Multiple GPU and OpenCL performance

[art3mis]

I know GPU renderers take advantage of multiple GPU's, but does H16 OpenCL performance also benefit?

[ikoon]

OpenCL microsolvers in DOP are not yet multiGPU. With DOP (I have tried pyro upres with opencl), for me (5820K / GTX1070) the cpu is still the limit. CPU is fully loaded and GPU is quite idle. Even if I switch all the microsolvers OpenCL on with a script. Maybe sobebody has another experience?

Maybe the "RAW" OpenCL kernel can use 100% GPU, if you have optimized snippet, I don't know:
http://www.sidefx.com/docs/houdini/nodes/sop/opencl

[rod.vfx.td]

 

Blender Cycles and Octane take advantage of all gpus, even on the ones on risers. it is very nice on my tests, almost absolut speed increase for each GPU.

I am imagining near future possibility to use this on FEM for muscle simulations H17??, Houdini segmenting problem on multi-gpu system, this can open a myriad of possibilities on complex simulations that fit on GPU memory.

[martinkindl83]

some results from my new machine using TITAN Xp and dual Xeons and pyro solver, no micro solvers involved

OpenCL on GPU - almost no GPU usage, high GPU ram usage, took 12:43
OpenCL on CPU- no GP usage, no GPU ram usage, took 14:25
no OpenCL - after 14:25 i had only 47% done (so its 1/2 slower compared to OpenCL)
OpenCL on GPU , all inputs for dopnet precached (cca 4min to precache all inputs) - almost no GP usage, high GPU ram usage, took 10:55

Edited September 28, 2017 by martinkindl83