Open abhinavdangeti opened 4 months ago
@metonymic-smokey would you share the numbers you have comparing the considered index classes for this doc count at the unit level.
Numbers from go-faiss unit level testing -
nq = 100, k = 100
256 dims
nvecs |
QPS without HNSW | QPS with HNSW |
---|---|---|
200k | 425 | 311 |
500k | 257 | 244 |
1M | 170 | 145 |
512 dims
nvecs |
QPS without HNSW | QPS with HNSW |
---|---|---|
1M | 75 | 62 |
1024 dims
nvecs |
QPS without HNSW | QPS with HNSW |
---|---|---|
1M | 20 | 15 |
2048 dims
nvecs |
QPS without HNSW | QPS with HNSW |
---|---|---|
200k | 14 | 5 |
500k | 11 | 5 |
1M | 9 | 4 |
These numbers were prior to the latest go-faiss PRs being merged. I'm not sure that 100k is a high enough number to use the coarse quantiser and the number should depend on the dims too since it's most effective for high dims.
Hmm, but per the numbers you've shared it seems even at higher dims the QPS seems to fall with the coarse quantizer?
Bringing back the coarse quantizer into use per observations Aditi recorded earlier. A slight adjustment now is to deploy it only over a vector count of 100000.