Hi, Jason It's a GPU issue. I just want to verify the scalability of multi-GPU in windows, and the answer is disappointing. Few NVIDIA GPU cards support GPU-affinity in Windows, and SLI mode is the only way to use multi-cards. In fact, Windows NVidia driver will try to send OpenGL cammands to all GPU when SLI model disabled.
In linux, the default behavior is each GPU is bind to a screen. On Mon, Jun 25, 2012 at 10:18 PM, Jason Daly <[email protected]> wrote: > On 06/25/2012 12:59 AM, GeeKer Wang wrote: > >> Hi, all >> >> I want to use two GTS 250 cards to do the parallel rendering job based on >> OpenSceneGraph。 >> >> In my experiment, a scene full of complex models is split to 2 parts >> according to viewport。 >> The 2 parts are rendered in seperated windows (actually two slaves). >> I hope the LEFT part is rendered in the first screen and by the first >> GPU, and the RIGHT part in the second screen and by the second GPU。 >> The results are a little wierd and disappointing. >> > > > > Hi, Bob, > > The additional card will only help if your application is GPU-bound (that > is, waiting on the GPU to finish rendering a frame before the next can > start). Another possibility is that it is CPU bound (the CPU is taking > most of the time and the GPU is waiting on it), and another is that it is > interconnect bound (the bus between the CPU or RAM and the GPU is the > bottleneck). > > Just adding another card will not help if you're CPU-bound or interconnect > bound, and it can actually make things worse if your scene isn't set up in > an optimal way. You might end up trying to cram twice the amount of data > down the bus in order to try and keep both GPUs fed, and this will kill > performance even more than with a single card. > > One quick fix you can try is to run the osgUtil::Optimizer on your scene > with the option "VERTEX_PRETRANSFORM | INDEX_MESH | VERTEX_POSTTRANSFORM". > This will optimize the vertex data for the vertex cache common on modern > GPUs, and may help with bus bandwidth. The other key is to be sure that > your scene is organized and subdivided spatially, so that when you're > viewing the scene, you don't end up duplicating draw calls across the GPUs. > > Hope this helps, > > --"J" > > > ______________________________**_________________ > osg-users mailing list > osg-users@lists.**openscenegraph.org <[email protected]> > http://lists.openscenegraph.**org/listinfo.cgi/osg-users-** > openscenegraph.org<http://lists.openscenegraph.org/listinfo.cgi/osg-users-openscenegraph.org> > -- Bob
_______________________________________________ osg-users mailing list [email protected] http://lists.openscenegraph.org/listinfo.cgi/osg-users-openscenegraph.org
