You can absolutely train a nontrivial task model on a gaming gpu (image classifier, sound classifier, text model with a very structured input and output, etc). You can also post-train addon layers on top of existing open-weight models (LORA).