Embedding Dimension Storage Calculator
Compare vector storage at float32, float16 and int8.
With the default inputs, Storage by vector precision — 1,000,000 vectors x 1,536 dims. Enter your own numbers below to recompute instantly; the full step-by-step math is shown under the worked example.
Results update automatically as you type.
| float32 | 4 | 8.01 GB | $2.00 |
| float16 | 2 | 4.01 GB | $1.00 |
| int8 | 1 | 2.00 GB | $0.5007 |
Quantizing vectors can slash storage. Compare the size and monthly cost of your vector set at float32, float16 and int8 precision.
How this is calculated
For each precision we compute raw size = vectors × dimensions × bytes (4 for float32, 2 for float16, 1 for int8), add ~40% index overhead, and price it at your $/GB-month. It shows the storage trade-off of quantizing your vectors side by side.
Is this a good result? What to do next
Quantization is a big storage lever with a small quality cost: float16 is usually lossless for retrieval, and int8 trades a little accuracy for a 4× reduction versus float32. On large indexes that's a meaningful monthly saving — test int8 on your own queries first.
Typical planning ranges
- float32 → float16
- 2× smaller
- float32 → int8
- 4× smaller
- Index overhead
- ~40%
Ranges are typical planning figures to sanity-check your result, not authoritative benchmarks. Your numbers will vary with use case, volume, and vendor.
How to improve this number
- Move to float16 (usually lossless) as a safe default.
- Use int8 on large indexes after testing recall.
- Reduce dimensions where the embedding model supports it.
Common mistakes
- Storing float32 when float16/int8 is accurate enough.
- Assuming quantization is free of any recall impact.
When to use a different approach
For storage plus query cost, use the vector database cost calculator. For the full RAG pipeline, use the RAG cost calculator.
Worked example (defaults)
With the default inputs above, here is the result:
| float32 | 4 | 8.01 GB | $2.00 |
| float16 | 2 | 4.01 GB | $1.00 |
| int8 | 1 | 2.00 GB | $0.5007 |
Frequently asked questions
Does quantization hurt accuracy?+
Float16 is usually lossless for retrieval; int8 trades a little accuracy for big storage savings. Test on your data.
What's the overhead factor?+
Indexes (like HNSW) add roughly 40% on top of raw vector bytes, which this includes.