Array programming seems adaptable to any form of parallel processing: SIMD instructions, multi-core, or GPUs. I think it's because, in addition to their inheren

Array programming seems adaptable to any form of parallel processing: SIMD instructions, multi-core, or GPUs. I think it's because, in addition to their inherent parallelism, primitives are so simple that they make very few demands on an implementation. GPU code can be a lot more flexible than array programming because the processors have a bit of independence, so array primitives are not necessarily the best way to use a GPU, but they're pretty good and the limitations might make it easier for a human to think about.