I actually don't think that would be easier to program. The current thread abstraction won't be efficient with 1024 cores. OpenCL is much better suited for that. But then you might as well run it on a GPU.