As artificial intelligence (AI) and machine learning see broader application, traditional network architectures are beginning to choke performance, said Rockport Networks CTO Matthew Williams, who argues it’s time to ditch the switch in these environments.

Founded in 2012, the Ottawa, Canada-based networking vendor champions a distributed switching model. The technology integrates switching functionality into every server in the cluster, doing away with traditional leaf-spine topologies in favor of a mesh network. This approach, the company claims, improves operational efficiency and eliminates the long tail latency challenges that plague high-performance compute (HPC) workloads.

“When you look at the overall architecture of the network, it's very much focused on optimizing for north-south traffic,” Williams said, adding that applying this approach to HPC applications is at best challenging if not impractical.

Because HPC workloads are spread across the cluster, any new instructions or data has to be distributed to the nodes before any work can be done.

It’s sort of like a party that can’t start until the last guest arrives, and that guest just got caught in traffic, Williams explained. Specifically, the guest, or in this case the data, got held up by network congestion on the switch.

“Every time a message is stalled, a whole set of resources is wasted. They’re idle,” Williams said, adding that this problem only gets worse the larger the cluster gets.

To solve this problem, Rockport cut the switch out entirely. “Building a better switch doesn’t solve the architectural challenges,” Williams said. “We do not have traditional, centralized switches at all in our architecture.”

Instead, switching is handled by two components that interface directly with each compute node. The first is Rockport’s NC1225 network card, which consists of a NIC ASIC, an FPGA running the company’s network operating system, and a set of optical transceivers.

Each card features a single interface that breaks out into 12 fiber pairs capable of 25 Gb/s of throughput each for an aggregate of 300 Gb/s per node.

The second component is Rockport’s Shuffle, a passive breakout box that turns the single interface on the network card into 12 optical ports, enabling each node to interface directly with a dozen others.

“Think of it like the six degrees of Kevin Bacon. Everyone has 12 friends and all of a sudden you can get anywhere you want in a very small number of hops,” Williams said. “It’s very different from a how traffic moves point of view, but it still looks like standard Ethernet.”

So while the topology may be radically different, customers don’t have to learn an exotic new protocol to take advantage of the technology. It looks and functions like any other Ethernet network.

In real-world, multi-workload environments, Rockport claims customers are seeing a 28% reduction in workload completion times under load on average.

“A lot of networking vendors will ask you to test their technology in a dedicated environment for getting baseline, best-possible results,” Willaims said. “But really, what people care about is in my multi-workload environment, when the network is supporting lots of workloads concurrently, when the network is always busy, when I have noisy-neighbor traffic, how does it work? That’s where we really shine.”

In fact, the technology has already been deployed in the Texas Advanced Computing Center’s Frontera supercomputer.

The company this week snagged $48 million in a funding round led by Northern Private Capital, bringing the company's total raised to just over $100 million.

In addition to expanding its engineering focus, the new funding will enable Rockport to accelerate its go-to-market strategy, newly appointed Co-CEO Marc Sultzbaugh told SDxCentral.

Prior to joining Rockport, Sultzbaugh spent nearly 20 years at Mellanox, where he most recently served as SVP of sales and marketing.

After departing Mellanox following the successful acquisition by Nvidia in early 2020, Sultzbaugh joined Rockport’s board of directors, and this week joined the company’s C-suite by CEO and co-founder Doug Carwardine’s side.

“I was introduced to Doug and the Rockport team through the investors and really initially fell in love with the product and the technology and the way that they were trying to disrupt a very large market,” Sultzbaugh said.

Moving forward, Sultzbaugh will be responsible for driving Rockport’s growth including technology and product management, marketing, sales, manufacturing, and customer relationships. This includes taking advantage of trends surrounding machine learning, AI, and deep learning to break into the lucrative enterprise space.

“There are, if you think about it, tons of modeling, simulation, and fluid dynamics [applications] that are being run in enterprises today,” Sultzbaugh said, adding that there are a substantial number of enterprises which have whole divisions devoted to performant applications.

Editors note: This story has been updated to reflect Rockport's headquarters in Ottawa, Canada. We regret the error.