A data science team is preparing to fine-tune a large language model on a single workstation. During their first attempt, the training job crashes immediately with an out-of-memory error before completing even one iteration. The model weights, optimizer states, and a modest batch of training data collectively exceed what can be held on the accelerator. The team wants to understand which GPU characteristic is the primary constraint they have hit, so they can plan a fix. Which hardware attribute is the limiting factor in this situation?