plan.md

Experiment 15: Bandwidth Saturation Sweep

1. Focus

• Find the workload size where simple contiguous kernels reach a sustained bandwidth plateau.

2. Question

• How large must a memory workload be before fixed dispatch overhead is mostly amortized?

3. Variants

• read_only
• write_only
• read_write_copy
• dense size sweep from small to large buffers

4. Method

• Reuse the simple contiguous memory modes from the baseline experiment with a denser size sweep.
• Keep timing limited to GPU dispatch so the plateau reflects kernel execution rather than staging cost.

5. Outputs

• GB/s vs size for each memory mode.
• Plateau onset estimate.
• Sustained-region median GB/s.

6. Interpretation

• The plateau is practical measured bandwidth for the tested device, not proof of theoretical peak.
• Later experiments should prefer sizes in this stable region when the goal is to study steady-state behavior.