Titan X and 64bit OpenCl computation

[sekow]

Hey, 

 

at the brink of having Octane in Houdini, and a (kinda) Nvidia vram support for 64bit opencl, I am thinking of upgrading my GPUs to two Titan X cards.

I am aware that Sli wont do much with OpenCl, but alot for gpu rendering. Octane Benchmarks are batshit crazy with such a rig.

 

What I am concerned about are some reports of the double precision capabilities with Maxwell chips, as these wont match those of the first Titan card which had aimed at the computation market.

How would this affect the OpenCl computation within Houdini?

I believe that the Maxwell Quadros would have the same issue. 

 

 

 

 

[malexander]

From what I've heard, Houdini's OpenCL code only uses double precision in a few limited areas where such precision is needed - such as keeping track of total error. Overall this shouldn't impact the speed much, even with FP64 computation running 32x slower than FP32. I think the other gains from the Maxwell architecture would easily make up for these occasional double precision ops.

[lukeiamyourfather]

The Titan X works just fine with OpenCL enabled for FLIP and Pyro simulations. Not sure exactly what is going on under the hood but the performance I'm getting from it would indicate that it's mostly FP32 being used in Houdini. I'm a little surprised the Quadro M6000 (workstation cousin) uses the exact same GM200 chip and has the same abysmal FP64 performance.

Edited May 28, 2015 by lukeiamyourfather

[kurtw]

Thats interesting I was under the impression I'd have to track down a used Titan GTX/first generation board for the FP64 support.

 

So you can get away with openCL pyro/flip on a TitanX/Z without any performance loss? That's a huge savings over the M6000.

[johner]

You should be in great shape with a Titan X.  Recent NVIDIA OpenCL drivers give access to the full 12GB for good size pyro sims, and you don't need fast 64-bit support.  As Mark said Houdini mostly uses 32-bit floats internally.  You might need to switch the FLIP Viscosity solver to 32-bit if accelerating with OpenCL (a really good idea), but otherwise it's all 32-bit by default.

[haggi]

From what I could read in the docs, OpenCL is used only in a few areas. And it has the best effect if you use a very reduced setup like the example with smoke solver. But as soon as you introduce non OpenCL accelerated nodes, the speed drops a lot. That would mean that in more complex simulations the speed benefit will be quite low. I've done some tests with my Titan X and there was no real benefit except for the grain solver. But maybe my tests are too naive.

 

@edit: Just tested a high viscose flip again and got a speed increase 40%.

Edited September 7, 2015 by haggi

[Guest tar]

But as soon as you introduce non OpenCL accelerated nodes, the speed drops a lot. 

 

Worth considering too that the future will more than likely have more OpenCL nodes developed, the hardware of PCIe buses and memory transfers will become more efficient, i.e. PCIe v4 & NVLINK,  and probably a whole host of other factors will make OpenCL incredible important. Definitely worth building the infrastructure for.