← All questions
Retrieval & EmbeddingsMedium· Memory Estimation· Vector Databases

How much memory does a 10-million-chunk index at 768 di...

A strong answer refuses to provide a single number without knowing the index type, precision, and whether the payload is resident. It builds the estimate per vector: 3,072 bytes of float32, plus 260 bytes of HNSW links at M=32, plus 8 bytes for an ID in the IVF case, multiplied by N, plus fixed centroids and rebuild/replica overhead. A weak answer simply calculates 10M * 768 * 4 bytes, which fails to account for the structural overhead that leads to OOM errors.

Code playground (Python)