GPU Architecture
Now we work with a very different architecture. To deal with this complexity, we write a program that runs on the CPU, which writes and runs a program on the GPU. This both addresses the architectural differences, and because the GPU program is compiled at runtime, cross platform (ATI, Nvidia, etc) issues are also dealt with.
The difference is size for the arrows to and from the GPU is not a conicidence. It is far faster to load data onto the GPU than to read data back.