AMD wants to take on NVIDIA's CUDA monopoly with ROCm, its open-source GPU computing platform. The pitch is simple: an alternative that doesn't lock you into a single vendor's compiler stack. For security-focused teams who can't trust proprietary binaries, that matters. But the reality of using ROCm tells a different story.
Nightmare. That's how developers describe porting projects like TheRock to ROCm. Over 30 dependencies. Custom LLVM builds. Fighting with toolchains like musl and mimalloc just to get deterministic builds working. A Hacker News commenter working on security workloads said they wrestled with every layer of the stack. That's the gap between "open source" and "actually usable."
AMD has been here before. The Boltzmann Initiative in the mid-2010s tried to unify CPU and GPU programming through HIP, a tool for porting CUDA code. It didn't take off, partly due to fragmented tooling. ROCm inherited HIP but shifted to an "upstream-first" philosophy, leaning on LLVM to avoid maintaining proprietary forks. The goal was avoiding the technical debt that plagued earlier efforts. Adoption remains slow regardless.
The bigger problem is hardware support windows. ROCm support for consumer devices reportedly lasts 3 to 5 years. CUDA support stretches far longer. For teams building infrastructure that needs to run for a decade, that's a dealbreaker. For those avoiding long support cycles, local options like AMD's Lemonade provide an alternative.
Some community discussion floated the idea of AI agents helping close the code parity gap between ROCm and CUDA. Perhaps tools like Modo could provide the structure needed to manage complex toolchains. But until AMD commits to longer support cycles and the tooling gets easier, CUDA's grip on AI compute isn't loosening.