resources - Maximum instructions per CUDA kernel? Maximum operations per CUDA stream? -
is there maximum number of cuda operations may pending specific cuda stream? haven't seen such limit in documentation.
i interested in related figure of maximum number of instructions per cuda kernel.
there maximum number of cuda ptx instructions per kernel:
- 2 million gpus compute capability under 2.0 (i.e. before fermi microarchitecture)
- 512 million gpus compute capability 2.0 or higher (e.g. fermi, kepler, maxwell, ...)
this information can found in cuda c programming guide, "maximum number of instructions per kernel".
as streams, if kernels run on given stream respect limit, there no such stream instruction limit. @talonmies pointed out, streams host side queue of operations, have nothing loading code onto gpu.
Comments
Post a Comment