GPU stuff and ideally easier parallelism. The same binary could be executed on a GPU from any vendor, any CPU, anything that supports OpenCL, and could maybe even be extended to support FPGAs in the future.