WG211/M23Scholz

Stream it if you can - not just when your GPU runs out of memory (Sven-Bodo Scholz)

This talk presents our work on enabling out-of-core computations on GPUs, i.e., computations on arrays that are bigger than what can be held within the device memory of a given GPU. We propose code generation schemes for streaming map-like operations as well as stencil computations across GPUs. We also provide a performance evaluation of these two schemes. As it turns out, computations of all sizes benefit from the streaming approach.