980ti. Houdini 14 and up supports higher-end GEForce cards. If you're doing sim work, try to find an 8GB variant of the 980ti, or go with the 12GB TitanX.

The M4000 isn't a very good card for the price when the Titan and GEForce Ti's are factored in.

Moritz Schneider
But i don't know which specs are the most important for houdini to simulate fast. “Core Clock”? “Memory Clock” or maybe “Memory Size”?

For sims, Memory size foremost (8GB+) because it doesn't matter how fast the card is if the sim won't run on the GPU Then shader count (more = better), and finally shader clock. Memory clock is pretty consistent at the high end.

Does Houdini use SLI or does it work with two different cards with different specs?

Houdini does not use SLI (doesn't work with windowed apps), but it can use a different GPU for simulation than graphics. In your houdini.env file (in your preference dir, $HOME/houdiniX.X), add:

HOUDINI_OCL_DEVICENUMBER = 1

to use the secondary graphics card for OpenCL simulation.

The TitanX is a really good card for simulation.

twod

Vivo3d
I mainly do VFX and fluid sims and I was thinking to upgrade to a Quadro m4000 or a GTX 980 TI.

I would go with the 980ti. Houdini 14 and up supports higher-end GEForce cards. If you're doing sim work, try to find an 8GB variant of the 980ti, or go with the 12GB TitanX.

The M4000 isn't a very good card for the price when the Titan and GEForce Ti's are factored in.

Awesome, thanks a lot for your answer. Coming from you, I know that it is a reliable information.
Did they ever produce the 8GB version of the 980ti? Couldn't find them online.

Vivo3d
Did they ever produce the 8GB version of the 980ti? Couldn't find them online.

Actually, looks like the 980Ti has a standard 6GB - so the 12GB version is likely the TitanX sorry about that.

thank you for the great answer!
So do you think the titan x is so much better than the 980 ti?

I mean cost effectiveness. Is the titan so much faster?

Thank you for your help!

No, at the high-end of things you pay much more for smaller performance gains. The price/performance sweet spot is usually the 960 or 950 series for Nvidia. Above that you pay more for smaller performance gains. Below that, the base price is generally too high for the performance you get.

Whether you should go for the TitanX or the 980ti depends on what you're simulating. The TitanX gives you the extra headroom to handle very large sims. But for the same sized sims that both the 980ti and the TitanX can both process, the TitanX won't be much faster.

It even looks like that on certain compute benchmarks the 980 ti is faster than the Titan X. On more memory hungry ones, obviously the Titan X has an advantage. Seen my actual GPU has only 2GB, I never tried to use openCL to run a flip sim. So wondering how big of a sim you can load into a 6GB GPU compared to a 12GB CPU. Usually I let my dual CPU do the long sim jobs.

Not all of Flip is OpenCL accelerated. You can dig through the nodes to see what is a OpenCL branch.

If I activate openCL on it, will it still use the CPUs for those nodes too or just the GPU?

Damn, just found a new m4000 for less than a 980 ti + water block.

The nodes that are OpenCL will run on the GPU, but there are Flip nodes that are CPU only.

You can also run OpenCL on the CPU. You will get a speed boost and can use all the CPU ram too.

I remember doing some benchmark tests (years ago) and my Xeons where outperformed by my old Firepro. Going to see to do some new tests

The biggest gain with FLIP is if you are using viscosity. If you aren't using viscosity, the OpenCL doesn't buy you too much, IIRC.

The grain solver gains substantially by using OpenCL. So much so we'd like to have it on by default at some point. It also works with smaller memory sets.

The Pyro Solver can have significant OpenCL improvements, especially when troublesome operations like Turbulence are avoided. But this is where the memory requirements are most important.

Too bad that the GPU can't access the system memory for OpenCL computation. But I think that 6 GB should be enough for quiet a few scenes.

Sorry for interrupting, but I have a question after following this thread.
Does pyro use openCL as default computation system or should I do some changes every time I setup a fluid simulation?
Thanks