Qdrant, the developer of a high-performance open-source vector database, today introduced its graphics processing unit accelerated vector indexing capability that will make scaling up artificial intelligence applications easier.

A vector index is a data structure used in information retrieval to store and pull out high-dimensional data, or a dataset with many features or variables, more efficiently. It enables fast similarity search and nearest-neighbor queries, which are fundamental for generative AI and large language models. Essentially, it allows AI models to “understand” context among words, sentences and concepts to generate realistic and interactive text, images, video and audio from a variety of data sources.

Andre Zayarni, co-founder and chief executive of Qdrant, told SiliconANGLE in an interview that the new GPU-accelerated feature optimizes the company’s Hierarchical Navigable Small World vector index building, one of the world’s most resource-intensive steps in the vector search pipeline.

“It’s pretty fast and flexible, but it takes the time to build this index,” Zayarni said. “It’s also very performant and suitable for near-real-time use cases.”

Depending on the use case, a vector index may be rebuilt as soon as the data changes, such as a piece of information. For example, if there is a database with a great deal of data, such as a content management system with a semantic search or AI on top of it, the index may need to rebuilt when users alter or adjust the information. For smaller cases, a few hundred vectors this is not a problem, but when updates scale up to millions or hundreds of millions of updates, the sheer number can become a bottleneck.

According to Zayarni, with the new feature the company can accelerate index building up to 10 times by using GPUs instead of central processing units.

The implementation uses the low-level Vukan application programming interface, which allows it to run on varied hardware, including GPUs from Nvidia Corp., AMD Corp. and Intel Corp. and even Apple Inc. silicon. This gives customers the freedom to choose the cloud vendor and hardware that best fits their needs.

It also helps maintain feature parity across platforms, Zayarni explained. The solution ensures that all the features of the core product remain the same and are supported between CPU and GPU irrespective of the hardware. As a result, customers do not need to compromise on any essential features or capabilities when moving between CPU or GPU acceleration.

In line with Qdrant’s open ethos, the hardware independence of GPU acceleration means customers can avoid vendor lock-in. It also enables local development and testing of vector databases on local machines including laptops with integrated graphics. This allows customers to develop and test their applications more easily before deploying them on more powerful production hardware.

Image: SiliconANGLE/Microsoft Designer