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/