IVF per vector memory budget seems wrong?

[ssk01]

the picture is 【Decision tree to choose a Faiss index】https://github.com/facebookresearch/faiss/wiki/Guidelines-to-choose-an-index
it shows
【memory budget is
M + 16 bytes
per vector】

my problems is why 16bytes?
it should only a ids(8 bytes) + code(M)

[satymish]

For IVF_HNSW, HNSW is the quantizer for IVF centroids. So the indices of the centroids also need to be stored for HNSW quantization

[ssk01]

example. 1m vector in a IVF1024_HNSW index.
i guess hnsw quantization only use centroids(1k), which is much smaller than vector nums(1m). it can't give addtional 8bytes per vector.

[mdouze]

Hmmm indeed this is a mistake, it should be 8 bytes.