plan.md

Experiment 29: Shared Memory Bank Conflict Study

1. Focus

• Measure how shared-memory stride changes throughput because of bank conflicts.

2. Question

• How severe is the slowdown from conflict-heavy strides, and how much does padding recover?

3. Variants

• stride_1
• stride_2
• stride_4
• stride_8
• stride_16
• stride_32
• padded_fix

4. Method

• Load one workgroup tile into shared memory and reread it with a configurable stride.
• Keep global-memory payload fixed so the difference is dominated by on-chip behavior.

5. Outputs

• Median GPU time by stride.
• Slowdown relative to stride_1.
• Padding recovery relative to stride_32.

6. Interpretation

• Shared memory is not automatically fast; bank conflicts can erase the intended gain.
• Padding should be justified by a measured recovery, not by habit.