Benchmarks

Two kinds of measurement, kept apart. The FastHenry comparison is operator-run, by hand, on one machine, and CI never runs FastHenry. The criterion suite and the scaling run are scripted. Everything on this page is rendered by site/build.py from docs/benchmarks.md; no number is typed into the page.

Machines

FastHenry comparison and the criterion suite: Apple M3 Ultra, 28 cores, host under background load (~2× slowdown typical; both engines equally affected).
Scaling run: AWS EC2, 8 vCPU (Xeon-class), 30 GB RAM, Linux. One thread: the measuring process was CPU-budgeted to a single core by its cgroup, so rayon ran one worker and *both* arms are single-threaded, like the "fasterhenry (dense, 1 thread)" column above. Treat the crossover as a per-thread figure; a many-core run moves both arms, not obviously by the same factor.
SIMD: the partial-inductance kernels use 4-wide f64 lanes (wide::f64x4), as in neumann.rs.

Speed against FastHenry

deck	filaments	FastHenry (multipole, 1 thread)	fasterhenry (dense, 28 threads)	fasterhenry (dense, 1 thread)
spiral	48	31 ms	10 ms	—
plane_trace	~250	21 ms	13 ms	—
serpentine	1 000	69 ms	73 ms	—
serpentine	5 000	4.52 s	2.06 s	31.6 s
serpentine	19 600	103.7 s	83.4 s	1 475 s
serpentine	80 000	did not finish (> 30 min)	not attempted (dense)	—

Operator-run; method and caveats are in docs/benchmarks.md. The decks and the build of FastHenry are described in its Setup section.

Criterion bench suite

Not yet regenerated by CI — run .github/workflows/bench.yml (workflow_dispatch) or cargo bench -p fasterhenry --bench assembly --bench solve && python3 tools/bench_table.py locally to fill this in.

Regenerated by .github/workflows/bench.yml via tools/bench_table.py, which parses criterion’s own JSON output. Locally: cargo bench -p fasterhenry --bench assembly --bench solve && python3 tools/bench_table.py.

Scaling: dense vs pFFT + GMRES

Wall clock of assemble + one-frequency sweep, single thread. The dense arm stops where its model memory no longer fits; the table below is the data.

filaments	loops	dense s	dense model	dense peak	iterative s	iterative model	iterative peak	GMRES its
960	720	0.84	24 MB	61 MB	1.71	53 MB	55 MB	5
2 400	1 800	4.15	150 MB	227 MB	4.69	129 MB	130 MB	5
4 760	3 570	20.9	589 MB	635 MB	8.95	240 MB	223 MB	5
9 800	7 350	156	2 497 MB	2 198 MB	19.4	487 MB	416 MB	5
29 928	22 446	not attempted	23 GB	—	60.9	1 474 MB	1 204 MB	5
99 224	74 418	not attempted	256 GB	—	207	4 803 MB	3 831 MB	5

Reproduce: cargo bench -p fasterhenry --bench scaling -- __report_only__. CI does not run this suite; it gates the headline with iterative_path_holds_its_scaling_headline in perf_smoke.rs.