The NVVM library compiles NVVM IR (a subset of LLVM IR) into PTX code which can then be executed on NVIDIA GPUs. In contrast to the standard NVPTX target built in to the LLVM toolchain, NVVM includes a set of proprietary optimisations which are otherwise only available by compiling CUDA code with the nvcc compiler. The resulting PTX code can be loaded onto the GPU and executed using the cuda package: https://hackage.haskell.org/package/cuda The NVVM library is a compiler component available a part of the CUDA toolkit: https://developer.nvidia.com/cuda-toolkit See the travis-ci.org build matrix for tested CUDA library versions.