Performance Analysis of Solidigm SSDs in RAG Applications

This technical report presents the performance analysis and real-world application of Solidigm SSDs in Retrieval-Augmented Generation (RAG) solutions. It addresses the challenges posed by memory limitations in deploying AI workloads, particularly in managing large-scale vector databases and AI models. The report outlines how offloading data storage to SSDs can provide a scalable and cost-effective alternative to high-memory GPUs, thus enhancing AI capabilities. Performance tests were conducted using VectorDBBench to evaluate the effectiveness of Solidigm SSDs in vector search workloads. Key findings include significant improvements in query speed, with SSD offloading achieving higher queries per second (QPS) compared to traditional in-memory indexing methods. The report also details the reduction in DRAM usage and maintains high recall accuracy, demonstrating that SSD offloading does not compromise performance. Overall, the findings support the use of Solidigm SSDs for efficient and scalable AI applications.