Americas

  • United States

Enterprise adoption of smartNICs faces challenges

News Analysis
Aug 15, 20223 mins
Networking

Dell'Oro: Cost and expertise are two obstacles to enterprise use of smartNICs, but the NIC makers are working to eliminate them.

network speed

Growth in hyperscaler data centers and processor-intensive enterprise workloads, such as high-performance computing (HPC) and AI, is set to drive broadscale adoption of SmartNICs.

The Ethernet controller and adapter market will grow from $3.2 billion in 2021 to $5 billion in 2026, with smartNICs contributing significantly to the growth, according to Dell’Oro Group. In addition, server network connectivity will transition to higher speeds, according to the research firm, which predicts 100 Gbps and higher-speed ports will account for 44% of the shipments in five years.

The standard issue 25Gbit NIC port is more than enough for enterprises doing basic networking tasks such as database and Web serving, but Dell’Oro Group sees an opportunity at the higher end for smartNICs, which better manage network traffic and take the networking load off CPUs.

“SmartNICs will displace traditional NICs for most of the hyperscale cloud infrastructure for general-purpose and high-end workloads such as accelerated computing,” says Baron Fung, research director at Dell’Oro Group.

However, smartNIC vendors such as Intel, AMD, Nvidia, and Marvell Technologies will need to address the costs of ownership and implementation in order to promote broader smartNIC adoption beyond hyperscalers, Fung says.

A smartNIC will be three to 10 times the cost of a regular NIC. A standard 25Gbit NIC is usually $60 or $70 a port, whereas smartNICs range from $300 to $500. “So you’re looking at a pretty substantial multiple increase,” Fung says.

There’s also the challenge of programming a smartNIC. Hyperscalers like Google have a massive infrastructure and strong engineering resources, so they can really maximize the use of a programmable NIC. The average enterprise won’t have the same resources. “Because they may not be able to fully optimize the part, it will be difficult to justify the higher price point and higher power consumption that smartNICs are associated with,” Fung says.

He says some vendors are trying to support to customers through SDKs with specific smartNIC application support for building more turnkey solutions, which would be required to see more smartNIC adoption elsewhere.

But that will take time. Right now, 25Gbit Ethernet is standard and will be for at least the next three to five years. From there, Fung predicts the next upgrade speed will be 100Gbit as smartNICs are shipped with high-end storage devices, such as NVMe over fabric arrays and AI systems. Every Nvidia DGX AI system is equipped with its Bluefield DPU, he says.

It seems a bit odd that four semiconductors giants would battle it out for a relatively modest enterprise market, but Fung says there are two reasons:

“This is a growth market, with 25% CAGR over the next few years, which is faster than the overall industry,” Fung says. “It also creates a complete solution. Nvidia has talked the most about the three-legged stool of CPU, GPU, and DPU. This lets all of them go to market with a complete portfolio of products.”