GPU Indexing

Build Vector Indexes Up to 4x Faster

GPU accelerates HNSW index construction as your data is indexed, built for sustained high write load, bulk loads, and re-embeds. Queries keep running on CPU.

Start Free View the Docs

Available on Qdrant Cloud AWS clusters with 16 GiB RAM or more.
No GPU infrastructure to operate, up to 4x faster builds based on Qdrant benchmarks.

Qdrant Cloud cluster settings showing the 4 vCPU plus GPU node option

Faster time to production.
Bulk loads of large-scale vector datasets complete significantly faster. Model swaps and full-corpus re-embeds benefit from accelerated indexing.
No infrastructure to manage.
Qdrant Cloud provisions and configures the GPU instances automatically. Enable it in the Console, and the GPU settings are managed for you.
Stop indexing at night.
Because the build runs on GPU, it does not take CPU from your live queries. Index and serve at the same time, instead of scheduling rebuilds for off-hours.

Enable GPU

GPU is a cluster-level setting. Select a node configuration with 16 GiB RAM or more in the Console, where GPU becomes available. All nodes in the cluster run on GPU-capable instances.

Indexing Accelerates

HNSW index construction runs on GPU across the cluster as new data is indexed. Query traffic is unaffected, because queries always run on CPU.

Accelerated HNSW Builds

GPU-accelerated index construction across all indexing workloads, from continuous high-volume ingestion to bulk loads and migrations.

Managed GPU Provisioning

Qdrant Cloud provisions and configures the GPU instances automatically.

Queries Stay on CPU

GPU accelerates the build; search runs on CPU, so there is no GPU cost at query time and no change to query behavior.

Quantization Compatibility

GPU accelerates the build, and quantization reduces the resident memory of the resulting index. Use them together.

NVIDIA T4 on AWS

Qdrant Cloud uses NVIDIA T4 instances for GPU-accelerated indexing on AWS. (In open source, GPU indexing is vendor-agnostic via Vulkan: NVIDIA, AMD, or Intel.)

Available on AWS

Supported in select AWS regions, with more rolling out over time.

Quantization

Memory complement to GPU builds

Set Up Quantization

Multivector per Point

Large-payload workloads that benefit from GPU builds

Explore Multivector Search

FAQs

Which clusters can use GPU indexing?

Any Qdrant Cloud cluster with 16 GiB RAM or more per node. Select a qualifying node configuration and enable GPU at the cluster level. Available on AWS today.

Where is GPU indexing available?

Available on AWS clusters today. GCP and Azure are on the roadmap. Need GCP or Azure with GPU-accelerated indexing? Contact us.

Do I need to manage the GPU instances?

No. Qdrant Cloud provisions and configures them automatically. Select a node configuration with 16 GiB RAM or more, and the GPU option appears in the package selector at that threshold. GPU is enabled at the cluster level; everything else is managed.

Does GPU indexing speed up my queries?

No, and that is by design. GPU accelerates HNSW index construction; queries always run on CPU. You get faster builds without paying for GPU on every query.

Which GPU does it use?

NVIDIA T4 instances on AWS Cloud. In open source, GPU indexing is vendor-agnostic through the Vulkan API and works with NVIDIA, AMD, or Intel GPUs.

How much faster is it?

Up to 4x faster HNSW index builds on dedicated GPUs, based on Qdrant benchmarks. The benefit is largest for sustained write load, bulk loads, and full-corpus re-embeds.

Can I turn GPU off when I don't need it?

Yes. GPU is a cluster-level setting you can disable, returning the cluster to standard instances. Useful for one-time migrations; for sustained write load, leave it on.

Does GPU indexing work with quantization?

Yes. The two complement each other: GPU accelerates the build, and quantization reduces the memory footprint of the resulting index.

Build Vector Indexes Up to 4x Faster

No Infrastructure to Operate, Faster Time to Production

Faster time to production.

No infrastructure to manage.

Stop indexing at night.

Flip One Switch. Index Faster as You Write.