Ilmu Komputer & AI editorial
Open AccessOA2026
When More Cores Hurts: The Vector Database Scaling Paradox in HPC
A large-scale evaluation of Qdrant, Milvus, and Weaviate on production supercomputers reveals that cloud-oriented vector databases fail to scale efficiently on HPC systems, with additional cores reducing query throughput by up to 30.67%.
Seth Ockerman; Song Young Oh; Amal Gueroudji; Rochana Chaturvedi; Philip Carns; Nicholas Chia; Matthieu Dorier; Robert Latham; Tanwi Mallick; Swan Perarnau; Robert Underwood; Kyle Chard; Ian Foster; Robert Ross; Shivaram Venkataraman· 2026· DOI 10.48550/arXiv.2606.08950
The core problem
Vector databases have become essential for managing high-dimensional embeddings in AI applications, but their design has primarily targeted cloud environments. Emerging scientific AI workloads—such as molecular search, meteorological trajectory detection, and literature-driven hypothesis generation—require efficient, scalable execution on high-performance computing (HPC) systems. This paper presents a large-scale evaluation of three state-of-the-art vector databases (Qdrant, Milvus, and Weaviate) on two production supercomputers, scaling to 256 distributed workers across 64 compute nodes. The authors aim to uncover how these systems perform under HPC conditions and identify fundamental mismatches between cloud-oriented designs and HPC architectures.
Innovation
The results reveal a scaling paradox: workload characteristics can limit latency reduction, and adding more cores can actually reduce query throughput by up to 30.67%. Furthermore, scaling from 16 to 256 workers (a 16x increase) only yields a 5.46x improvement in performance. This sublinear scaling indicates significant overhead and contention in the vector databases when deployed on HPC systems. The throughput degradation with additional cores is particularly striking, as it contradicts the typical expectation that more resources lead to better performance.
Mathematically, the scaling efficiency can be expressed as:
where is the speedup and is the number of workers. For the observed 5.46x speedup at 16x workers,
, indicating poor scaling efficiency.
Vector databases have become essential for managing high-dimensional embeddings in AI applications, but their design has primarily targeted cloud environments. Emerging scientific AI workloads—such as molecular search, meteorological trajectory detection, and literature-driven hypothesis generation—require efficient, scalable execution on high-performance computing (HPC) systems. This paper presents a large-scale evaluation of three state-of-the-art vector databases (Qdrant, Milvus, and Weaviate) on two production supercomputers, scaling to 256 distributed workers across 64 compute nodes. The authors aim to uncover how these systems perform under HPC conditions and identify fundamental mismatches between cloud-oriented designs and HPC architectures.
The evaluation was conducted on two production supercomputers, using up to 256 distributed workers across 64 compute nodes. Three vector databases—Qdrant, Milvus, and Weaviate—were tested under two representative workload patterns: mixed read/write and write-then-read. The workloads included popular benchmarks, multimodal embeddings, and a novel real-world scientific dataset. Performance metrics focused on latency reduction and query throughput as the number of workers scaled from 16 to 256. The experimental design allowed the authors to assess scaling efficiency and identify bottlenecks in HPC environments.
Why it matters
The scaling paradox exposes a fundamental mismatch between cloud-oriented vector database designs and HPC systems. Cloud environments typically assume homogeneous, loosely coupled resources with high-latency storage, while HPC systems feature tightly coupled nodes, high-speed interconnects, and parallel file systems. The authors argue that current vector databases are not optimized for HPC's communication patterns, leading to contention and inefficient resource utilization. This highlights the need for new, HPC-aware vector database designs that can leverage the unique characteristics of supercomputers. Potential directions include redesigning indexing and query processing to minimize cross-node communication and adapting to HPC storage hierarchies.
Who should read this
CS practitioners and researchers
Opening member content…