Elasticsearch GPU HNSW indexing cuts 138M-vector build below 10 minutes
Elasticsearch describes GPU-accelerated HNSW construction using NVIDIA cuVS. On its Rally benchmark with 138M 1,024-dimensional MS MARCO vectors, eight RTX PRO 6000 GPUs cut index time from about one hour to under 10 minutes at a 95% recall target; p90 search latency under concurrent indexing was 6x lower.
🔗 www.elastic.co
#Elasticsearch #vector-search #HNSW
Post #322
202
