Shattering the Limits of Search with Domain Specific Computing
summary
Claims a purpose‑built search chip and SPU architecture (FPGA instances on AWS) that runs hybrid lexical+vector search, rebuilding indexing, candidate‑generation then multi‑threaded ranking, and tiered memory placement to shrink indexes and avoid GC. Reports 10x faster throughput, costs cut by half, and a customer case hitting sub‑20ms on 1B docs across 50–200 QPS versus a 100ms target; exposes Elasticsearch‑compatible APIs as a managed cloud service.
AI-written summary · about