Engineering case studies¶
Porting and performance work — moving code across GPU vendors, fixing GPU kernels, building software from source — done end-to-end through Trinity's agentic workflow, with the debugging story and validated numbers, not just a final checkmark.
-
BEM_2D: 1,041x speedup porting a 2008 PhD code to CUDA
A 1970s special-function routine gets a from-scratch GPU rewrite, verified bit-identical to the CPU reference at every stage.
-
BEM_3D: from a 40GB memory wall to a 2TB-equivalent problem
A matrix-free GMRES solver, written from scratch, reaches mesh resolutions no dense method could ever store.
-
GeoSeries: a 7,638x speedup from four stacked performance bugs
Systematic profiling — not guesswork — finds four independent bottlenecks piled on top of each other.
-
Porting a multiphysics framework across GPU vendors — and a 209x kernel speedup
Two weeks, three obscure vendor bugs, and a bypass of the framework's own bottleneck loop.
-
Porting an AlphaFold2 reimplementation to Intel GPUs
A fix to the obvious call site wasn't enough — the same CUDA kernel was hardcoded a second time, deeper in the model.
-
Adding GPU support to GROMACS on Frontier — a work in progress
Both preferred GPU backends turned out to be unavailable; the honest current result is a real production slowdown, not a win yet.
-
Porting a spatiotemporal model from AMD to NVIDIA — in under an hour
No model code changed — just the scheduler and hardware-detection glue around it, fixed across 10 job submissions.
-
Getting a lightweight protein-folding model running on Aurora
A modest deployment story: a silent CPU-only install trap, a network config fix, and a clean inference smoke test.
-
Building PyTorch from source for Intel GPUs
A dead-end build arrangement abandoned for a newer one, resolved autonomously across 23 iterations.
-
Building a GPU compiler stack — and confirming a known bug still exists
Targeted test kernels isolate exactly which patterns trigger a known closed-source compiler bug, and which don't.
-
Building PyTorch where the usual submission paths didn't work
Both standard job-submission routes failed outright; a dependency trap silently mismatched GPU toolkit versions.
-
Building a cosmology code with no direct system access
Four build failures diagnosed and fixed purely by reading logs through a file-transfer API — no terminal access at all.
-
Designing a physics experiment from a single sentence
Given only "show PV=NRT," Trinity designed the densities, the runs, and the analysis, and got the right physical trend.