Couchbase Hyperscale Vector Index (HVI) Delivers Breakthrough Performance Against MongoDB in Independent Billion-Scale Benchmark

SAN JOSE, Calif. – October 23, 2025 – Couchbase, Inc., the developer data platform for critical applications in our AI world, today announced results from a Couchbase 벤치마크 test using an independent tool that demonstrates Couchbase’s Hyperscale Vector Index (HVI) delivers exceptional performance advantages over competitive solutions at billion-vector scale. The comprehensive testing, conducted using the industry-standard VectorDBBench methodology, shows Couchbase achieving over 700 queries per second (QPS) with sub-second latency and higher accuracy, which is 350 times faster than MongoDB Atlas under identical conditions.

The rigorous head-to-head comparison tested Couchbase and MongoDB against datasets of 100 million vectors with 768 dimensions, and 1 billion vectors with 128 dimensions, measuring QPS, response latency and recall accuracy across multiple retrieval configurations. The benchmark results reveal dramatic performance differences that become even more pronounced as applications require higher recall accuracy levels.

“When you’re building AI applications at scale, performance isn’t just a nice-to-have; it determines whether your application provides value that makes it worth adopting,” said BJ Schaknowski, CEO of Couchbase. “We see this with customers every day. They’re making infrastructure decisions right now that will either enable their AI initiatives or create bottlenecks they’ll struggle with for years. These benchmark results show what we already know: at billion-vector scale, architectural choices have massive implications for what you can actually deliver as GenAI-powered applications. Couchbase gives enterprises the performance and accuracy they need without the traditional trade-off between the two – all at lower total cost of ownership.”

Key Benchmark Findings

The testing demonstrates that Couchbase’s architecture enables organizations to simultaneously achieve both exceptional performance and high accuracy in vector search applications.

At optimized speed settings, Couchbase delivered 19,057 QPS with 28-millisecond latency at 66% recall accuracy, while MongoDB managed only 6 QPS at 57% recall with 62.6-second latency, representing a 3,176 times performance advantage for Couchbase. When configured for high-accuracy retrieval, Couchbase maintained 703 QPS with sub-second latency of 369 milliseconds at 93% recall accuracy. In contrast, MongoDB’s performance dropped to just 2 QPS with over 40 seconds of latency at 89% recall – giving Couchbase a 350 times throughput advantage while also delivering higher accuracy.

Technical Approach Drives Performance Advantages

Couchbase’s HVI leverages the DiskANN nearest-neighbor search algorithm with the Vamana directed graph construction algorithm, providing flexibility to operate in-memory and across partitioned disks for distributed processing and to maintain exceptional query performance while it scales. The architecture includes support for scalar quantization (SQ4), which reduces memory footprint while maintaining accuracy.

The benchmark testing used hardware-equivalent configurations on Amazon Web Services (AWS) infrastructure for both platforms, with four query nodes (64 cores, 128GB RAM each) and three data nodes (32 cores, 128GB RAM each) for each system. Both systems were tested against the same workloads with parameter configurations optimized for comparable breadth-of-search levels, where wider search settings drive higher recall accuracy.

The complete testing methodology and detailed results are available in the technical 보고 “Comparing Vector Search Capabilities of MongoDB and Couchbase by Benchmarking Using VectorDBBench.” The results show that Couchbase delivers more work, with less latency, and significantly lower costs per operation.

Implications for Enterprise AI Applications

Every second of delay in AI applications translates to frustrated users, abandoned transactions and lost revenue. For enterprises deploying AI at scale, database performance is not just a technical detail, it is the difference between AI that delivers business value and AI that becomes expensive shelfware. The benchmark results demonstrate that architectural choices have profound implications for application responsiveness, infrastructure costs and scalability.

가용성

Couchbase 8.0 with HVI capabilities, also 발표 this week, is now generally available for both self-managed and Capella-based deployments. Organizations can deploy across on-premises, cloud and edge environments to support their AI application requirements.

추가 리소스

• To learn more about Couchbase’s vector search capabilities and access the complete benchmark report, visit https://www.couchbase.com/products/vector-search/.
• To learn more about how Couchbase helps customers address their most critical AI challenges, click 여기.