Vector Search Bench

Vector Search Bench is a high-signal, no-nonsense benchmark suite for vector similarity search on iPhone. Measure how fast and stable your device can score embeddings using the CPU (Accelerate/vDSP) versus the Apple Neural Engine (Core ML)—with real metrics you can trust. What you can benchmark Exact Top-K vector search (dot-product similarity) CPU path: Accelerate / vDSP optimized scoring NPU path: Core ML Dense inference tuned for Neural Engine scheduling Query-by-query execution with repeatable timing Intricate performance metrics Latency distribution: median, p95 Throughput: QPS Recall@K vs CPU reference (sanity check) Memory footprint and dataset sizing knobs Built for real-world AI systems