Embedding Storage Calculator

How much space do your embeddings take? Enter how many vectors you will store, their dimensions and the precision they are kept in, and this calculator multiplies it out exactly: vectors ร— dimensions ร— bytes per value. Add a metadata allowance if each vector carries an id or payload. For the operational size of a vector database with an index and replicas, use the vector database size estimator.

Processed locally in your browser. Your data is not uploaded to our servers.

One vector per chunk, document or item you embed.

Your embedding modelโ€™s output size, from its documentation.

Ids, text snippets or payload stored beside each vector.

Result

Exact arithmetic on your numbers

Enter the number of vectors and dimensions to see the storage.

Raw storage only: no index, replicas, compression or database overhead is included.

Examples

Worked examples (no metadata):

VectorsDimensionsPrecisionPer vectorRaw vectors
1,000,0001,536FP326 KiB (6.14 KB)5.72 GiB (6.14 GB)
10,000,000768FP161.5 KiB (1.54 KB)14.31 GiB (15.36 GB)
100,000,0001,024INT81 KiB (1.02 KB)95.37 GiB (102.4 GB)
1,000,000,0001,024Binary128 B119.21 GiB (128 GB)

Formula & Methodology

Raw storage is exact arithmetic on the numbers you enter:

bytes per vector = dimensions ร— bytes per valueraw vector bytes = vectors ร— dimensions ร— bytes per valuetotal bytes = raw vector bytes + vectors ร— metadata bytes per vector

Good to know

  • Bytes per value: FP32 is 4, FP16 and BF16 are 2, INT8 is 1 and binary quantization stores one bit per dimension (0.125 bytes).
  • GB (10โน bytes) and GiB (2ยณโฐ bytes) differ by about 7%; the results show both.
  • Some systems keep the original full-precision vectors as well as a quantized copy. Count both if yours does.
  • Lower precision saves space but can reduce retrieval quality. Test on your own data before committing.
  • This is the size of the vectors themselves. A vector database adds an index, replicas and other overhead: see the vector database size estimator.

How to Use

  1. Enter the number of vectors you will store.
  2. Choose the dimensions of your embedding model and the precision of the stored values.
  3. Optionally add the metadata bytes stored with each vector.
  4. Read the storage per vector and in total, in both GB and GiB.

Frequently Asked Questions

How do I calculate embedding storage?

Multiply the number of vectors by the dimensions and by the bytes per value. For one million 1,536-dimension FP32 vectors that is 1,000,000 ร— 1,536 ร— 4 bytes, about 6.1 GB (5.7 GiB).

How many bytes is one embedding?

Dimensions ร— bytes per value. A 1,536-dimension FP32 vector is 6,144 bytes; at FP16 it is 3,072; at INT8 it is 1,536; with binary quantization it is 192.

Does quantization reduce storage?

Yes, in proportion to the bytes per value: INT8 is a quarter of FP32 and binary is 1/32. It can lower retrieval quality, and some systems also keep the original vectors, so check both size and quality.

Is the metadata allowance required?

No. Leave it empty or zero to see the vectors alone. If each vector carries an id, source or text, add the average bytes so the total reflects what you actually store.

Is this the size of my vector database?

Not exactly. This is the raw vector storage. A database adds an index, replicas and other overhead, which depend on the system and settings. Use the vector database size estimator for a range.

Last updated: