Specialized Data Processor (DPU) Performance Benchmark Methodology and Implementation

The document presents the first industry whitepaper on DPU (Data Process Unit) performance benchmarking, jointly authored by the Institute of Computing Technology, Chinese Academy of Sciences, and China Yushu. It aims to establish a fair, open, comprehensive, and objective DPU evaluation framework.

To address the challenge of comparing heterogeneous DPU products, the authors propose a set of "POC" (Proof‑of‑Concept) principles: Practicality (real‑world relevance), Orientation (targeted performance focus), and Comprehensiveness (covering network, storage, and compute aspects, with combined tests such as network+storage and network+compute).

Four evaluation dimensions are defined:

Network dimension : measures packet processing capabilities (throughput in PPS/BPS), latency, observability, and deterministic QoS features.

Storage dimension : assesses I/O performance for NVMe‑oF (IOPS, MBps), latency, and QoS guarantees such as bandwidth and rate limiting.

Compute dimension : evaluates DPU’s ability to accelerate heterogeneous workloads (CPU, GPU, FPGA, AI accelerators) and to offload services like service‑mesh sidecars.

Security dimension : examines hardware‑root‑of‑trust, encryption/decryption throughput, policy‑size scalability, and concurrent session handling.

The whitepaper details how to construct test systems for each dimension, design representative workloads, and select appropriate metrics, providing a practical reference for DPU users and a guidance for future DPU standardization.

Readers can download the full whitepaper via the provided links.