AMD ROCm

Raised as a usability problem by 14 people on Hacker News over 12 years (2015–2026), in 15 threads.

People raised many different problems with AMD ROCm, but no single problem has yet been raised by 3 people in 3 different months and threads.

Unreviewed issues were grouped by machine: Claude Haiku labelled each comment as a usability complaint, named what it's about and summed up the problem in a line, and a small model grouped similar problems. Nobody has checked them by hand. Expect mistakes.

The newest 12 complaints

Bugs in the compute stack go unfixed for years, and it barely works on Windows or consumer GPUs2026-06-11
Developer experience is horrible; driver crash bug reports not accepted2025-11-15
Setup needs googling error messages and random GitHub issues to find fixes2025-04-24
Borderline unusable for developing new model architectures, unreliable beyond running popular pretrained models2024-06-15
AMD GPU software stack is too hard to use for DNN work at scale2024-05-03
Confusing naming and poor docs; users can't tell what ROCm is or if their card works2024-03-19
Most AI tools assume CUDA, so getting them to run on ROCm takes manual effort2023-08-24
Getting the 7900XTX to compute is an obnoxious rabbit hole, unlike CUDA's fuss-free setup2023-07-26
Hard to get ROCm working; users must be engineers to make it run2023-06-06
Consumer Ryzen APUs are effectively unusable for compute because the software stack doesn't support them2023-03-01
AMD GPGPU works only on few distros and new hardware gets slow support2022-07-05
Heterogeneous compute power seems unusable outside Windows drivers2015-03-01