Memory footprint, build time, and query throughput on SIFT-100K and SIFT-1M.
Memory is the cost lever: a smaller footprint means lower infrastructure cost for large vector workloads. Try it in pilot.