Hello,

I build new machine specifically for the work with pyro and explosions in Houdini. What is better for me?
2x 1070 or 1x 1080? (both from Gigabyte)

I want new Titan X but the price and availability is really bad.

Technical specifications of new PC:
Asus X99
i7 5960X
64GB RAM DDR4

Thanks so much

I will boldly say use 1 1080GTX. The 1080 is faster than the 1070, it uses a little more power so make sure your power supply can handle that. Also not many software, (even games) can use 2 cards even if you have them. I have seen no claim that Houdini can make efficient use of 2 GPUs in one computer. Some GPU rendering engines can, however, but not Mantra.

In my experience OpenCL enabled for simulations only gets you about 30% faster. So no matter what GPU you choose the benefit is only marginal for simulations.

I have a 1080GTX and it works very well with Houdini other than a couple minor annoyances. If the screen goes into power saving mode Houdini will crash, I still need to log that bug. The only reason to have multiple video cards would be if you plan on using GPU accelerated renderers like Red Shift and Octane

Thanks so much. I am thinking about new TitanX Pascal, but for this moment I just need GPU for viewport. How many % can help TitanX Pascal to i7-5960X in pyro simulation?

Maybe 2x old TitanX? I find on Amazon for 500€.

Houdini 16 Should take more advantages of GPUs for what I heard. Also if you are want to use any of the GPU renders out there RedShift, etc … 2 cards might be worthy.

30%? I tried enabling opencl on CPU which lets you use the full amount of system memory while being up to 2 or 3 times faster. The smallest improvement you actually get with flip. I have the feeling it doesnt really matter wheter opencl is activated or not.

At a guess OpenCL will become more important as multiprocessor heterogeneous and threaded chipsets are the only way we are going to get more performance in the future.

30% was the amount of usage I could see reported by my GPU monitoring software when I enable Use OpenCL on the solver and let the sim cook. I can see the GPU is in use by that amount.

I guess what I don't get is why only 30%? If there are still numbers to crunch why not try to send more of that to the GPU per-frame? Or better yet, why can't I increase the resolution of my FLIP sim by 70% and see 100% usage on the GPU?

More importantly Realflow 10 is claiming great GPU speed 'creases of 3x etc.

30% sounds like a ratio of what is processed on the GPU vs CPU.