TechPrufer
Get started
Qdrant
Qdrant logoQdrant · Feature trail

CPU Performance for Vector Search

4 posts · Jan 2024 – Aug 2026

What this trail covers

Qdrant continues to optimize performance for vector search, demonstrating significant speed improvements for disk-based retrieval. This post details benchmarks showing Qdrant achieving 2x higher throughput and 50% lower latency than Elastic's DiskBBQ at the same recall target, using significantly less CPU and RAM per node. This is achieved through advanced quantization (TurboQuant 4-bit), async disk scoring, and a two-stage retrieval process. The results highlight Qdrant's efficiency and perform. This post specifically highlights the performance gains achieved by leveraging Intel's new generation Xeon processors (Emerald Rapids), showing up to 1.38x faster query execution and a 2.79x drop in latency compared to previous generations for machines in the 8-64 core range. It emphasizes that CPUs are well-suited for concurrent vector search requests and recommends Intel's chips for optimal performance and cost-efficiency in enterprise AI/ML applications.

All Qdrant trails
  1. Qdrant 1.19 - TurboQuant Datatype & Memory Tiers - QdrantIntroduced the TurboQuant datatype for 4-bit vector compression, reducing storage by up to nine times. Unified memory tier configuration with `pinned`, `cached`, and `cold` tiers. Enabled Per-Tenant IDF Statistics for improved BM25 scoring in multi-tenant environments. Added prefix matching to keyword fields and a slice filter condition for deterministic point partitioning. Enhanced the Web UI with live resharding progress, a scalable Collection Visualizer, and payload index management.
  2. Qdrant Beats Elastic’s DiskBBQ at 2x Throughput, Half the Latency, and 1/3 the Compute - QdrantThis post details a benchmark comparing Qdrant's disk-based retrieval performance against Elastic's DiskBBQ. It demonstrates that Qdrant, using its two-stage retrieval process with TurboQuant 4-bit quantization and async disk scoring, achieves approximately 2x higher throughput and 50% lower latency than Elastic's DiskBBQ at the same recall target (0.96). Crucially, Qdrant achieves this on significantly less compute resources (2 vCPU / 8 GiB RAM per node) compared to Elastic (7 vCPU / 26 GiB RAM per node). The post highlights the efficiency gains from Qdrant's open-source nature, advanced quantization, and optimized disk I/O handling, contrasting it with Elastic's proprietary solution and higher resource requirements.
  3. Qdrant 1.18 - TurboQuant - QdrantThis post introduces TurboQuant, a new quantization method that offers twice the compression ratio of scalar quantization with similar recall and speed. It also details memory monitoring capabilities for collections, the ability to add/remove named vectors without recreating collections, improvements to audit logging (querying logs and request tracing IDs), per-collection API metrics, and new strict mode guardrails for memory and search batch size.
  4. Qdrant 1.15 - Smarter Quantization & better Text Filtering - QdrantIntroduced 1.5-bit and 2-bit quantization for improved accuracy and compression, and asymmetric quantization for smarter memory use by allowing different encoding for stored vectors and queries. Enhanced the text index with multilingual tokenization, stop words, stemming, and phrase matching capabilities. Added Maximal Marginal Relevance (MMR) reranking to balance relevance and diversity in search results. Implemented HNSW healing for index reuse and migration to Gridstore for faster ingestion.
  5. Intel’s New CPU Powers Faster Vector Search - QdrantThis post benchmarks Qdrant's performance on Intel's 5th gen Xeon processors (Emerald Rapids), demonstrating up to 1.38x faster query execution and a 2.79x reduction in latency compared to the 4th gen. It highlights that CPUs are a cost-effective and performant choice for vector search, especially for concurrent requests, and recommends Intel's chips in the 8-64 core range for mass-market use cases.
  7. Qdrant Updated Benchmarks 2024 - QdrantThis post updates Qdrant's benchmarks, introducing a new dataset of 1 million OpenAI embeddings relevant for RAG applications. It separates latency and requests-per-second (RPS) cases to better reflect diverse application requirements. The benchmarking principles remain consistent: focus on open-source solutions, exclude libraries (like FAISS) to focus on vector databases as a whole system, and use ann-benchmarks datasets for comparison.
  8. Binary Quantization - Andrey Vasnetsov | Vector Space Talks - QdrantThis post introduces and explains binary quantization as a method to reduce vector storage size and increase search speed by up to 30x. It details how float point elements are converted to binary (0 or 1) and how this approximates dot product/cosine similarity using Hamming distance, leveraging optimized CPU instructions (Pixor and Popcount). The post also covers the concept of oversampling for real-time precision control during search and discusses model compatibility, highlighting that OpenAI models are more compatible with binary quantization than other open-source models.