Vx – One Language, Every Chip

Writing code that runs on GPUs or other accelerators becomes significantly less error-prone.

Vx, a new systems programming language, has been released, aiming to simplify heterogeneous computing across various hardware accelerators like CPUs, GPUs, and NPUs. Its core innovation is integrating memory locations and accelerator specifics into the type system. This approach aims to prevent common runtime errors such as segfaults or silent data corruption by catching them at compile time. The language uses a data-oriented, parallel frontend and supports compilation to MLIR, which then interfaces with LLVM for code generation across different architectures including x86-64, ARM64, NVIDIA GPUs, and Apple's hardware. Vx emphasizes explicit data locality and transfers, ensuring that data movement is provable from the source code. It provides machine files for specific hardware, allowing the compiler to check resource constraints like memory capacity and bandwidth against the target hardware's specifications before compilation.

Vx's type system, which incorporates memory location and accelerator topology, represents a significant departure from how heterogeneous computing is typically handled. By front-loading memory-related bugs into compile-time checks, Vx aims to drastically reduce the debugging burden for developers working with diverse hardware. This strict compile-time enforcement of data locality and transfer operations, even on unified memory architectures, offers a new paradigm for building reliable and performant systems on specialized silicon.

The language's design, prioritizing ahead-of-time compilation and static regioning, positions it for applications where correctness and speed are paramount. While this approach might introduce a steeper learning curve for developers accustomed to the dynamic nature of frameworks like PyTorch, it promises to deliver robust solutions for complex computational tasks across a wide range of hardware. Vx's extensibility through MLIR pass plugins also suggests a flexible ecosystem for supporting future accelerators.