Brief analysis: What exactly does a smart network card do?

What exactly is SmartNIC (Intelligent Network Card) and what can it do? SmartNIC based on DPU can not only realize the connection function of network card, but also realize the network traffic processing usually performed by CPU. SmartNIC can perform encryption/decryption, firewall, TCP/IP and HTTP processing. This article explains from 5 aspects why the use of SmartNIC has been increasing in recent years.

[[383255]]

NO.1 The Internet is facing unprecedented pressure

The rise of trends such as artificial intelligence, machine learning, cybersecurity, hyperscale architectures, and cloud services are placing unprecedented demands on networks, especially in terms of performance and uptime. These factors, combined with a surge in network usage due to remote work, are driving increases in network bandwidth, the number of users, and the amount of active network traffic—all of which increase computational complexity. The growth in network traffic and the increasing sophistication of attack vectors are placing tremendous pressure on the CPUs of the compute nodes of the server infrastructure.

No.2 What drives the need for network performance

Many services currently require lower latency to support real-time applications and services deployed at the edge of the network, such as video conferencing (Zoom, Microsoft Teams, etc.), 5G, and self-driving cars.

Other factors include the need to support traditional network services as well as cope with the growth of services such as 5G and IoT, which place high performance demands on the network.

No.3 Reduce CPU burden

SmartNICs enable the offloading of packet processing workloads from the CPU to FPGA-based SmartNICs. By offloading network processing workloads and tasks from the server CPU, SmartNICs improve server performance in cloud and private data centers. Driven by the growing volume of data center network traffic and computational complexity, the adoption of SmartNICs provides a processing architecture that provides computation for certain workloads through SmartNICs and offloads these workloads from general-purpose computing cores, thereby improving the efficiency of the overall solution.

No.4 SmartNICs can better support data plane workloads

In any virtualized network infrastructure, there are significant data plane network requirements inside the server. Network workloads are particularly expensive in terms of computation. Virtual switching alone can consume more than 90% of a server’s available CPU resources. Offloading network tasks can return these important resources to the application layer.

Cryptographic algorithms are one of the fastest changing aspects of data plane processing, and are also the most complex and computationally intensive. SmartNICs can offload this expensive task while being programmable, with the option to deploy new cryptographic algorithms in hardware simply by updating the SmartNIC software.

No.5 The truth about SmartNIC

Enterprises need to understand the alternatives and standards for SmartNIC deployment. The simplest is to add computing power to the next generation of x86 processors, but there is no denying the fact that Moore's Law no longer applies. Therefore, the idea of ​​just "throwing computing" will not work.

Compared with standard NICs, SmartNICs are competitive in price and power consumption, which eliminates the debate that SmartNICs are too expensive and consume too much power. SmartNICs are also not complicated, and the ease of deploying SmartNICs is similar to that of using standard NICs and software.

The use of SmartNICs greatly reduces the TCO of large-scale deployment of network services. By using SmartNICs to increase the computing power of each computing node, fewer servers are required to provide the same computing power compared to standard network cards, thereby reducing upfront costs, space, power and cooling requirements.

SmartNICs are future-proof, they are fully programmable, so investments in network deployments will stand the test of time. ASICs offer good performance, but they take a long time to design and are completely static, which is not conducive to future development. With the rapid changes in networks, protocols, encapsulation, and encryption algorithms, the ability to change hardware at the speed of software is critical to success.