Vector Search

Guide / 1536 dimension embedding memory calculator

1536 Dimension Embedding Memory Calculator

1536-dimensional embeddings are common enough that their memory cost deserves a simple first-principles budget.

Why 1536 dimensions matters

At 1536 dimensions, every float32 vector stores 6144 bytes of coordinate data. That means 1 million vectors start at about 6.14 GB before the database adds anything else.

The dimension count is fixed per model, so vector count and precision become the fastest levers for estimating storage and RAM.

Calculation by vector count

For 1536-dimensional float32 vectors, multiply vector count by 6144 bytes. For float16, multiply by 3072 bytes. For int8, multiply by 1536 bytes.

Then add index overhead. HNSW graph links can make the real in-memory footprint meaningfully larger than the raw payload, especially when metadata filters and replicas are included.

How to use the calculator

Start with raw payload, then choose the index type and compression assumption. Add separate line items for metadata, filter indexes, replicas, shard headroom, and rebuild margin.

If the calculated payload already consumes most of the instance RAM, the production design needs compression, sharding, fewer dimensions, or a different retrieval architecture.

Complexity Table

Vector countFloat32 raw RAMFloat16 raw RAMPlanning risk
100KAbout 0.61 GBAbout 0.31 GBUsually manageable
1MAbout 6.14 GBAbout 3.07 GBOverhead becomes important
10MAbout 61.44 GBAbout 30.72 GBRequires careful instance sizing
100MAbout 614 GBAbout 307 GBCompression and sharding are usually required

When to Use This

  • Use this page when your embedding model outputs 1536 dimensions and you need a quick memory estimate.
  • Use it before comparing managed vector database pricing tiers.
  • Use it to decide whether float16, int8, or product quantization should be tested.

When Not to Use This

  • Do not use it for models with 384, 768, 1024, or 3072 dimensions without changing the dimension value.
  • Do not assume the raw payload includes HNSW graph links, payload indexes, metadata, replicas, or rebuild buffers.

Production Failure Modes

Teams often size 1536-dimensional vectors from raw payload alone, then discover that HNSW, metadata filters, and replicas exceed the instance.

Another failure is shrinking precision without measuring recall. Lower precision saves memory, but it can change ranking behavior and may require reranking.

Animated SVG Diagram

1536-dimensional embedding memory calculation A fixed 1536 dimensions multiplied by vector count and precision becomes the raw payload before index overhead. Query intent + constraints Lexical Signal terms, filters, IDs Vector Signal embeddings, ANN Ranked Context candidates + evidence
A fixed 1536 dimensions multiplied by vector count and precision becomes the raw payload before index overhead.

Next Topics

vector database memory calculatorhow much RAM for 1 million embeddingsproduct quantization for vector compressionproduction RAG retrieval architecture

FAQ

How many bytes is one 1536-dimensional float32 vector?

One 1536-dimensional float32 vector is 6144 bytes because each dimension uses 4 bytes.

How much RAM do 10 million 1536-dimensional embeddings need?

Raw float32 payload is about 61.44 GB before index overhead, metadata, replicas, and headroom.

Is 1536 dimensions too large for vector search?

Not automatically. It depends on vector count, latency targets, index type, compression, filters, and available RAM.