GPU’s, implementing common graphics tasks on specialized circuits, speed up display tasks involved in personal computing.
High data parallelism
GPU’s follow a dataflow architecture: Highly pipelined, parallel with many small cores - much more than CPU’s. Usually, these parallel cores are divided profitably and easily: Like one core for a bunch of pixels.
Earlier these processing elements were specialized for graphics, now these shaders are more programmable.
For general computing
With GPGPU, one would Disguise program as geometry computation. Now, can do such computation directly: eg: NVidia CUDA.