Skip to content

Benchmarks

Current numbers for what the README and the docs quote, measured on one machine. Regenerate them after a change that moves them (the tables below are llevant-bench report's Markdown output):

cargo build --profile dist -p llevant-bench
python tools/synthetic/gen_models.py untracked/synthetic
target/dist/llevant-bench report untracked/big_example_asembled.mcnp untracked/synthetic/*.i \
  --json report.json --markdown report.md

The reference model (big_example_asembled.mcnp, built by gitronics, not in git) is a fusion reactor model:

  • 378 MB of input, 425 023 cells, 592 588 surfaces;
  • 187 universes nested up to 4 levels deep, and six LAT=1 cells;
  • mostly general planes and GQ cylinders converted from CAD.

The synthetic models (tools/synthetic/gen_models.py) are other shapes:

  • tokamak: toroidal sectors with slab fillers;
  • core: a PWR-like lattice of lattices;
  • sparse: 1 000 spheres in a 5 km world;
  • bricks: axis-aligned boxes;
  • tori: torus shells.

What report measures:

  • Workload. Points and 100 cm lines are drawn uniformly inside each model's contents (Model::extent) with a fixed seed. Lines go in random directions. "One thread" runs a tenth of the batch on a pool of one thread. Rates are the best of three calls.
  • Staged compile and recompile. The compile up to the BVHs, without octrees (Model::compile_staged: what the viewer shows first). Then the first surface of the input's last cell moved by 1e-3 cm and the model recompiled against the compiled one (Model::recompile), staged and then complete, with the number of universes it built again (decision 0010).
  • Drag. The same surface dragged live (decision 0011):
  • Draft::begin;
  • the 95th percentile of 41 steps moving its last coefficient by up to 1 cm either way, with the cells it moves;
  • with a GPU, the best of three writes of a step into the upload (GpuModel::patch, waited for).
  • Slice. 1000 rows across the middle of the contents, at z = mid.
  • CPU render. 800 × 600, solid cells, from the direction (0.6, 0.7, -0.4).
  • Check. Model::volumes over the contents, on lines about 1/400 of their longest side apart, rounded to 1, 2 or 5.
  • GPU. The upload from the model's lean export, and the best of four 1920 × 1080 frames of the render's view, solid cells, traced back to back after a 5 s warm-up (frames read back one at a time let the driver drop the clock to a third, decision 0015).

Engine, CLI and GPU

Measured 2026-11-03 at commit cbc5379, dist profile, on AMD64 Family 25 Model 80 Stepping 0, AuthenticAMD (16 threads, windows), GPU NVIDIA GeForce RTX 3060 Laptop GPU.

big_example_asembled.mcnp bricks.i core.i sparse.i tokamak-coarse.i tokamak.i tori.i
Cells / compiled surfaces / levels 425023 / 592637 / 4 32001 / 103 / 1 16 / 23 / 5 1002 / 1001 / 1 7805 / 337 / 2 6293 / 308 / 2 470 / 50 / 1
Compile (no cache) 37.81 s 0.06 s 0.00 s 0.15 s 0.05 s 0.02 s 0.01 s
Compile to a staged model (no octrees) 9.55 s 0.05 s 0.00 s 0.00 s 0.02 s 0.02 s 0.00 s
Recompile, one surface moved: staged / complete (universes rebuilt) 2.44 s / 2.64 s (1) 0.05 s / 0.05 s (1) 0.00 s / 0.00 s (1) 0.00 s / 0.18 s (1) 0.01 s / 0.01 s (1) 0.01 s / 0.01 s (1) 0.00 s / 0.01 s (1)
That surface dragged: start / step, 95th percentile (cells moved) 90 ms / 0.3 ms (16) — — — 1 ms / 0.0 ms (2) 1 ms / 0.0 ms (2) —
Read the compiled model (cache) 1.41 s (1462 MB) 0.00 s (6 MB) 0.00 s (0 MB) 0.00 s (0 MB) 0.00 s (4 MB) 0.00 s (1 MB) 0.00 s (1 MB)
Points, all threads / one 7.02e6 / 6.18e5 per s 1.12e7 / 1.26e6 per s 1.96e7 / 2.44e6 per s 5.91e7 / 4.71e6 per s 1.20e7 / 1.37e6 per s 1.04e7 / 1.05e6 per s 1.01e7 / 1.11e6 per s
100 cm lines, all threads / one 4.12e5 / 4.22e4 per s 8.96e5 / 1.06e5 per s 1.96e5 / 3.17e4 per s 1.51e6 / 1.84e5 per s 7.05e5 / 8.45e4 per s 1.40e6 / 1.85e5 per s 6.84e5 / 7.42e4 per s
Slice, 1000 rows 93 ms 4 ms 42 ms 1 ms 16 ms 3 ms 22 ms
CPU render, 800 × 600 2027 ms 447 ms 188 ms 934 ms 1324 ms 473 ms 1025 ms
Fingerprints 0.43 s 0.01 s 0.00 s 0.00 s 0.00 s 0.00 s 0.00 s
Gap and overlap check 91.2 s (h = 5.0 cm) 1.0 s (h = 1.0 cm) 14.2 s (h = 1.0 cm) 3.5 s (h = 0.2 cm) 26.6 s (h = 2.0 cm) 9.9 s (h = 2.0 cm) 7.4 s (h = 2.0 cm)
GPU upload 1.4 s (1512 MB) 8.2 s (5 MB) 0.1 s (0 MB) 7.7 s (1 MB) 0.0 s (5 MB) 0.0 s (1 MB) 7.8 s (1 MB)
GPU frame, 1920 × 1080 280 ms 13 ms 9 ms 16 ms 10 ms 11 ms 10 ms
GPU: a drag's step written into the upload 7.4 ms — — — 0.5 ms 0.4 ms —

Against the previous table (a719ae9), the drag rows are new (decision 0011); the rest is unchanged within run-to-run noise. The 8 s uploads of three small models are most likely the driver compiling their shaders again on a first run (see below); the upload now also reserves 256 KB of slack. The surface measured on the reference model moves 16 cells; llevant-bench drag measures the others (steps of 0.5–13 ms at the 95th percentile, 36 ms where a roi cascade bounds 6 330 cells again). bricks.i, core.i, sparse.i and tori.i cannot be dragged on that surface: it bounds the problem's IMP:N=0 cell.

The table at a719ae9, against the previous one (f118416, v0.1.0 before the release commit), the staged compile and the recompile are new (decision 0010); the rest is unchanged within run to run noise (a back-to-back A/B of v0.1.0 and this commit, release profile, reference model: points 6.86 / 6.65·10⁶ per s, lines 3.95 / 3.99·10⁵ per s, check 94.4 / 94.2 s). The saved model is 1462 MB: v0.1.0 saves 1474 MB (the 1577 MB of the previous table predates it), and node-box ids local to each cell, small numbers, pack a little tighter.

Point and line rates are the best of three calls, as the other timings. The GPU upload is erratic on a first run: the driver compiles the shader for each new depth of programs and BVHs (8–9 s once, then cached by the driver; staged and complete models share it), and one report took 28 s for the reference model.

Viewer

Reference model, solid cells, default window, GPU backend, reprojection on. Measured with:

  • llevant-viewer <model> --orbit 10 --screenshot out.png for the orbit;
  • llevant-viewer <model> --storm for the edit storm (the UI thread's own work per frame).
Orbit, 10 s: frame time median / 95th percentile 16.1 ms / 21–22 ms (display rate 60 Hz)
Previews while orbiting ("While moving" at its default, 66 ms) ~11 per second, mean scale 0.84–0.86 (at 33 ms: ~20, 0.45–0.50)
A still view refined to full resolution 0.30–0.38 s (one job, blocks shown as they land)
Edit storm: worst frame of the UI thread's own work 18.6–19.4 ms (631 frames, median 0.35 ms)
The same orbit on the CPU backend (16 threads): 95th percentile / refinement 32.5 ms / 4.9 s

Measured 2026-11-14 at 70182a6, release profile, against 13f08c2 the same day: previews 12 a second at scale 0.07, refinement 0.65 s, orbit 19.6 ms at the 95th percentile; CPU backend 32.3 ms / 6.35 s (decision 0013). Frame times include the display's vsync; the window waits a little longer behind a GPU that is busy more of the time. The storm's frames over 8 ms are the label rules over nearly every cell (select $ e, isolate $ e, 13–19 ms on the UI thread), added after the previous table.

History

Earlier numbers, and what was tried and did not help, are in the decision records (decisions/), and in the git history of docs/measurements.md. The 96-thread workstation numbers of October 2026 are kept there too: about 7·10⁶ points/s and 2.5·10⁵ lines/s on the reference model, and a 5 cm gap check of its 36 × 37 × 35 m building in 45 s.