Work stealing makes it arguably much easier to program than CUDA or OpenCL.