Why does GPU development still feel so infrastructure-heavy?
It’s interesting how much progress we’ve made in software development overall, yet GPU-based workflows still feel very infrastructure-heavy compared to everything else. Instead of focusing purely on code and experiments, a lot of time still goes into managing environments, configuring systems, and handling setup details. Even simple experiments often require multiple steps before anything actually runs. I’ve been exploring more minimal approaches that try to reduce this overhead and make GPU usage feel more like a natural extension of development rather than a separate system. One example I came is swmgpu which focuses on simplifying the workflow into a more direct, CLI-driven experience. Do you think this is something that will naturally improve over time, or is GPU computing always going to carry this level of complexity?