Reproduce “To GPU or Not to GPU: Vector Search in Relational Engines” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Vasilis Mageirakos, Joel André, Marko Kabić, Bowen Wu, Yannis Chronis, Gustavo Alonso
This research from ETH Zürich investigates the performance of integrating vector search with analytical SQL queries on heterogeneous CPU-GPU database systems, introducing a benchmark and an engine to evaluate various execution strategies. The study empirically demonstrates that non-data-owning vector indexes and unified-memory architectures enable efficient GPU acceleration for combined SQL+VS workloads, often with greater gains in relational operations than in vector search itself.
Key insights
- GPU acceleration for SQL+VS queries is beneficial, with relational operators often exhibiting proportionally larger performance improvements than vector search operators.
- Traditional "data-owning" vector indexes incur substantial data transfer bottlenecks to the GPU, primarily due to unoptimized data layouts and per-call overheads, making them inefficient unless already GPU-resident.
- Non-data-owning index designs, combined with host-residency optimizations on unified-memory architectures, enable efficient GPU acceleration for large-scale SQL+VS workloads without requiring full embedding residence in GPU memory.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.