higherorderco/bend

Bend is a high-level programming language designed to automatically run your code across thousands of parallel processors, including GPUs, without requiring you to manually manage threads, locks, or any of the usual parallel programming complexity. Think of it like writing regular Python or Haskell-style code that silently scales across over 10,000 concurrent threads on its own. The way it works is surprisingly straightforward: if your algorithm can be broken into independent tasks (like splitting a calculation into two halves that don't depend on each other), Bend detects that and runs those tasks simultaneously. You write the logic, Bend figures out the parallelism. It's powered by an underlying runtime called HVM2, which handles the parallel execution using a model called Interaction Combinators. You'd reach for Bend when you have computationally heavy problems, sorting, simulation, tree traversals, number crunching, and want to squeeze the most out of modern hardware without rewriting your code in CUDA or dealing with multi-threading bugs. It's especially useful for divide-and-conquer style algorithms that naturally split into independent subtasks. The tech stack is built in Rust, runs via C and CUDA runtimes, and currently supports NVIDIA GPUs only. Windows users need WSL2. Single-core performance is still being improved, but multi-core and GPU speedups are already demonstrated with real benchmarks.