[Bf-cycles] Cycles use in Visualization

Brecht Van Lommel brechtvanlommel at pandora.be
Sat Oct 1 12:11:37 CEST 2011


Hi,

On Thu, Sep 29, 2011 at 11:35 PM, Matheson, Michael A.
<mathesonma at ornl.gov> wrote:
> At the OLCF at Oak Ridge National Laboratory ( an Open Science
> Supercomputing site for the Department of Energy ), we routinely use blender
> ( along with other packages ) for scientific visualization.  Our models
> probably differ than many users of blender in that they are the output of
> simulations from some of the large computers in the world.  Over 100M
> polygons is common with some in the billion polygon range ( and increasing
> all the time ) as one use.  We’ve recently been experimenting with Cycles
> and are impressed with capabilities it would give us.  We find that we are
> often constrained with blender-internal by the time to create the
> acceleration structures ( seems to be a serial process ) and by the memory
> footprint.

At the moment, Cycles isn't any better at handling big datasets than
Blender Internal, and building the acceleration structure is still
serial. It's intended to become multithreaded though, but scaling this
across many more cores than a typical desktop computer would probably
still require a more advanced algorithm.

> All of our polygons have vertex colors so possible future plans
> for Cycles using packed byte colors is a huge win for us ( I’m lobbying for
> the need ).  However, we don’t understand how to render vertex colored
> models within Cycles and we’d like to know whether it is implemented.

Vertex colors aren't supported yet, but they will be soon, code is
already there in the core.

> We also see a different type of cpu load with cycles versus blender internal –
> it seems more difficult to fully utilize all cores ( although it clearly is
> much faster ).  I’m sure we use hardware resources that are not at all
> typical or representative of blender/cycles users.  Our newest computer is
> around the corner with 100s of thousands of AMD cores and 1000s of nVidia
> GPUs so the models will explode in size.  I appreciate any information we
> receive.

In principle a brute force path tracer can scale well across a large
amount of cores, but in practice it probably requires quite some work
to get this practical. I'd like to do some experiments with this once,
getting a single image to render interactively on the blender
institute renderfarm, but again it's not something that already works,
just plans and ideas at this point.

Brecht.


More information about the Bf-cycles mailing list