Data centres are going through a transformation. Gradually, we will see a new type of equipment attached to servers in almost every data centre: smartNICs are here. They will be the enablers of converged data centres where common infrastructure tasks are offloaded from a host server to attached network interface cards (NIC).
Think of it this way: host servers will only run the applications that drive the actual return on investment. Operating the infrastructure itself will be the task of a fleet of smartNICs, effectively helping out the servers with the necessary networking, storage, and security tasks, among many others. In other words, data centres will do what they are actually supposed to do: run applications. The overhead will be offloaded, driving efficiency.
What is a smartNIC?
The obvious question is then: how can a simple interface card operate all the common infrastructure tasks? The answer is: a smartNIC is not a simple interface card. It is a new type of NIC that is equipped with more than just the conventional chips and circuitry that enables wired connectivity. These cards have various types of accelerators on them, speeding up commonly performed tasks on packets, and much more.
With more sophisticated but highly efficient board designs, smartNICs have evolved over recent years, and are now referred to as data/information processing units (DPU/IPU). The main advancement is the addition of an array of CPUs, besides the packet processor chip and the accelerators onboard. This makes it possible to expedite common data processing tasks in the interface card before actually delivering the useful information to the host server.
Consider a DPU/IPU as a separate computer, effectively assisting the operations of the host server it is attached to. This latest and most advanced form of smartNICs is the focus of this blog, and we will simply refer to a DPU/IPU unit as a smartNIC here.
How it all started
Hyperscalers were behind it all – there was a simple observation: around 30% of compute cores were busy dealing with infrastructure tasks. That is a big waste for an industry that gets its revenues from running virtualised customer workloads, relying on a fleet of servers powered by CPU cores.
From 2018 onwards, the industry saw changes at hyperscaler data centres, increasingly adopting smartNICs and dedicating CPUs on servers entirely to customer workloads. Hyperscalers achieved this transformation by working with smartNIC vendors through their partnership programs. Custom smartNIC types were designed, each of which is now dedicated to the use of a specific hyperscaler. Moreover, these cards are operated with proprietary software. Unsurprisingly, this has left any other industry other than hyperscalers without a cost-effective solution. There was a need for smartNIC solutions at smaller scale data centres operated by enterprises in various sectors.
Generally available SmartNICs
The good news is that various vendors now offer open and programmable smartNIC products. Software to drive smartNIC operations is also available. These vendors have one clear target: enterprises and private clouds. Customers in these markets can now source and adopt off-the-shelf smartNICS and transform their infrastructure.
Innovation for SmartNICs has been rapid, and products can now provide complex features driven by competition. However, there is fragmentation. Despite a number of different products now being available on the market, adoption has been rather slow. This is because there is still a need for common open source software and APIs, industry-standard hardware, and orchestration systems that use common interfaces.
Ultimately, smartNICs must be plug and play devices that are interoperable and easy to operate, despite being sourced from different vendors. That calls for standardisation and an ecosystem of open source developers that will drive the technology forward.
A growing ecosystem
SmartNIC and CPU vendors, application providers, operating system vendors and system integrators form a wide ecosystem that is collaborating to advance the use of smartNICs across sectors.
Open source projects facilitate common APIs and bring the community together as well. Common themes in community discussions include standardisation and open APIs, application offloading, artificial intelligence, operating systems, and latency reduction with acceleration, among many others.
There are now also events specifically targeting the smartNIC community, such as the SmartNICs Summit, which brings the community and industry players together annually, with hardware and software vendors showcasing their products and solutions. Canonical has been part of the event’s sponsors for two years.
SmartNICs will be common in next-generation data centres, enterprise networks, and edge clouds. In the area of conventional networking, the SmartNIC market is expected to overtake the standard NIC market by 2027, according to Dell’Oro Group’s market research.
A large proportion of this market will have some element of artificial intelligence (AI), especially generative AI. Top public clouds will continue to deploy servers with smartNICs and will represent the majority of the market for the next few years. The rest of the data centre market is expected to slowly adopt the new technology, but the pace can speed up if solutions are provided to make it easier to use and manage smartNICs for smaller organisations.
Canonical provides open source software solutions and helps data centre operators in their journey to adopt smartNICs. To learn more about Canonical’s work on smartNIC enablement in virtualisation platform solutions, check out our earlier blog. You can also read more about the smartNIC technology on our data centre networking series.
SmartNICs have a wide range of applications across industries. Contact us to learn more.