Performance

Runtime depends on the measures, sample count, dimensions, neighborhood size, memory budget, and device. Measure both the first evaluation and subsequent evaluations if you plan to reuse a runner.

Measure your workload

Using the arrays from the quickstart:

from time import perf_counter
from statistics import median
from zadu import ZADU

specs = [{"id": "tnc", "params": {"k": 10}}, {"id": "stress"}]
start = perf_counter()
runner = ZADU(specs, original)
runner.measure(projection)
first = perf_counter() - start

times = []
for _ in range(5):
    start = perf_counter()
    runner.measure(projection)
    times.append(perf_counter() - start)
print({"construction_and_first_seconds": first, "median_reused_seconds": median(times)})

This measures one process. Framework imports and initialization may already have happened if you ran other code first. Use a fresh process when measuring startup or process peak RSS. Compare score differences alongside timings when changing dtype or backend.

Published measurements

The 0.5.1 report records measurements from 2026-08-23 at revision 94e71a9, compared with v0.5.0 and v0.1.1 under one dependency environment. These are historical results, not measurements of the current checkout. Later distance kernels and memory plans can change runtime.

The repository also contains selected-rank measurements from revision 34730bb4 on an Apple M4. The file records its environment, budget, score deltas, and timing ranges. See the benchmark instructions for reproduction commands.