induwara.lk
induwara.lkAI · Infrastructure

Vector Database Storage & RAM Calculator

Find out how much disk, RAM, and monthly money your embeddings will cost to store and serve. Enter a vector count, model, and precision to see raw size, HNSW overhead, total RAM, and an estimated Pinecone, Qdrant, Weaviate, or self-hosted bill — no signup.

By Induwara AshinsanaUpdated Jun 6, 2026
Size your vector index
Verified · Qdrant 1.5× rule
Common setups

One per embedded chunk / row.

Stored payload per vector (IDs, text, tags). 0 if none.

CBSL indicative rate — edit for today's value.

Total RAM needed
878.9 MiB
Fits a 1 GiB instance
Total disk
585.9 MiB
Vectors + metadata, no graph
Raw vectors
585.9 MiB
100,000 × 1,536 × 4 B
Index overhead
293.0 MiB
HNSW graph (≈0.5× raw)
Estimated monthly cost — Self-host (pgvector)
No managed fee

Self-provisioned — a 1 GiB RAM instance covers this index.

No managed fee — provision your own instance from the RAM figure.. Storage / RAM tier only — excludes read, write, and query units.

Same vectors at every precision

PrecisionBytes/dimRaw vectorsSaved vs float32
float32 (full precision)4585.9 MiB
float16 / half (halfvec)2293.0 MiB50%
int8 (scalar quantization)1146.5 MiB75%
binary (1-bit quantization)0.12518.3 MiB96.88%

Byte sizes follow Qdrant's capacity-planning and quantization guides. Provider rates are published list prices verified on 2026-06-06 and cover storage / RAM tiers only — confirm against each vendor's live pricing before you commit.

How it works

Vector search costs are mostly storage costs, and storage is just byte arithmetic. Every embedding is an array of numbers; this tool multiplies that out and adds the overhead each index type carries. All sizes use binary units — 1 GiB = 2³⁰ = 1,073,741,824 bytes.

  1. Bytes per dimensionfollow the precision, from Qdrant's quantization guide: float32 = 4 bytes, float16 = 2, int8 = 1, and binary = 0.125 (one bit).
  2. Raw vector storage = num_vectors × dimensions × bytes_per_dim. A 1536-dim float32 vector is 6,144 bytes; a million of them is 5.72 GiB.
  3. Metadata = num_vectors × metadata_bytes_per_vector — the IDs, text snippets, and tags you store alongside each vector.
  4. Index overhead. A Flat (exact) index adds nothing. An HNSW graph adds roughly 0.5 ×the raw vector footprint for its links, which is where Qdrant's capacity rule of thumb — memory ≈ 1.5 × num_vectors × dim × 4 for float32 — comes from.
  5. Total RAM = raw + metadata + overhead (what the engine holds in memory to serve queries). Total disk = raw + metadata (the graph is rebuilt or memory-mapped as needed).
  6. Managed cost. Each provider bills differently: Pinecone Serverless charges per GB of stored data ($0.33/GB-month), Qdrant Cloud is sized by RAM, and Weaviate Serverless bills per stored dimension ($0.05 per million dimensions-month). The tool applies the right basis per provider and converts to LKR using the editable CBSL indicative rate.

The float32 + HNSW path is cross-checked against Qdrant's published 1.5 × n × d × 4 identity to the byte, the same way a tax tool reconciles against the official table. Managed prices are storage / RAM tier only and exclude query, read, and write units — stated plainly so the estimate is never oversold.

Worked examples

100k vectors · text-embedding-3-small (1536) · float32 · HNSW

  1. Raw: 100,000 × 1536 × 4 = 614,400,000 B = 0.57 GiB
  2. HNSW overhead: 0.5 × 614,400,000 = 307,200,000 B = 0.29 GiB
  3. Total RAM: 921,600,000 B = 0.86 GiB (= 1.5 × raw ✓ Qdrant rule)
  4. Total disk: 0.57 GiB — fits a 1 GiB instance
  5. Pinecone: 0.6144 GB × $0.33 = $0.20 / month

1M vectors · text-embedding-3-large (3072) · HNSW · float32 vs int8 vs binary

  1. float32 raw: 1,000,000 × 3072 × 4 = 12.29 GB = 11.44 GiB; RAM ×1.5 = 17.17 GiB
  2. int8 raw: 1,000,000 × 3072 × 1 = 3,072,000,000 B = 2.86 GiB; RAM ×1.5 = 4.29 GiB
  3. int8 saving vs float32 raw: 1 − 1/4 = 75%
  4. binary raw: 1,000,000 × 3072 × 0.125 = 384,000,000 B = 0.36 GiB
  5. binary saving: 1 − 0.125/4 = 96.875% (4 / 0.125 = 32× smaller ✓)

Edge case — Flat index with metadata, 100k vectors · 1536 · float32 · 500 B/vector

  1. Raw: 100,000 × 1536 × 4 = 614,400,000 B
  2. Metadata: 100,000 × 500 = 50,000,000 B
  3. Flat index overhead: 0 (no graph)
  4. Total disk = total RAM: 664,400,000 B = 0.62 GiB
  5. Switching to HNSW would add 0.29 GiB of graph on top of this.

Frequently asked questions

Sources & references

Related tools

Rate this tool
Be the first to rate

Comments & feedback

Spotted a bug or want an improvement? Tell us — our team reviews every comment, and good ideas get built. Comments are public and anonymous.

Found a bug, edge case, or want a provider added?

Email me at [email protected] — most fixes ship within 24 hours.