Artificial intelligence (AI) and machine learning (ML) were unsurprising topics of popularity at Cisco's annual sales kickoff, in terms of how those technologies will impact both the industry at large and Cisco's networking business. While AI has existed in some form since the 1960s, its current iterations are finally able to change the definition of what is possible in terms of unlocking the value of data, Cisco VP Thomas Scheibe, who leads data center networking product management, told SDxCentral.

Enterprises in financial services, health care, automotive and government domains looking for a competitive advantage tend to make "good [network] customers" for Cisco, Scheibe said.

In these verticals, "it became very clear that [AI] was actually real and can be used relatively fast to unlock the value of the data," he explained. "The verticals that will move the fastest are the ones that have a lot of data sitting there. They don't have to generate it."

Over the next few years, however, he said he anticipates "this will be pretty much every customer environment."

So what does that mean for enterprise networks? Since AI/ML applications work by running a large cluster of specialized processes at a very high IO, each component of the network needs to communicate constantly and with low latency – a job ideal for Ethernet networks, Scheibe said.

As AI clusters grow from 256 accelerators today to somewhere around 4,000 accelerators per cluster in 2025 – "we can probably argue the number of GPUs you need on a cluster" – it's evident enterprises plan to invest in these specialized processors, he said. And given the need for accelerated communication between cluster nodes, these investments "will have a really important effect on the ROI."

To connect their GPUs, enterprises are looking at deploying Ethernet fabric, but "even more the question is does it have to be a dedicated Ethernet fabric or can it be a shared Ethernet fabric," Scheibe noted.

Specialized dedicated networks "are really not it," he said. "The cost of operating dedicated infrastructure makes very little sense, and we have seen this over the years," he explained.

Rather, enterprises want one network that can connect CPUs, GPUs and data processing units (DPUs) and is simple to operate, maintain and scale. Many existing capabilities of Ethernet fabric technology, like lossless connection and high performance, are a fitting answer to these needs. "There was actually a lot of work that went on before AI/ML became popular," and now Ethernet fabrics can prioritize specialized workloads ahead of others on a shared network, Scheibe said.

"Those technologies are available today, not just from Cisco, but across the board," he said, adding that Ethernet is steadily becoming the dominant network fabric technology.

The Ultra Ethernet Consortium, for example, is comprised of "a whole spread of different organizations" interested in advancing Ethernet technology. "This is going to be the future," he said. "If you look at the larger industry, it's very clear they're leaning in this direction."