Turbopuffer vs. Manticore Search on a couple of cheap VPS
A head-to-head benchmark comparing turbopuffer (serverless managed vector search) against Manticore Search running on two $8.25/month Hetzner VPS, using 975K 1536-dimensional DBpedia vectors. Both systems achieved nearly identical recall (~0.966). Manticore delivered ~4.5× lower cost ($16.50 vs $75/month), faster end-to-end latency when compacted (10–13 ms vs 21 ms) due to eliminating the network round-trip, and comparable unbatched throughput (~329 vs ~296 QPS). turbopuffer wins on zero-ops burden, elastic scaling, and batched throughput ceiling (~800–1,100 QPS). The cost difference is driven almost entirely by turbopuffer's per-GB write charge ($61.44 of the $75 bill). The benchmark concludes that for bounded workloads around 1M vectors at a few QPS, self-hosting Manticore is cost-effective, while turbopuffer justifies its premium for teams wanting fully managed infrastructure or expecting large-scale growth.