nvidia.com/object/nvlink.html
So basically until this kind of technology will enter our computers,
we will end up using lagging sops ?
Just for general knowledge, what are the limitations for creating\manipulating
geometry only with the gpus ?
when and why and in which cases we really need to cpu ?
is this just matter of work force and time consuming developing or real
technical\computational problem to solve ?
for example the “add” and “point” sop nodes, why are we creating and manipulating
the points\vertexs with the cpu ?
someone can try and explain\direct me to understand this issue ?
GPUs are incredibly fast at doing the things that they are good at, and slow at doing things they aren’t. GPUs excel when the data is many->many and there isn’t much interaction between each data elements. When you need to parse over entire sets of data to get a single result the benefit of GPUs greatly diminishes. For example calculating the bounding box of a geometry is a process that can be done on the GPU, but it’s far more complex to do and I’d be surprised if it’s faster than a multithreaded CPU approach.
The main reason we haven’t done GPU SOPs is purely due to how we are focusing our resources, but they would be a different set of operations than SOPs, not a 1:1 mapping.
Thanks for the reply Malcolm,
My intention of this question is to try and understand where things are going,
I know a little about the Parallel Computing limitations,from reading here computing.llnl.gov/tutorials/parallel_comp/
but still not sure how easy or not the implementation of stuff,and i`m not try to
get inside your considerations
I guess I will have to focus and learn GLSL,
Btw:
We already have 21 people signed for the course.
will SOPs be replaced by GOPs or something - like COPs were replaced with TOPs?