In the last example, we saw that we can use the (overloaded) Python multiplication operator (* ) to multiply each element in a gpuarray object by a scalar value (here it was 2); note that a pointwise operation is intrinsically parallelizable, and so when we use this operation on a gpuarray object PyCUDA is able to offload each multiplication operation onto a single thread, rather than computing each multiplication in serial, one after the other (in fairness, some versions of NumPy can use the advanced SSE instructions found in modern x86 chips for these computations, so in some cases the performance will be comparable to a GPU). To be clear: these pointwise operations performed on the GPU are in parallel since the computation of one element is not dependent on the computation of any other element. 

To get a feel for how the operators work, I would suggest that the reader load up IPython and create a few gpuarray objects on the GPU, and then play around with these operations for a few minutes to see that these operators do work similarly to arrays in NumPy. Here is some inspiration:

Now, we can see that gpuarray objects act predictably and are in accordance with how NumPy arrays act. (Notice that we will have to pull the output off the GPU with the get function!) Let's now do some comparison between CPU and GPU computation time to see if and when there is any advantage to doing these operations on the GPU.