Vector Database Search

Overview

Vector database search is a core component of AI inference pipelines, performing approximate nearest-neighbor (ANN) similarity search over high-dimensional embedding vectors. It imposes a distinctive storage access pattern: random-read (RR) at 512-byte granularity, with vector DB query serving at low throughput and vector DB index construction at high throughput. These fine-grain random access requirements are poorly served by conventional enterprise SSDs optimized for 4KB I/O.

RAG (Retrieval-Augmented Generation) over GNN (graph neural network) architectures add further demand, requiring LLM+GNN and GNN+LLM workloads with 512-byte RR access at high throughput alongside fast random access to graph metadata and low-latency constraints. SK hynix's AIN P is purpose-built to address the IOPS demands of vector database search, targeting 50M IOPS at 512-byte access on PCIe Gen6 — 7x more than a conventional Gen6 enterprise SSD.