Estimated footprint
RAM (resident index)
—
Disk (serialized)
—
Per vector
—
Per 1M vectors
—
Index overhead share
Composition breakdown
Vector payload
—
Index structure
—
ID map
—
Total RAM
—
Precision × index comparison (RAM)
| Flat | HNSW | IVF-PQ | |
|---|---|---|---|
| FP32 | |||
| FP16 | |||
| INT8 |
Click any cell to apply that configuration.
Footprint by configuration
Assumptions & notes
- Flat stores every vector verbatim; RAM ≈ raw bytes + optional 64-bit IDs.
- HNSW adds bidirectional level-0 links: ≈ (8·M + 8) bytes per vector, i.e. 136 B at the default M=16.
- IVF-PQ compresses each vector to one byte per PQ segment (⌈d÷segDims⌉ segments) plus nlist·d·4 B of coarse centroids; nlist follows the √N heuristic.
- Disk figures add ~2% serialization/container overhead; memory-mapped engines may run below the RAM figure on disk-backed indices.
- FP16/INT8 assume engine-side quantized storage with rescoring — verify recall on your own workload before shipping.
- HNSW construction can transiently need ~1.5–2× resident graph memory during build.