notwitcheer/sm120-field-guide

The sm_120 field guide is a written collection of real world problems and fixes for running modern AI software on consumer NVIDIA RTX 50 series graphics cards, specifically the RTX 5090. Most AI tools and tutorials are built and tested on expensive datacenter grade cards, and consumer cards from the same generation look similar on paper but fail in confusing ways: pieces of software silently ship without the right low level support for these cards, and speed claims from other guides often assume a much more expensive card than most people actually own. Every entry in the guide documents something the author personally ran into on their own machine, including the exact error message, why it happens, the fix that worked, and the date and software versions it was tested with. Nothing is included secondhand or without having actually been run. The guide is organized as 80 short entries spread across seven topic areas: getting the core AI software stack and its accelerated math libraries working, serving AI models for other programs to use, training smaller helper models on a single card, shrinking large models to fit in limited graphics memory, building software from its source code on this hardware, and honestly measuring whether a performance claim actually holds up. The guide opens with a five minute test anyone can run before trying to set up a larger AI project, to quickly check whether a piece of software's required version of the underlying deep learning library will work on this card at all, since some older versions simply do not support it and cannot be patched around. The project describes itself as reproducible and honest, meaning every entry is dated and versioned so readers know how much to trust it as tools change over time, and every technique described only counts if it works on a single card the author actually owns, with no reliance on cloud computing.