Show HN: CuMetal – Run CUDA Programs on Apple Silicon via Metal

CuMetal, an experimental CUDA compatibility layer for Apple Silicon, was posted to Hacker News on August 25, 2026 by developer lulzx. The project, hosted on GitHub, aims to run a subset of CUDA programs on Apple Silicon GPUs by translating CUDA calls through Apple's Metal graphics and compute API.

Per the Hacker News submission, CuMetal is described as targeting "a useful subset of CUDA programs" rather than full CUDA compatibility — a meaningful scope limitation given the breadth of the CUDA ecosystem. The project is presented as experimental, and no production-readiness claims are made.

CUDA is NVIDIA's proprietary parallel computing platform and API, and programs written for it have historically required NVIDIA hardware to run. Apple Silicon Macs ship with Apple's own GPU architecture, which uses Metal as its native low-level compute API. The gap between the two ecosystems has been a persistent friction point for developers porting machine learning workloads, scientific computing tools, and other GPU-accelerated software to Apple hardware.

CuMetal sits in a space alongside other efforts to bridge CUDA and non-NVIDIA hardware, including AMD's HIP/ROCm platform and Intel's oneAPI CUDA compatibility tools. Apple itself has promoted Metal and its machine learning frameworks as native alternatives, but direct CUDA compatibility layers for Metal remain relatively rare and largely community-driven. The GitHub repository does not list prior funding, a company affiliation, or a team beyond the individual author.

The submission drew no comments on Hacker News at the time of posting. No roadmap, contributor guidelines, or further forward-looking statements were included in the available source material.