Back to the question about what type of parallel programming model CUDA fits under, this slide shows that in a way, it has aspects from all three models. Although it primarily seems to be a data parallel SPMD programming model, the communication between separate host and device memory is reminiscent of message passing, and threads in the same thread block have shared memory as well.

To add to jchen: it seems like we need to be clear that different layers of the CUDA abstraction uses a different memory model. And I would presume this is important to keep in mind as we get into implementation with CUDA