Skip to main content
With LanceDB’s GPU-powered indexing you can build vector indexes for billions of rows in just a few hours—dramatically improving ingestion speed.
Internal tests show GPU indexing processing billions of vectors in under four hours.

​
Automatic GPU Indexing in LanceDB Enterprise

Automatic GPU indexing is currently available only in LanceDB Enterprise. Contact us to enable the feature.
Whenever you call create_index, Enterprise automatically selects GPU resources to build IVF or HNSW indexes. Indexing is asynchronous; call wait_for_index() to block until completion.

​
Manual GPU Indexing in LanceDB OSS

Use the Python SDK with PyTorch ≥ 2.0 to manually create IVF_PQ indexes on GPUs. GPU indexing currently requires the synchronous SDK. Specify the device via the accelerator parameter ("cuda" on Linux/NVIDIA, "mps" on Apple Silicon).

​
GPU Indexing on Linux

​
GPU Indexing on macOS (Apple Silicon)

​
Performance Considerations

  • GPU memory usage scales with num_partitions and vector dimension.
  • Ensure GPU memory comfortably exceeds the dataset you’re indexing.
  • Batch size is tuned automatically based on available GPU memory.
  • Larger batches further improve throughput.

​
Troubleshooting

If you encounter AssertionError: Torch not compiled with CUDA enabled, install a PyTorch build that includes CUDA support.