Hi,
Is houdini designed to use all the available cuda cores from all the available GPU cards on the system if these cards are not connected together (via nvlink or SLI connection)?

If yes, should we expect any performance improvements in OpenCL simulations if these GPUs are linked together?

OpenCL can only address one device at a time.

Not sure if there are easy answers to this but is that OpenCL in general or the current implementation in Houdini? Would hypothetical CUDA implementations of the OpenCL accelerated tools fare any differently?

In theory, from OpenCL development aspect, it's doable by making separate command queue for each device, then they would work as with multithreading, but my question here is if houdini was developed with this taken in account?

I wish to know to see if installing 4 cards in my server will make a difference.

From the system requirements page [www.sidefx.com], at the bottom:
“GPU acceleration currently does NOT make use of multiple OpenCL devices (i.e. two or more graphics cards) but this may change at a future date.”

christopherw
From the system requirements page [www.sidefx.com], at the bottom:
“GPU acceleration currently does NOT make use of multiple OpenCL devices (i.e. two or more graphics cards) but this may change at a future date.”

Thank you… That what I was looking for to know… I missed that info.

But one more question… If two cards are linked via NVLink, then is this will be supported due to the link that makes the two as one?