GPU instances work best for applications with massive parallelism such as workloads using thousands of threads. Graphics processing is an example with huge computational requirements, where each of the tasks is relatively small, the set of operations performed form a pipeline, and the throughput of this pipeline is more important than the latency of the individual operations. To be able build applications that exploit this level of parallelism, one needs GPU device specific knowledge by understanding how to program against various graphics APIs (DirectX, OpenGL) or GPU compute programming models (CUDA, OpenCL).
