Thanks Marty! I'll try as soon as possible!

It works!
There's the result of the first test with an identical scene (standard sphere, standard volcano settings, just OpenCL enabled) running under Ubuntu
CPU with OpenCL = 1m01s (same as Win7)
Memory used = 8.1GB (on Win was almost 9GB)

Now testing the memory limits of my GPU (3GB) with Ubuntu … let's seeeeee …
Actually it seems it's going way better, it's now on frame 57 with 0.025 as subdivisions, but I can't see the console with the OpenCL memory usage. I've set the variable HOUDINI_OCL_REPORT_MEMORY_USE=1
but it doesn't appear … any idea?

You can double check that the OCL report memory is being used by the terminal and the command ‘hconfig’

It'll show you what is set, if the OCL mem. is set there, I'm not sure right now.

You should also double check the the Intel OpenCL is being used in Houdini and not the AMD OpenCL library. In Houdini - go to About Houdini, under help menu, and the OpenCL info should be something like:

OpenCL Platform Intel(R) OpenCL
Platform Vendor Intel(R) Corporation
Platform Version OpenCL 1.2 LINUX

This is really helpful and exactly what I was looking for! Phew!

@merlino3d, I know it's quite a old topic right now but I was curious to know if you managed to get way more out of your gpu memory on Linux than you did on Windows?

Have you guys also had the chance to compare with Windows 10 too since then?

I'm currently testing OpcenCL on GPU and it seems like I'm capped to 550ish megs while my graphics card has 3gb of gddr5 memory. I know that houdini, my dual screen setup and other app could take some memory but I'm surprised how few memory seems to be left for houdini…

the 550MB is the memory pool - OpenCL likes memory to be unfragmented.

On OsX you can use the OpenGL Drive Monitor to see actual device memory usage. i.e. Purple line is OpenCL using up all the available memory in the screensnap.

Regarding GPU memory, that may not be as much of an issue as getting data to and from the GPUs. The more memory in the GPU the more the data channel bottleneck becomes apparent. That is where I have found the bottleneck to be using GPUs.