Vx – One Language, Every Chip

Vx is a systems programming language designed to make heterogeneous computing - CPU, GPU, NPU and other accelerators - an explicit part of program semantics by encoding memory location and topology into the type system. Data pinned to different memories (host DRAM, HBM, NPU-local) are distinct types and crossing them requires an explicit transfer(), so illegal host dereferences of device pointers, capacity overcommit, missing transfer visibility, or transfers across unreachable topology paths are compile-time errors rather than runtime crashes. The compiler enforces linear ownership, region-aware borrowing, SMT-checked “seam contracts” for asynchronous visibility, autodiff safety across regions, and capacity admission against a declared machine description; machine files precisely record memory hierarchies, bandwidths and interconnects (with exact integer units) for real parts like H100, M4, MI300X, etc.

Compilation uses a data-oriented parallel frontend where every symbol is a flat 256-bit identifier and monomorphized units emit deterministic MLIR; the pipeline is designed to scale without lock contention and to produce bit-identical IR across parallel builds. Backends target native CPU (x86-64, ARM64), NVIDIA GPUs (NVVM→PTX→SASS), Apple AMX/ANE via CoreML dispatch, and distributed manifests; vendors extend via MLIR pass plugins. Vx favors ahead-of-time, statically regioned programs that must be correct and fast across diverse silicon, and is deliberately unsuited to highly dynamic, exploratory workflows that mutate shapes or control flow at runtime.