Closed alexklibisz closed 7 months ago
Results on cx41 were:
Model | Parameters | Recall | Queries per Second |
---|---|---|---|
eknn-l2lsh | L=100 k=4 w=1024 candidates=500 probes=0 | 0.379 | 133.326 |
eknn-l2lsh | L=100 k=4 w=1024 candidates=1000 probes=0 | 0.446 | 118.878 |
eknn-l2lsh | L=100 k=4 w=1024 candidates=500 probes=3 | 0.635 | 101.896 |
eknn-l2lsh | L=100 k=4 w=1024 candidates=1000 probes=3 | 0.717 | 94.297 |
eknn-l2lsh | L=100 k=4 w=2048 candidates=500 probes=0 | 0.767 | 115.363 |
eknn-l2lsh | L=100 k=4 w=2048 candidates=1000 probes=0 | 0.846 | 104.785 |
eknn-l2lsh | L=100 k=4 w=2048 candidates=500 probes=3 | 0.921 | 79.075 |
eknn-l2lsh | L=100 k=4 w=2048 candidates=1000 probes=3 | 0.960 | 73.672 |
Performance in Hetzner was terrible compared to AWS. I'll open a PR shortly that adds a terraform file for spinning up a development and benchmarking instance in AWS.
Related Issue
...
Changes
...
Testing and Validation
...