Comparison with Microsoft DiskANN #

DiskANN #

DiskANN (Disk-based Approximate Nearest Neighbor) is Microsoft’s original disk-based approximate nearest neighbor search solution designed for large-scale indices. It excels in scenarios requiring massive vector datasets that don’t fit in memory and forms the backbone of Microsoft’s search infrastructure.

Benchmark Limitations #

DiskANN could not be included in the current benchmark results due to compatibility issues with the ann-benchmarks framework. The DiskANN implementation within ann-benchmarks appears to be hardcoded to require AVX512 CPU instructions without providing fallback support for other x86_64 processors. This issue has been documented in ann-benchmarks GitHub issue #240.

Future Work #

We plan to address the DiskANN compilation issues to support a wider range of CPUs in future benchmark iterations. Based on official ann-benchmarks results, PatANN likely outperforms DiskANN as well, but direct comparative testing in our environment will be conducted once we resolve the DiskANN and ann-benchmarks compatibility issues.