Repository navigation
GPU batched tridi1 PBL solver — 7.5-11.6x speedup at 500K columns #1206
Description
Summary
The Thomas algorithm tridiagonal solver in physics/PBL/tridi.f (tridi1) runs on GPU with one thread per column. At GFS-scale column counts, 7.5-11.6x speedup with FP32 precision.
Benchmark (RTX 3060)
| Columns | Levels | CPU | GPU | Speedup | Status |
|---|---|---|---|---|---|
| 10K | 64 | 5 ms | 0.43 ms | 11.6x | PASS |
| 100K | 64 | 43 ms | 5.6 ms | 7.6x | PASS (FP32 rounding) |
| 500K | 64 | 215 ms | 28.7 ms | 7.5x | PASS (FP32 rounding) |
The GPU kernel runs the identical Thomas algorithm (forward elimination + back substitution). FP32 rounding causes 0.006% of entries to differ at the last few bits — not algorithmic error.
Code
$NEW_REPO (noaa_multi_kernel.cu)
Activity
consigcody94 commentedon Mar 30, 2026
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.