GPU batched TOPMODEL — 31.3x speedup at 2.7M catchments, PASS

[consigcody94]

Summary

The TOPMODEL runoff generation kernel (exponential transmissivity + saturation deficit) can run on GPU with one thread per catchment. At the NWM's 2.7M catchments, this achieves 31.3x speedup with max relative error 5.87e-07 (FP32 precision).

Benchmark (RTX 3060)

Catchments	CPU	GPU	Speedup	Max rel error
100K	7 ms	0.40 ms	17.7x	4.76e-07
1M	69 ms	2.36 ms	29.2x	5.30e-07
2.7M (NWM)	184 ms	5.87 ms	31.3x	5.87e-07

Zero NaN, zero failures. The GPU uses CUDA __expf() for the transmissivity exponential which gives slightly different rounding than CPU expf() — both are within FP32 precision.

Code

https://github.com/consigcody94/parallel-prefix-rt/blob/master/benchmarks/cuda/owp_extended_kernels.cu

[consigcody94]

Code has been reorganized to: https://github.com/consigcody94/noaa-gpu-kernels (water-prediction/owp_extended_kernels.cu)

[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.