MundoKit

Vector Database Storage Calculator

Enter vector dimensions, point counts and storage choices to estimate cluster disk and RAM.

Inputs
For differing dimensions, calculate each vector separately; count payload and IDs only once.
0 disables the copy. Example scenario: 0.5 for a 4-bit copy. Originals remain on disk.
Adds about 4 RAM bytes per point per index, regardless of memory tier.

Results

Cluster disk with headroom
10.76 GiB
Cluster RAM budget with headroom
7.33 GiB
Per-element index RAM before headroom
0 B
Raw original vectors
5.72 GiB
Across all copies, before headroom
ComponentDisk bytes
Original vectors6,144,000,000
Quantized copy0
HNSW graph307,200,000
Payload storage3,072,000,000
Payload indexes0
ID tracker104,000,000
  • RAM includes graph, payload indexes and IDs. Cold originals or payloads can still use OS cache. Sparse vectors, backups and workload-specific overhead need separate measurement. Binary units: GiB = 1,073,741,824 bytes.

Results update as you type. The share link stores your inputs after the # in the address, which is not sent to our server, but anyone you send it to can read them.

How it works

Uses the component formulas in Qdrant’s capacity planning guide. Original bytes = points × copies × named vectors × dimensions × bytes/dimension. Quantized copies are additional storage. Each equal-sized named vector gets its own graph.

HNSW = points × copies × named vectors × m × 2 × 4 × 1.2

Payload disk = points × copies × average bytes × 1.5; cached payload RAM is three times that. Payload indexes use twice the indexed bytes. Each per-element index adds about 4 RAM bytes per point per copy, regardless of memory tier; this separate allowance receives the same headroom. The index count defaults to zero; IDs use 52 bytes per replicated point. Add the entered headroom to each total. Graph, indexes and IDs are included in RAM. Round only displayed values to two decimals; byte calculations remain unrounded.

Worked examples

Raw vectors with two copies

One million points × 768 dimensions × 4 bytes × 2 copies = 6,144,000,000 original bytes. With default graph, payload and headroom, disk is 11,552,640,000 bytes and RAM is 7,866,240,000 bytes.

Quantized copy

With the same defaults and 0.5 quantized bytes per dimension, the extra copy uses 768,000,000 bytes. Disk with headroom becomes 12,474,240,000 bytes.

Frequently asked questions

Does quantization replace originals?
This model keeps originals on disk and adds a compressed copy.
Are replicas extra copies?
Total copies includes the primary: 2 means one primary and one additional copy.
Can I use different named vector dimensions?
Run each vector separately, adding payload and ID storage only once.
Is this measured RAM usage?
No. It is a component budget; measure your workload and cold-page cache separately.

Sources

Last reviewed 2026-11-14; next review due by 2027-02-12. The method is written out above so you can check our math.