Skip to content

Latest commit

 

History

History
24 lines (18 loc) · 780 Bytes

File metadata and controls

24 lines (18 loc) · 780 Bytes

Experiment 06: AoS vs SoA

1. Focus

  • Measure how record layout changes memory efficiency for field-oriented GPU work.

2. Question

  • Does struct-of-arrays outperform array-of-structs for the same logical particle update?

3. Variants

  • aos
  • soa

4. Method

  • Run the same logical kernel over identical particle data in both layouts.
  • Keep the fields touched, arithmetic, and output contract fixed across both variants.

5. Outputs

  • Median GPU time by layout.
  • Useful-payload GB/s by layout.
  • Speedup of soa relative to aos.

6. Interpretation

  • Field-wise access usually favors SoA because threads read contiguous values from the same field.
  • Any AoS win would need to be explained by a different access shape or reduced address-generation cost.