Performance: Spatial

Benchmarks (Intel i7‑12700K, 32 GB, Python 3.12, metainformant 0.8.0)

All tests on a single Visium section (~5 200 spots × 36 500 genes, sparse matrix ~4 % fill).

Operation	Median	95th pct	Peak RAM	Notes
`load_visium()`	2.1 s	2.4 s	180 MB	includes Space Ranger JSON parse
`neighbors(method='knn')`	180 ms	210 ms	50 MB	sklearn KDTree, expression PCA‑50
`neighbors(method='radius', r=100)`	420 ms	480 ms	55 MB	radius search
Spatial Leiden (res=0.8)	320 ms	380 ms	80 MB	igraph partition
Moran's I (single gene, 999 perm)	350 ms	400 ms	10 MB	permutation loop
Local Moran (single gene, 99 perm)	200 ms	240 ms	10 MB
Stereoscope (5 k spots × 20 types, 400 epoch)	6.2 min	—	450 MB (CPU)	90 s on CUDA A100
Tangram (same)	1.4 min	—	380 MB	CUDA 45 s
`spatial_scatter()` (matplotlib)	120 ms	140 ms	20 MB	figure render only

Spots	Graph (knn)	Moran (999 perm)	Stereoscope (CPU)
5 k	0.18 s	0.35 s	6 min
20 k	0.75 s	1.4 s	25 min
100 k	4.2 s	7.8 s	2 h +

Memory scales linearly with N_spots for dense graphs (~N² for radius method).

Graph sparsification — reduce n_neighbors to 10 for very large N; speeds clustering ×1.5 with minor accuracy loss.
Permutation count — for discovery use n_permutations=99; for publication use 999. Analytic Moran's I (no permutation) is available but less accurate for small N.
Deconvolution epochs — early stopping monitors ELBO; typical convergence plateau at 300–400 epochs. Reduce to 200 for prototyping.
Device — CUDA accelerates Stereoscope×8; Tangram×4 (memory bound).

Loader I/O — raw matrix HDF5 read dominates initial load; place data on NVMe.
Neighbour graph — radius method O(N²); prefer KNN for N > 20 k.
Deconvolution — GPU memory can OOM with >100 cell types; lower batch_size.
Permutation tests — auto‑parallelise across genes via joblib if you have many cores: morans_i(..., n_jobs=8).

python -m cProfile -o spatial.prof myscript.py
snakeviz spatial.prof

Look for accumulated time in: