Abstract
Advanced SmartNICs known as Data Processing Units (DPU) enable in-network data analytics, being equipped with standard processors and accelerators capable of doing custom computation on-NIC. These SmartNICs are advantageous because the host CPU can delegate simpler data analytics tasks, like filtering, to the NIC where the data first arrives. Only data needing further refinement must be passed on to the host CPU. Our benchmarks focus on NVIDIA's commercially available Bluefield-3 DPU. Similarity search, particularly Approximate Nearest Neighbor (ANN) search, is an important domain with wide usage across numerous applications. We explore ANN search on SmartNICs, providing insight into the performance of various ANN algorithms on the DPU when compared to a standard CPU device. We evaluate common ANN methods such as LSH, PQ, IVFPQ, and HNSW via the FAISS (Facebook AI Similarity Search) library on a host x86 CPU and a DPU using GloVe-200 vectors. In short, HNSW delivers the best latency-recall tradeoff on both platforms; LSH suffers the steepest recall degradation. The host/DPU performance gap varies by algorithm, reflecting the devices' different strengths.
Recommended Citation
S. Bhoria et al., "Performance Evaluation Of Approximate Nearest Neighbor Search On NVIDIA BlueField-3 DPU," Proceedings of the 35th ACM International Symposium on High Performance Parallel and Distributed Computing Hpdc 2026, pp. 613 - 614, Association for Computing Machinery, Jul 2026.
The definitive version is available at https://doi.org/10.1145/3806645.3818794
Department(s)
Computer Science
Publication Status
Free Access
Keywords and Phrases
Approximate Nearest Neighbor; BlueField-3 DPU; FAISS; GloVe; HNSW; LSH; Product Quantization; Recall@k; Vector Search
Document Type
Article - Conference proceedings
Document Version
Citation
File Type
text
Language(s)
English
Rights
© 2026 The Author(s), All rights reserved.
Creative Commons Licensing
This work is licensed under a Creative Commons Attribution 4.0 License.
Publication Date
13 Jul 2026
