aws AWS Big Data Blog ·

Amazon OpenSearch Service Accelerates Vector Indexing with GPUs

blogaiawsgaengineer
feature

Amazon OpenSearch Service now supports GPU acceleration for building large-scale vector indexes, addressing a key bottleneck for generative AI applications. This feature, powered by NVIDIA cuVS, offloads compute-intensive tasks to GPUs, enabling faster and more cost-efficient construction of indexes containing billions of vectors without impacting query performance. The decoupled architecture separates indexing from serving, with automatic scaling and pay-per-use billing for GPU resources. This is particularly beneficial for use cases involving rapid reindexing, absorbing write bursts, and optimizing cluster sizing.

  • GPU-accelerated vector indexing for large-scale search
  • Decoupled architecture for indexing and serving
  • Automated GPU scaling and resource management
  • Pay-per-use billing for GPU resources
  • Automatic activation and operational considerations
Features (1)
  • GPU-accelerated vector indexing for large-scale search

    Amazon OpenSearch Service and Amazon OpenSearch Serverless now offer GPU acceleration for k-NN (vector) indexing, enabling efficient scaling to billions of vectors. This capability offloads index building to specialized GPU workers, utilizing NVIDIA cuVS, while CPU nodes continue to serve search queries, resulting in faster and more cost-effective index construction.

Enhancements (3)
  • Decoupled architecture for indexing and serving

    The service employs a decoupled architecture where vector index building is offloaded to dedicated GPU workers. This separation allows existing CPU data nodes to continue handling ingestion, search, and non-vector workloads without interference, improving overall system performance and efficiency.

  • Automated GPU scaling and resource management

    GPU workers automatically scale up and down based on the number of pending index build jobs, ensuring resources are available for peak loads and scaling back to zero when idle. The service also automatically selects appropriate GPU instance types for each build job, eliminating the need for manual capacity planning.

  • Pay-per-use billing for GPU resources

    Customers are charged only when GPUs are actively building indexes. Costs are measured in OpenSearch Compute Units (OCUs) and apply only when segments reach the activation threshold, meaning there are no standing GPU infrastructure costs. Billing scales directly with indexing activity.

Notes (1)
  • Automatic activation and operational considerations

    GPU acceleration activates automatically upon opting in via the Vector Acceleration option on OpenSearch Service domains or is enabled by default for NextGen vector search collections in OpenSearch Serverless. The service automatically routes indexing operations to GPU or CPU based on segment size, with configurable windows for GPU activation.

Read the original announcement →

https://aws.amazon.com/blogs/big-data/how-gpu-acceleration-builds-billion-scale-vector-indexes-on-amazon-opensearch-service/

Related releases