FIG. 07 // HSPANN.MACHINES FINEWEB · 100M VECTORS · R6I.4XLARGE × 1–8
One 100M index served from 1, 2, 4 and 8 machines
More machines share the scan, but routing remains: memory per machine falls and throughput rises, while single-query latency improves less because every query routes on one machine.
[
Each machine: 16 vCPUs, 128 GB · posting lists and full vectors in a block store, cached in memory · 1,000 queries · latency one query at a time on warm caches · throughput at 32 queries in flight. Select ⓘ for each measure’s definition.