Jetson shared memory is nearly full: will adding swap solve an inference workload?

Inventory model weights, runtime workspaces, input and output tensors, decoded frames and application copies. Check which allocations must be device-accessible or otherwise cannot benefit from ordinary swapping. A board's shared-memory architecture does not make every allocation interchangeable.

Measure peak use during startup, engine creation and steady operation separately. For example, four 1920 × 1080 RGB frames occupy about 24.9 MB before alignment and additional processing buffers. Several pipeline stages and queued frames can multiply that footprint.

Reduce unnecessary copies and bound queues first where possible, then evaluate model or resolution changes against accuracy requirements. Swap may help some CPU memory pressure but can introduce storage traffic and latency; it is not a substitute for a measured GPU and pipeline memory budget.