All device operations (kernels and data transfers) in CUDA run in a stream. When no stream is specified, the default stream (also called the “null stream”) is used. The default stream is different from other streams because it is a synchronizing stream with respect to operations on the device: no operation in the default stream will begin until all previously issued operations in any stream on the device have completed, and an operation in the default stream must complete before any other operation (in any stream on the device) will begin.