25 lines (19 loc) · 872 Bytes

Experiment 08: std430 vs std140 vs Packed

1. Focus

Separate API-driven layout padding from useful logical payload in storage buffers.

2. Question

How much cost comes from std140 and std430 padding compared with a tightly packed representation?

3. Variants

packed
std430
std140

4. Method

Keep the same logical fields and arithmetic while changing only the storage layout.
Measure both runtime and physical bytes per record so padding cost stays visible in the analysis.

5. Outputs

Median GPU time by layout.
Physical storage bytes per record.
Useful-payload GB/s and relative slowdown vs packed.

6. Interpretation

Padded layouts can simplify interoperability, but they should not be mistaken for free memory traffic.
The important comparison is useful work delivered per unit time, not just raw storage stride.