xPUBench Shows DPUs Use Less Power for 10 Gbps, GPUs Better for 100 Gbps

A new tool called xPUBench tested network hardware. DPUs use much less power at 10 Gbps compared to other options.
By Newsroom | Technology, Networking |

Benchmarking Environment Unveils CPU, GPU, and DPU Efficiencies

A new benchmarking environment, xPUBench, has been developed to systematically measure the performance and energy efficiency of network function packet processing across different hardware architectures. This research highlights a critical trade-off: Data Processing Units (DPUs) offer superior energy efficiency for lower network throughputs, consuming a mere 24 Watts to sustain 10 Gbps connections. Conversely, Graphics Processing Units (GPUs) prove more effective for higher throughput scenarios. The findings suggest that for network speeds of 100 Gbps and beyond, traditional Central Processing Units (CPUs) falter without dedicated acceleration, indicating a growing need for specialized hardware solutions like DPUs, often found in SmartNICs.

The introduction of xPUBench allows for the evaluation of various Network Function Virtualization (NFV) processing models, including CPU+GPU hybrid, DPU-only, and GPU-only approaches. This systematic approach is vital as network speeds rapidly increase, making packet processing on general-purpose CPUs a significant challenge. The research underscores that DPUs, with their embedded ARM or RISC-V cores, can entirely offload packet processing tasks, a capability crucial for modern data centers.

Read More: 250 European Nanotech Firms Struggle to Sell New Tech in 2024 Due to Slow Management

Specialized Hardware's Growing Role

The insights from xPUBench align with broader trends in modern data center architecture. NVIDIA's efforts in "accelerated networking" combine CPUs, GPUs, and DPUs into a unified fabric tailored for networking workloads. This approach optimizes not just processors but the entire network stack, including network interface cards (NICs), switches, and software.

The broader landscape of network and AI systems research, as evidenced by numerous papers presented at conferences like SIGCOMM 2025, indicates a strong focus on specialized hardware acceleration. Many of these works explore the efficient handling of massive data volumes and complex computational tasks, from AI inference to high-speed network functions.

Background: The Shifting Network Landscape

The demand for higher network speeds, driven by the proliferation of data-intensive applications and the rise of AI, has pushed the limits of conventional computing. General-purpose CPUs, while versatile, are increasingly struggling to keep pace with the throughput and latency requirements of modern networking. This has spurred the development and adoption of specialized hardware accelerators.

Read More: Open-Source Router Settings Lost After Reboot: Users Face Storage Errors

DPUs, for instance, are designed to offload network and storage tasks from the main CPU, thereby improving overall system efficiency and freeing up CPU resources for application processing. GPUs, traditionally known for graphics rendering, have also found significant application in high-performance computing and data processing due to their massive parallel processing capabilities. xPUBench aims to provide a clear, comparative understanding of how these different architectures perform under various network processing loads.

Frequently Asked Questions

Q: What is xPUBench and why was it made?
xPUBench is a new tool that tests how well network hardware works and how much power it uses. It was made because network speeds are getting faster and faster, and regular computer chips struggle to keep up.
Q: How much power do DPUs use for 10 Gbps speeds according to xPUBench?
According to xPUBench tests, DPUs use only 24 Watts to handle network speeds of 10 Gbps. This shows they are very good at saving energy for lower speeds.
Q: When are GPUs better than DPUs or CPUs for network speeds?
xPUBench found that GPUs work better for very fast network speeds, like 100 Gbps and even faster. For these high speeds, regular CPUs don't work well without special help.
Q: Why is specialized hardware like DPUs and GPUs important for future networks?
As networks get faster with more data and AI, specialized hardware like DPUs and GPUs are needed. They can handle the heavy work of processing data much better than normal computer chips, making systems faster and more efficient.
Q: What does xPUBench say about CPUs for fast networks?
xPUBench shows that regular CPUs have a hard time with network speeds of 100 Gbps and above. They need special help, like from DPUs or GPUs, to work well at these high speeds.