i think the main idea is that by using MLIR[1] instead of the usual LLVM path it can target gpus directly

1. https://mlir.llvm.org/