Accelerating FP16 vector search performance using bulk SIMD in OpenSearch 3.5

OpenSearch has significantly improved the performance of vector search using 16-bit floating point numbers (FP16) over the last three releases. Starting with memory-optimized search in 3.1, they addressed FP16 bottlenecks by introducing SIMD (Single Instruction, Multiple Data) calculations in 3.4 and then further optimizing with bulk SIMD processing in 3.5. These optimizations resulted in a 310% increase in throughput and nearly a 300% reduction in latency for FP16 vector searches, transforming OpenSearch into a high-performance vector database.