Intel® Xeon® Processors Set the Standard for Vector Search Benchmark Performance

Authors:

Nithya Rao, System and Software Optimization Engineer, Intel

Cecilia Aguerrebere, AI Research Scientist, Intel

Mihai Capotă, AI Research Scientist, Intel

Hardik Jain, Software Engineer, HCL Technologies

Ishwar Singh Bhati, AI Research Scientist, Intel

As vector search becomes a core building block in artificial intelligence (AI)-powered systems — from semantic search and recommendation engines to anomaly detection and generative AI — the performance of the underlying hardware is increasingly important. These workloads depend on large-scale, high-dimensional similarity computations, where processor architecture can directly affect both latency and overall throughput.

In real-world vector search performance tests, Intel® Xeon® server architectures outperform AMD EPYC processors when running two commonly used vector search frameworks: Redis Open Source 8.2 with integrated support for Intel® Scalable Vector Search (Intel® SVS) and Faiss (Facebook AI Similarity Search)*.

Unlike traditional search approaches that rely on exact keyword matches, vector search enables applications to retrieve results based on semantic similarity. This makes it especially valuable for use cases such as contextual search, personalization, computer vision, retrieval-augmented generation (RAG), and agentic AI. As these applications grow in scale and complexity, efficiently executing vector search operations becomes essential to meeting performance and cost goals.

Vector databases also play a key role when paired with large language models (LLMs). While LLMs handle language understanding and generation, vector search provides a fast and scalable way to retrieve relevant context stored as embeddings. This combination enables RAG and agentic AI systems that ground model responses in external, domain-specific data, supporting more reliable and context-aware applications such as enterprise search, chatbots, and AI assistants.

Pushing the Boundaries of Vector Search with Intel® Scalable Vector Search (Intel® SVS)

In AI systems such as recommendation engines, RAG workflows, and agentic AI, vector search has become a critical bottleneck, especially when dealing with billion-scale, high-dimensional embedding collections. Intel SVS is a software library purpose-built to address this challenge. It combines highly optimized vector indexing, intelligent compression techniques, and dimensionality reduction algorithms to deliver strong performance on Intel Xeon platforms. This enables fast similarity search while using substantially less memory than alternatives.

Intel SVS enables high accuracy and blazing-fast search across billions of vectors through its expertly tuned implementation, far beyond the reach of brute-force scanning. Its design prioritizes both throughput and memory efficiency by reducing data movement via compression and leveraging single instruction multiple data (SIMD)-ready compute kernels. The result: developers can deploy scalable, low-latency similarity search directly on Intel Xeon hardware without resorting to expensive or memory-limited GPU alternatives.

Using highly optimized indexing and search algorithms combined with Intel’s novel vector compression techniques, including Locally-adaptive Vector Quantization (LVQ) and LeanVec, Intel SVS provides vector similarity search with the following benefits:

• Delivers strong vector search performance with fast retrieval while using significantly less memory and without sacrificing accuracy compared to the most popular alternative, HNSW.
• Scales effortlessly to handle billions of high-dimensional vectors dimensional vectors.

This enables application and framework developers using similarity search to unleash Intel SVS performance on Intel Xeon CPUs, delivering faster, more scalable vector search than competing architectures.

The following video explains how Intel SVS works by comparing Intel Xeon 6900 processors to AMD Genoa/EPYC Zen 4.

Comparison of Vector Search Frameworks on Intel Xeon and AMD EPYC

Redis is an in-memory database that provides native support for vector similarity search via the RediSearch module. It runs as a lightweight, stand-alone server, allowing multiple clients to connect and perform operations concurrently. It is well-suited for deployment in large-scale, real-time production applications, such as large RAG systems.

We used the Cohere 768 vector search dataset to benchmark Redis’s search performance on Intel Xeon 6980P and AMD EPYC 9965. The dataset contains 768-dimensional vector representations of Wikipedia articles produced by Cohere's embedding model. We tested search performance by indexing the embeddings in Redis using HNSW and SVS and measuring latency/throughput under varying query loads. HNSW is a widely used graph-based vector search algorithm that achieves both high speed and high recall.

Combined with the architecture-optimized performance of Intel SVS, Redis delivers up to 2x higher queries per second when running on Intel Xeon 6980P compared to AMD EPYC 9965 (see Figure 1)*.

 Figure 1: Intel Xeon 6980P delivers up to 2x higher Redis queries-per-second versus AMD EPYC 9965.

While Redis vector search offers fast in-memory search and Intel SVS delivers highly optimized performance, many AI pipelines also rely on Faiss for its flexible indexing options and support for both exact and approximate search at scale. Faiss is widely used in offline and batch-heavy inference scenarios, offering index types such as IVF, PQ, and HNSW, and supporting both CPU and GPU execution. Its modular design allows fine-grained control over recall, latency, and memory usage, making it a strong fit for experimentation and hybrid pipelines.

In addition to Cohere 768, we used two other popular vector search benchmarking datasets to test the vector search performance of Faiss when running on Intel Xeon 6980P and Intel Xeon 6972P, and on AMD EPYC 9965, AMD EPYC 9755, and AMD EPYC 9655 processors. SIFT 128 provides a set of 128-dimensional scale-invariant feature transform descriptors for images, while DBpedia 1536 has 1536-dimensional OpenAI text embeddings of entities from the DBpedia knowledge base. We found that Intel Xeon 6980P delivers up to 1.9x higher Faiss performance compared to AMD EPYC 9965 (see Figure 2)*.

When running Faiss on Intel Xeon 6 platforms , we consistently observed superior performance across both exact and approximate indexing modes. Thanks to high memory bandwidth with MRDIMMs and large cache architecture, Intel Xeon processors accelerate distance computations and reduce memory bottlenecks, especially in high-dimensional, large-batch search scenarios.

 Figure 2: Intel Xeon 6980P delivers up to 1.9x higher Faiss performance versus AMD EPYC 9965.

A Clear Advantage: Vector Search Performance on Intel Xeon Processors

Whether it’s architecture-optimized performance with Intel SVS or flexible indexing with Faiss, Intel Xeon platforms consistently deliver. With high core scalability, SIMD acceleration, and efficient memory architecture, Intel Xeon processors can provide a unified, high-performance foundation for vector search at any scale.

Learn how Intel Xeon processors can improve the performance of your vector search workloads.

*Product and Performance Information

Redis Vector Search

Tests by Intel as of December 2025. Your results may vary. Intel technologies may require enabled hardware, software, or service activation.

Hardware Configurations

Intel Xeon 6980P: 1-node, 2x Intel(R) Xeon(R) 6980P 128-Core Processor, 128 cores, HT On, Turbo On, Total Memory 1536GB (24x64GB DDR5 8800 MT/s [8800 MT/s]), BIOS BHSDCRB1.IPC.3544.O01.2503080923, microcode 0x10003d0, 2x Ethernet Controller E810-C for QSFP, 2x Ethernet Controller X710 for 10GBASE-T, 1x 3.5T SAMSUNG MZWLJ3T8HBLS-00007, Ubuntu 22.04.4 LTS, 6.5.0-21-generic. Test by Intel as of December 2025.

AMD EPYC 9965: 1-node, 2x AMD EPYC 9965 192-Core Processor, 192 cores, SMT On, Boost On, Total Memory 3072GB (24x128GB DDR5 6400 MT/s [6000 MT/s]), BIOS 1.4, microcode 0xb101047, 2x MT2910 Family [ConnectX-7], 2x BCM57416 NetXtreme-E Dual-Media 10G RDMA Ethernet Controller, 1x 3.5T Micron_7450_MTFDKCB3T8TFR, Ubuntu 22.04.4 LTS, 6.5.0-21-generic. Test by Intel as of December 2025.

AMD EPYC 9755: 1-node, 2x AMD EPYC 9755 128-Core Processor, 128 cores, SMT On, Boost On, Total Memory 1536GB (24x64GB DDR5 6400 MT/s [6000 MT/s]), BIOS 1.1, microcode 0xb002116, 2x Ethernet Controller E810-C for QSFP, 2x Ethernet Controller X710 for 10GBASE-T, 1x 1.7T Micron_7450_MTFDKBG1T9TFR, Ubuntu 22.04.4 LTS, 6.5.0-21-generic. Test by Intel as of December 2025.

Software Configuration

Redis Open Source 8.2.0, vector-db-benchmark 0.1.0, python 3 3.12.

HNSW TYPE=FLOAT16, M=32, EF_CONSTRUCTION=200. SVS-VAMANA TYPE=FLOAT16, GRAPH_MAX_DEGREE=64, CONSTRUCTION_WINDOW_SIZE=200, COMPRESSION=LeanVec4x8, REDUCE=256. Search top_k=10, recall/precision=0.95.

Faiss Vector Search

Tests by Intel as of December 2025. Your results may vary. Intel technologies may require enabled hardware, software, or service activation.

Hardware Configurations

Intel Xeon 6980P: 1-node, 2x Intel(R) Xeon(R) 6980P 128-Core Processor, 128 cores, HT On, Turbo On, Total Memory 1536GB (24x64GB DDR5 8800 MT/s [8800 MT/s]), BIOS BHSDCRB1.IPC.3544.O01.2503080923, microcode 0x10003d0, 2x Ethernet Controller E810-C for QSFP, 2x Ethernet Controller X710 for 10GBASE-T, 1x 3.5T SAMSUNG MZWLJ3T8HBLS-00007, Ubuntu 24.04 LTS, 6.8.0-78-generic. Test by Intel as of December 2025.

Intel Xeon 6972P: 1-node, 2x Intel(R) Xeon(R) 6972P 96-Core Processor, 96 cores, HT On, Turbo On, Total Memory 1536GB (24x64GB DDR5 8800 MT/s [8800 MT/s]), BIOS 1.1, microcode 0x010003d0, 2x Ethernet Controller X550, 1x 3.5T SAMSUNG MZWLJ3T8HBLS-00007, Ubuntu 24.04 LTS, 6.8.0-78-generic. Test by Intel as of December 2025.

AMD EPYC 9655: 1-node, 2x AMD EPYC 9655 96-Core Processor, 96 cores, SMT On, Boost On, Total Memory 1536GB (24x64GB DDR5 6400 MT/s [6000 MT/s]), BIOS 1.1, microcode 0xb002116, 2x Ethernet Controller X710 for 10GBASE-T, 1x Ethernet Controller E810-C for QSFP, 1x 1.7T Micron_7450_MTFDKCC1T9TFR, Ubuntu 24.04 LTS, 6.8.0-78-generic. Test by Intel as of December 2025.

AMD EPYC 9755: 1-node, 2x AMD EPYC 9755 128-Core Processor, 128 cores, SMT On, Boost On, Total Memory 1536GB (24x64GB DDR5 6400 MT/s [6000 MT/s]), BIOS 1.1, microcode 0xb002116, 2x Ethernet Controller E810-C for QSFP, 2x Ethernet Controller X710 for 10GBASE-T, 1x 1.7T Micron_7450_MTFDKBG1T9TFR, Ubuntu 24.04 LTS, 6.8.0-78-generic. Test by Intel as of December 2025.

AMD EPYC 9965: 1-node, 2x AMD EPYC 9965 192-Core Processor, 192 cores, SMT On, Boost On, Total Memory 1536GB (24x64GB DDR5 6400 MT/s [6000 MT/s]), BIOS 1.4, microcode 0xb101047, 2x MT2910 Family [ConnectX-7], 2x BCM57416 NetXtreme-E Dual-Media 10G RDMA Ethernet Controller, 1x 3.5T Micron_7450_MTFDKCB3T8TFR, Ubuntu 24.04 LTS, 6.8.0-78-generic. Test by Intel as of December 2025.

Software Configuration

Faiss 1.8.0, et_xmlfile 2.0.0, numpy 1.26.4, openpyxl 3.1.5, packaging 24.1, pandas 2.2.2, pip 24.2, psutil 7.0.0, python-dateutil 2.9.0, pytz 2024.1, setuptools 72.1.0, six 1.16.0, tzdata 2024.1, wheel 0.44.0.

HNSW M=32, efConstruction=200. Search k=10, recall=0.90.

Notices and Disclaimers

Performance varies by use, configuration, and other factors. Learn more on the Performance Index site. 

Performance results are based on testing as of dates shown in configurations and may not reflect all publicly available ​updates.  See backup for configuration details.  No product or component can be absolutely secure.

Your costs and results may vary.

Intel technologies may require enabled hardware, software, or service activation.

© Intel Corporation.  Intel, the Intel logo, and other Intel marks are trademarks of Intel Corporation or its subsidiaries.  Other names and brands may be claimed as the property of others.