Page 1 of 1

support data larger than the GPU memory?

PostPosted: Mon Jul 12, 2010 1:54 am
by ethanph
I would like to ask that whether CULA (any versions) can support LAPACK routings that the data cannot fit into the GPU memory (for example, the input matrix is too large to fit into the GPU memory)?

Re: support data larger than the GPU memory?

PostPosted: Mon Jul 12, 2010 8:32 am
by john
I'm afraid that at this time we do not support data spilling past the GPU's memory size. That is actually a rather difficult problem (closely related to multi-GPU work) that will take some time to have implemented. It is likely that it will be enabled on a per-routine basis, as each routine splits differently.

Re: support data larger than the GPU memory?

PostPosted: Mon Jul 12, 2010 8:52 am
by ethanph
I see, thanks for your answers. So, to support such large data processing actually requires a very different design? It looks like all existing GPU-based libraries cannot handle the data that cannot fit into the GPU memory.