Measures the per-epoch MSA load cost: parsing an A3M text file every
epoch vs reading the pre-tokenized MSA from a .ptt. Both yield identical
arrays (verified). Generated by python benchmarks/msa_benchmark.py.
Not measured here: MSA generation (JackHMMER / MMseqs2 / ColabFold). That is the large one-time cost cited in the README as a literature-based projection (~2.4 min/protein); it needs those tools + multi-GB databases and is not run or benchmarked by this script. No generation speedup is claimed here.
- Platform: Windows-10-10.0.26100-SP0
- Python: 3.11.9 | numpy 1.26.4 | zarr 2.18.7 | proteintensor 0.2.0
- Run: 2026-07-02T20:33:51+00:00
| Residues | MSA depth | A3M KB | A3M parse | .ptt read_msa | mmap open | load speedup | lossless |
|---|---|---|---|---|---|---|---|
| 76 | 512 | 43.5 | 6.3034 | 1.8655 | 0.5076 | 3.38x | yes |
| 300 | 1024 | 311.1 | 18.6002 | 4.7671 | 0.4993 | 3.9x | yes |
| 574 | 2048 | 1171.3 | 54.4643 | 11.0017 | 0.492 | 4.95x | yes |
| 1000 | 4096 | 4047.7 | 174.0184 | 29.6402 | 0.5199 | 5.87x | yes |