Request: guidelines for memory requirement when distributing index

[leothomas]

I'm working on an instance of a FAISS index distributed across EC2 instances in AWS.

Each EC2 instance has 8Gb of ram and I have ~250 million vectors to index, with 128 dimensions each (after PCA).

In theory, 8Gb could fit: 8E9bytes / (128 dim/vector * 4 bytes/dim) ~= 15.6 million vectors but I assume the index itself has additional memory requirements, and the search operations themselves consume additional memory.

Are there any guidelines for how many "shards" I should distribute my index into?

Cheers!