Capacity and movement
GPU memory holds more than model parameters. Runtime buffers, activations, attention state and training state can change the working set. Precision, context length and batching influence both capacity and throughput. State the complete software configuration before comparing memory figures.
A topology has boundaries
Memory attached to separate devices is not automatically one uniform addressable pool at the same performance. Parallelism can distribute a model but adds coordination and communication. Ask which links connect the devices and how the workload traverses them.
Evidence to request
Obtain a representative peak-memory trace, an out-of-memory recovery plan and a scaling test. Retain the model version, framework and allocation method. The readiness checklist records evidence status rather than estimating model fit from an incomplete parameter count.