How much does a vector cost to store?

A 1,536 dimension float32 vector is about 6 KB, so a million of them is roughly 6 GB before index overhead.

Is quantisation worth it?

At scale usually. int8 uses a quarter of float32 storage with a small retrieval accuracy cost, which is worth measuring on your own data before committing.

Technologyai

Embedding Storage Cost Calculator

Vector storage and embedding cost. Storage is vectors times dimensions times bytes, and quantising from float32 to int8 cuts it by four with a small accuracy cost that is usually worth taking at scale.

Also called: vector database cost, embedding cost calculator.

Vector storage

22.89

22.89 GB for 40,00,000 vectors of 1,536 dimensions at float32. Generating them costs ₹16,000 once, and storage runs ₹4,578 a month. Quantising to int8 would use 5.72 GB. This is raw vector storage. An index structure typically adds twenty to fifty percent on top.

Total vectors

40,00,000

Bytes per vector

6,144

One-off embedding cost

₹16,000

Monthly storage

₹4,578

If quantised to int8

5.72

Tokens embedded, millions

1,600

Annual storage

₹54,932

On index overhead

This is raw vector storage. An index structure typically adds twenty to fifty percent on top.

How this is calculated

Method and background

How this is calculated

Each vector is dimensions times bytes per component, so a 1,536 dimension float32 vector is about 6 KB. At millions of chunks that becomes substantial, and the index structure typically adds twenty to fifty percent on top of the raw vectors. Quantisation is the main lever: int8 uses a quarter of float32 with a small retrieval accuracy cost that most applications find acceptable. Embedding generation is a one-off cost per chunk and is usually small relative to ongoing storage.

vectors times dimensions times bytes per component, before index overhead

n: Vector count
d: Dimensions
b: Bytes per component

Worked examples

Each of these is asserted on every build. If a change to the engine ever moved one of these answers, the build would fail before the page could print it.

4 million vectors at 1536 dimensions

Documents: 5,00,000
Chunks per document: 8
Embedding dimensions: 1,536
Stored precision: float32
Tokens per chunk: 400
Embedding price per million tokens: ₹10
Vector store per GB month: ₹200

Vector storage22.89

1536 x 4 bytes; 4M x 6144 over 1024^3

Open this example

int8 quantisation uses a quarter

Documents: 5,00,000
Chunks per document: 8
Embedding dimensions: 1,536
Stored precision: int8
Tokens per chunk: 400
Embedding price per million tokens: ₹10
Vector store per GB month: ₹200

Vector storage5.72

boundary: exactly a quarter of float32

Open this example

Method and limits

What it assumes

Raw vector storage before index overhead, which is shown as a note.

What it deliberately does not model

Index structures add twenty to fifty percent depending on the algorithm and parameters.
Re-embedding after a model change repeats the full generation cost.
Quantisation accuracy loss depends on the data and should be measured, not assumed.

Formula version 1.0.0 · definition 1.0.0 · India · Report a problem with this calculator

Frequently asked questions

How much does a vector cost to store?: A 1,536 dimension float32 vector is about 6 KB, so a million of them is roughly 6 GB before index overhead.
Is quantisation worth it?: At scale usually. int8 uses a quarter of float32 storage with a small retrieval accuracy cost, which is worth measuring on your own data before committing.