My understanding is that using pinned memory can make repeated transfers to the GPU faster. I don't think we are at all memory bound, but it still might be useful to use pinned memory for the proposals and responses? I'm not sure it makes sense for the patch data since that is a single transfer per CPU process in the current architecture.
My understanding is that using pinned memory can make repeated transfers to the GPU faster. I don't think we are at all memory bound, but it still might be useful to use pinned memory for the proposals and responses? I'm not sure it makes sense for the patch data since that is a single transfer per CPU process in the current architecture.