LLVM actually has a native PTX backend, NVPTX, but since the Nvidia ISA is proprietary we use the CUDA driver API to generate the binary for the target GPU arch. (see here in MapD where it generates the PTX from the LLVM IR: https://github.com/omnisci/mapd-core/blob/ae88c486d1d790db54..., and here for where it calls the driver API to generate the native code: https://github.com/omnisci/mapd-core/blob/568e77d9c9706049ee...)