Intel’s smartNICs probably aren’t for you (yet) says Intel

'Right now, the focus is on the super-high volume'

Unless you happen to be running a cloud or hyperscale datacenter, Intel’s infrastructure processing units (IPU) probably aren’t for you, at least not yet.

Accelerators, like Intel’s Mount Evans chips, represent an extreme point on a spectrum. The IPU was built in collaboration with Google for their datacenters, Nick McKeown, who leads Intel’s Network and Edge Group, told The Register.

But rest assured, McKeown believes IPUs will find their way down market before long. “It’ll become a class of device that I think we will look back and say, ‘it's as important as the CPU, memory, and other accelerators,’” he said.

SmartNICs, IPUs, data processing units (DPUs), or whatever you or the vendors’ marketing department wants to call them have gained steady traction over the past few years as a means to extract value out of systems by offloading input/output (I/O)-intensive operations. Networking, storage, and security applications are common examples.

IPUs offer a way to physically separate the infrastructure layer from the components running customer code

The IPU essentially functions as a co-processor which handles the I/O operations so that the CPU can focus on doing work. This kind of functionality has made these accelerators popular in cloud and hyperscale datacenters.

“For the cloud providers, IPUs offer a way to physically separate the infrastructure layer from the components running customer code,” McKeown said. “The PCI operates almost as the DMZ that protects the user code and the client code that is rented out.”

In other words, these devices allow the cloud providers to manage and orchestrate the resources that make up their services, without taking away from or compromising the security of the revenue-generating CPU cycles sold to customers.

“Right now, the focus is on the super-high volume. Because, if you’re going to co-develop, you’re going to develop with someone who’s going to buy a lot of it,” he said.

Because of this, Mount Evans is bleeding edge out of necessity. But, as the technology finds applications down market — and McKeown believes it will — he expects a spectrum of devices with different core counts and performance targets will emerge from a variety of vendors.

A lot of what makes up Intel’s IPUs can be boiled down to suit less-demanding environments, he explained. So at one extreme is the IPU, and the other is a fixed-function NIC. “There are no major unknowns in that space in between.”

When it comes to software, there’s more work to be done. “I think what is perhaps even more important, dare I say, than the device itself, is the software ecosystem that grows up in support of it,” McKeown said.

SONiC for smartNICs

That’s where the Linux Foundation’s Open Programmable Infrastructure (OPI) project comes in.

Launched last month in collaboration with Intel, Nvidia, Marvell, and other smartNIC heavyweights, the project promises to build an open ecosystem of common software frameworks that can run on any DPU, IPU, or smartNIC.

The organization aims to address a common problem with smartNICs today: most rely on proprietary software stacks, RedHat’s Yan Fisher previously told The Register. "That was the fundamental thing that's slowed down the adoption" of smartNICs.

And while this hasn’t stopped hyperscalers or cloud providers, who are willing to dedicate the resources to building custom software around these devices’ unique capabilities, it poses a challenge for their application in other markets.

McKeown compares OPI’s efforts in many ways to how SONiC kickstarted the switching industry. The open source network operating system, developed by Microsoft, brought networking vendors together around a common platform and ultimately spurred innovation, he said. “What we're trying to do as a community is to create that same magical sort of momentum for the IPU class of product.”

However, more than simply fostering the development of novel smartNIC applications, the project also addresses another challenge: interoperability from one vendor to the next or from generation to generation.

“When you’re running an application on a NIC, you’re not really running an application, it’s a bunch of drivers,” McKeown said. “But now, when you talk about an IPU, you’re running whole microservices infrastructure; you might be running OVS… . You need to know that you can lift that up and move that forward to the next generation without starting over.”

As the smartNIC adoption grows, McKeown says the risk is some vendors may be tempted to lock down their software and hardware behind a walled garden. This kind of behavior has “held back the equipment industry; it's held back the switching industry; it’s held back many things in the past.”

If not the cloud, where next?

While applications for IPUs outside of hyperscale or cloud datacenters may be few and far between, that’s not to say there aren’t any in the near term.

“A lot of interest is towards the telco edge, where you can use such a class of device to take those cloud ideas and bring them to the edge,” McKeown said.

Whether the cloud providers are trying to extend their own infrastructure to the edge, or the telcos are trying to build out their own edge-compute offering, there’s still a need for highly-programmable, mesh networking, for which IPUs are well suited, he explained.

So while IPUs, DPUs, smartNICs, or whatever the marketing folks come up with next may not be a staple of servers just yet, that won’t always be the case. ®