GPU COARE 3.6 kernel — 1229x speedup at 1M ocean grid points

[consigcody94]

Summary

The COARE 3.6 bulk air-sea flux algorithm runs on GPU with one thread per ocean grid point. The 10-iteration Monin-Obukhov stability loop is compute-intensive (exp, log, sqrt, pow, atan, cbrt per iteration), making this well-suited for GPU.

Benchmark (RTX 3060)

Grid Points	CPU	GPU	Speedup	Max Rel Error (tau/hsb/hlb)
10K	18 ms	0.028 ms	643x	1.4e-06 / 2.3e-06 / 8.3e-04
100K	171 ms	0.171 ms	999x	1.8e-06 / 4.6e-06 / 1.0e-03
500K	854 ms	0.712 ms	1200x	3.0e-06 / 4.6e-06 / 2.1e-03
1M	1723 ms	1.40 ms	1229x	5.4e-06 / 4.6e-06 / 2.1e-03

Zero NaN across all configurations. The latent heat flux has slightly higher relative error (2.1e-03) due to GPU fast exp (__expf) in the saturation vapor pressure calculation. Momentum and sensible heat fluxes are accurate to < 6e-06.

Implementation

The GPU kernel implements the same iterative algorithm as module_coare36_model.f90:

• Saturation vapor pressure (Tetens-Magnus)
• Monin-Obukhov stability functions (psiu_26, psit_26)
• Charnock roughness length
• Transfer coefficients with stability correction
• Cool-skin terms (simplified in this benchmark)

Each grid point runs independently in registers. The 10 iterations per point give high arithmetic intensity (~200 FLOPs/point/iteration), which is why GPU utilization is so high.

Code

https://github.com/consigcody94/noaa-gpu-kernels/tree/main/ocean/coare

[consigcody94]

AI Attribution: This work was developed using Claude Code as an AI-assisted development tool. Code generation and benchmark automation were performed by Claude under my direction. I designed the research methodology, selected mathematical approaches, and validated all results on physical hardware (RTX 3060). All commits carry a Co-Authored-By: Claude tag. When results failed validation (e.g., tensor compression on rte-rrtmgp#394), they were publicly retracted.