Historical coagulation benchmark evidence
This is the 2026-07-10 machine-bounded decision record preserved from revision
7b064295037c3c9a137f5dd13c03f4a40da3f11b. Its timings are historical observations, not measurements
of the v0.3.0 tree. The mutable local artifact path may now contain a different
run; inspect its metadata before comparison. It is not resident/capture evidence.
The repository does not retain generated files under .artifacts/benchmarks/.
Optional benchmark runs may recreate that ignored output directory. The
documentation viewer uses the declared snapshot in
docs/data/gpu_benchmark_results.json; its recorded provenance is unchanged
and does not imply that it contains the July capture described below.
- Command evidence: run
pytest particula/gpu/tests/benchmark_test.py --benchmark -v -s. This remains opt-in benchmark evidence for local CUDA-capable environments, not part of the standard Warp CPU release-validation command set. The historical source-of-record artifact at.artifacts/benchmarks/gpu_benchmark_results.jsonpreserves the exact executed local command inbenchmark_metadata.command; the 2026-07-10 UTC capture came from a local coverage-wrapped invocation of that same benchmark target, but the plain user-facing reproduction command remains the one above. - Hardware/context: the same artifact records
warp_version=1.15.0,device.alias=cuda:0,device.name=NVIDIA GeForce RTX 5060, anddevice.total_memory_bytes=8082096128(about 8.08 GiB), withstarted_at=2026-07-10T01:59:34.215800+00:00andupdated_at=2026-07-10T02:00:43.918911+00:00. - Artifact output:
.artifacts/benchmarks/gpu_benchmark_results.json - Mixed-scale fixture note: the coagulation benchmark path uses a dedicated deterministic mixed NPF/droplet fixture aligned with the shipped coagulation rejection sampling baseline, while condensation benchmarks keep the generic helper.
Recorded one-thread-per-box tradeoff
Section titled “Recorded one-thread-per-box tradeoff”The recorded decision accepts the one-thread-per-box coagulation design with caveats: it is the current documented baseline for many independent boxes, especially when CUDA can deliver box-level parallel throughput, and for low-level direct-kernel workflows, but large single-box use remains bounded by the caution band below and does not imply a broader long-term scaling decision.
- Single-box caution band: treat the recorded
1x10kto1x50krange as the current caution band for large single-box workloads on the shipped one-thread-per-box kernel. The measured GPU time per step rises from0.6809sat1x10kto1.3689sat1x20kand3.3984sat1x50k, so this path is no longer a strong recommendation for large single-box work in that band. Smaller1x500to1x5kcases (0.0361s,0.1366s,0.3411s) remain a modest-throughput regime, not a decisive GPU win. - Multi-box effective region: independent-box parallelism remains
effective in the recorded many-box cases, where equivalent or larger
total-particle workloads stay near the smaller timings seen at lower box
counts:
10x5000.0359s,10x1k0.0699s,50x1k0.0714s,10x5k0.3437s,50x5k0.3528s,100x1k0.0766s, and10x10k0.6938s. Record this as the region where the current GPU path still benefits from many independent boxes keeping threads busy. - Machine-bounded caveat: this is measured evidence from the 2026-07-10 capture on the recorded Warp/device context above, not a universal cutoff for every GPU or every future kernel variant. Treat it as the current decision record for the shipped one-thread-per-box design, not as a final acceptance decision on the long-term scaling strategy.
The legacy viewer reads a locally recorded artifact. Resident benchmark qualification requires its separate schema and reviewed evidence.
Explore this topic: Performance and memory evidence