GPUs are used for embedding generation. If you use semantic search, embeddings are generated for documents at indexing time and also for the search queries are search time, and these will benefit from having a GPU, when using a built-in model. The actual nearest neighbor vector search does not require a GPU