The Futhark Programming Language (futhark-lang.org)

Futhark is a compact, statically typed, purely functional array language in the ML family that compiles ahead-of-time into high-performance parallel code for GPUs (CUDA/OpenCL) and multi-threaded CPUs. It’s designed for writing data-parallel numerical kernels more succinctly and safely than hand-coded CUDA/OpenCL, making it a practical tool for accelerating compute-heavy parts of applications rather than replacing general-purpose languages. The project is research-driven but production-ready enough to compile nontrivial programs and integrates easily: outputs include C or a Python module (via PyOpenCL), so Futhark kernels can be called from existing codebases. For the AI/ML community the key benefits are concise expression of array-parallel algorithms plus features directly useful for ML workloads: nested data-parallelism, automatic differentiation, and a uniqueness type system that enables in-place array updates without breaking purity—allowing both high performance and easier reasoning about code. The compiler is heavily optimizing, and the language deliberately omits some general-purpose features to enable aggressive parallel transformations. In short, Futhark offers a lightweight, learnable way to write GPU/CPU kernels with AD and safe in-place updates, ideal for accelerating training or inference hotspots without the complexity of manual kernel programming.