Benchmarks / Single-cell genomics
Python Scanpy methods Scanpy

Speed up Scanpy pca

Scanpy pca is one of the slower steps in many single-cell genomics workflows. AutoZyme ships a verified, drop-in patch that is up to 36.0× faster, returning output within a strict, verified tolerance with no change to how you call it.

Best speedup 36.0×
Median speedup 20.5×
Output equivalence Tolerance
Best runtime baseline 58.94 s optimized 1.63 s
Datasets 6
Pass rate 11/11

Benchmark charts

Switch benchmark platform; all charts update together
Platform
Speedup distribution
Each dot is one finalized dataset/thread run on Windows
log scale
pbmc200k_glaucomapbmc68kheart_adultsplitseq_rosenberggastrulation_pijuansa…tms_ss2
Thread sweep
Speedup across finalized thread counts on Windows
0×25×50×14full (32)pbmc200k_glaucoma · medium1 threads · 14.0× speedup1.50 min baseline → 6.43 s optimizedmemory 13 GB → 7.7 GBpbmc200k_glaucoma · medium4 threads · 24.7× speedup57.46 s baseline → 2.34 s optimizedmemory 13 GB → 7.7 GBpbmc200k_glaucoma · medium32 threads · 36.0× speedup58.94 s baseline → 1.63 s optimizedmemory 14 GB → 7.9 GBpbmc68k · small1 threads · 17.2× speedup35.89 s baseline → 2.10 s optimizedmemory 3.4 GB → 1.3 GBpbmc68k · small4 threads · 25.2× speedup20.90 s baseline → 829 ms optimizedmemory 3.5 GB → 1.6 GBpbmc68k · small32 threads · 12.0× speedup22.31 s baseline → 1.89 s optimizedmemory 3.6 GB → 1.6 GBheart_adult · large1 threads · 8.05× speedup1.97 min baseline → 14.83 s optimizedmemory 32 GB → 19 GBheart_adult · large4 threads · 14.3× speedup1.25 min baseline → 5.24 s optimizedmemory 32 GB → 19 GBheart_adult · large32 threads · 23.7× speedup1.28 min baseline → 3.27 s optimizedmemory 33 GB → 19 GBsplitseq_rosenberg · ood_large11 threads · 8.98× speedup42.10 s baseline → 4.70 s optimizedmemory 9.3 GB → 4.7 GBsplitseq_rosenberg · ood_large14 threads · 16.4× speedup27.21 s baseline → 1.66 s optimizedmemory 9.4 GB → 4.7 GBsplitseq_rosenberg · ood_large132 threads · 23.1× speedup27.56 s baseline → 1.19 s optimizedmemory 9.8 GB → 4.9 GBgastrulation_pijuansala · ood_large31 threads · 4.90× speedup26.48 s baseline → 5.41 s optimizedmemory 15 GB → 15 GBgastrulation_pijuansala · ood_large34 threads · 12.0× speedup20.70 s baseline → 1.72 s optimizedmemory 15 GB → 15 GBgastrulation_pijuansala · ood_large332 threads · 15.7× speedup18.52 s baseline → 1.17 s optimizedmemory 15 GB → 15 GBtms_ss2 · ood_large21 threads · 6.67× speedup23.80 s baseline → 3.54 s optimizedmemory 9.5 GB → 8.9 GBtms_ss2 · ood_large24 threads · 13.6× speedup17.89 s baseline → 1.34 s optimizedmemory 9.7 GB → 8.9 GBtms_ss2 · ood_large232 threads · 10.7× speedup14.88 s baseline → 1.39 s optimizedmemory 9.9 GB → 8.9 GB
pbmc200k_glaucomapbmc68kheart_adultsplitseq_rosenberggastrulation_pijuan…tms_ss2
Memory
Baseline vs optimized peak memory on Windows
0.0 GB25 GB50 GBheart_adult0.59×gastrulation_piju…1.02×pbmc200k_glaucoma0.56×tms_ss20.90×splitseq_rosenberg0.50×pbmc68k0.44×heart_adult · largememory 33 GB → 19 GBoptimized / baseline 0.59×23.7× speedup · 32 threadsgastrulation_pijuansala · ood_large3memory 15 GB → 15 GBoptimized / baseline 1.02×15.7× speedup · 32 threadspbmc200k_glaucoma · mediummemory 14 GB → 7.9 GBoptimized / baseline 0.56×36.0× speedup · 32 threadstms_ss2 · ood_large2memory 9.9 GB → 8.9 GBoptimized / baseline 0.90×10.7× speedup · 32 threadssplitseq_rosenberg · ood_large1memory 9.8 GB → 4.9 GBoptimized / baseline 0.50×23.1× speedup · 32 threadspbmc68k · smallmemory 3.6 GB → 1.6 GBoptimized / baseline 0.44×12.0× speedup · 32 threads
baselineoptimized

What is accelerated

The public API stays the same; AutoZyme replaces only the supported fast path.

This task targets pca in Scanpy. The benchmarked result preserves the declared scientific output gate while reducing CPU runtime on the listed datasets.

Also searched as: PCA, principal components, dimensionality reduction, RunPCA, pp.pca, tl.pca.

Supported scope

The fast path computes a full, zero-centered PCA via a Gram matrix (X^T X scaled by 1/(n_cells-1)) plus a symmetric eigendecomposition, then projects centered X onto the top n_comps eigenvectors. Read full supported scope

The fast path computes a full, zero-centered PCA via a Gram matrix (X^T X scaled by 1/(n_cells-1)) plus a symmetric eigendecomposition, then projects centered X onto the top n_comps eigenvectors. It correctly handles: dense or sparse adata.X (sparse is densified via toarray, line 159); any n_comps valid for the matrix; copy=True/False; and the implicit upstream zero_center=True / arpack-or-auto / full deterministic-solver case (output is mathematically equivalent up to sign, which is what the eval metric min_pc_cor>=0.95 checks). use_highly_variable is honored explicitly (lines 121-145): when an HVG annotation is present (or use_highly_variable=True) and the mask is a proper subset, it recurses on the HVG-subset matrix and lifts PCs back into full var space with zeros for non-HVG genes; use_highly_variable=False forces the full-gene path. Matrices with n_genes>8000 fall through to upstream ARPACK (line 154-156) where all kwargs are forwarded, and zyme=False (line 105-106) forwards everything to upstream. On macOS it uses Apple Accelerate cblas_sgemm + numpy.linalg.eigh; elsewhere numpy BLAS + scipy.linalg.eigh(driver="evr"). All computation is done in float32 internally regardless of requested dtype.

Out-of-scope behavior

silent fallback to upstream

Show detailed speedup table 11 runs
Dataset Tier Platform Threads Baseline Optimized Speedup Memory Concordance Pass
gastrulation_pijuansala ood_large3 Windows 32 18.52 s 1.17 s 15.7× 14.8 → 15.0 GB — pass
heart_adult large Windows 32 1.28 min 3.27 s 23.7× 32.7 → 19.3 GB — pass
pbmc200k_glaucoma medium Windows 32 58.94 s 1.63 s 36.0× 14.0 → 7.9 GB — pass
pbmc68k small Windows 4 20.90 s 829 ms 25.2× 3.5 → 1.6 GB — pass
splitseq_rosenberg ood_large1 Windows 32 27.56 s 1.19 s 23.1× 9.8 → 4.9 GB — pass
tms_ss2 ood_large2 Windows 4 17.89 s 1.34 s 13.6× 9.7 → 8.9 GB — pass
gastrulation_pijuansala ood_large3 macOS 4 8.75 s 815 ms 12.3× 14.4 → 14.7 GB — pass
pbmc200k_glaucoma medium macOS 4 28.54 s 976 ms 29.6× 16.5 → 10.3 GB — pass
pbmc68k small macOS 14 12.35 s 600 ms 20.5× 7.3 → 2.2 GB — pass
splitseq_rosenberg ood_large1 macOS 4 13.93 s 911 ms 17.0× 14.1 → 7.1 GB — pass
tms_ss2 ood_large2 macOS 14 7.47 s 766 ms 9.30× 9.2 → 9.0 GB — pass

Frequently asked questions

Speeding up Scanpy pca
Why is Scanpy pca slow?

Scanpy pca is CPU-bound, and the stock implementation in Scanpy leaves performance on the table in its core numerical work. On the benchmark datasets the original takes 58.94 s where the AutoZyme path takes 1.63 s (36.0× faster).

How do I make Scanpy pca faster?

Install AutoZyme and activate the Scanpy patch, then keep using Scanpy pca exactly as before. AutoZyme transparently substitutes the faster, output-validated path, up to 36.0× faster on the benchmark datasets, with no pipeline or API changes.

Does the AutoZyme speedup change the Scanpy pca output?

Effectively no. The output is tolerance-equivalent: held within a frozen concordance gate (up to about 0.6% drift from the original Scanpy result) on every benchmark dataset.

How do I install the Scanpy speedup?

In Python: pip install autozyme, then import autozyme and autozyme.activate("scanpy"). The patch applies automatically the next time you call pca.