Rockport Networks Names New Co

By Tiffany Trader

December 14, 2021

Rockport Networks, a new entrant in the HPC networking space with its switchless fabric offering, today announced the appointment of Marc Sultzbaugh to co-CEO. An active Rockport board member since December 2020, Sultzbaugh will lead the company alongside Rockport Networks Co-Founder and Co-CEO Doug Carwardine.

Sultzbaugh previously spent 20 years at HPC networking company Mellanox as a sales and marketing executive. In his new role at Rockport, he will be responsible for all aspects of the company's growth, including product management, marketing, sales, manufacturing and customer relationships. Carwardine will maintain the R&D and financial management of the company. The announcement comes a week after Rockport closed on a $48 million financing round, led by Northern Private Capital, bringing the company's total funding to just under $100 million.

The latest funding will go toward growing sales and marketing as Rockport transitions from an engineering firm to commercializing their product: the 300 Gbps switchless networking architecture launched in October. Instead of relying on a traditional switch architecture, Rockport distributes network functions to each node, which are direct-connected to each other over passive cabling. Although the company's technology is ‘switchless’ in the conventional sense of a hardware switch, the switching functionality is carried out at the node-level by Rockport NC1225 networking cards.

Sultzbaugh started out in engineering at Bell Labs some three decades ago but soon moved into business roles. He spent seven years at fabless chip company Brooktree, which went through a successful IPO. And he was with Mellanox from product launch all the way through the successful $6.9 billion handoff to Nvidia.

Sultzbaugh is confident he can lead a similar successful path at Rockport Networks, where he says the company is on a mission to help end users get the most utilization out of their networking. And that boils down to the way Rockport's switchless architecture handles congestion, according to the company.

"Our architecture is absolutely about performance," said Sultzbaugh, "But it just as importantly, if not more so, is about predictable performance. It doesn't matter what kind of workloads are in the fabric simultaneously, it doesn't matter how loaded your fabric is, you’re going to get very consistent runtimes."

Expanding on that point, he said, "Datacenter operators go through all kinds of gyrations to try to either limit the scale at which they run their applications or limit the kinds of traffic that are mixed in the network in order to overcome [congestion], or simply, they don't know why their applications are inconsistent or the performance isn't what they expect. And so the answer from traditional switch vendors is, let's just move you up the feeds and speeds. We’ll move you from 10 to 25, to 100, and now to 400 gigabits per second. But it really hasn't addressed the core issue on the architecture, and all they really accomplished is that they create congestion much faster. And worse, as you move up the theoretical performance of those switches, your cost is going up linearly. It's not like other technologies, where the costs tend to stay very neutral, if not go down."

The company says it has a number of engagements across multiple vertical markets in HPC and AI/ML, including within financial services, life sciences, energy, as well as the traditional supercomputing centers.

Early customers include the Texas Advanced Computing Center (TACC), which has installed the networking technology across 396 nodes of its Frontera system as an exploratory testbed, as well as DiRAC/Durham University, which is also trying out the networking gear. The high-performance networking group at Ohio State is also engaged with Rockport, lending its expertise with standards support.

Traditional HPC sites are Rockport's initial focus, but Sultzbaugh says there is very much an enterprise play. "There's more and more work being done in enterprises around what we call performance compute applications, which includes traditional HPC, design and simulation and modeling, and also new workloads around AI, machine learning, etc.," he tells HPCwire. "We will be 100 percent channel driven, but as you can imagine, with any kind of disruptive technology, we need to have high touch with those early adopters. And so we’ll be scaling out the salesforce, as well as scaling out our marketing efforts."

Education is a key part of many of the conversations Rockport is having.

"Anytime you have a transition or a disruption to the market, you have to educate customers, and part of our education for the industry is that we need to stop looking at the theoretical performance speeds and speeds that come from the traditional network vendors, and really look at and measure what's the impact in real-world applications when the network is loaded, when there's multiple different types of workloads being being run on that network, and that's really what matters to customers."

To assist with its performance validations and demonstrations, Rockport has lined up two Centers of Excellence (one at TACC and one at Durham), and plans to have several channel partners that will also validate the solution in their own testbeds. Dell, the system-maker behind the TACC Frontera system, is a key partner, and Sultzbaugh mentioned an early engagement with Lenovo. Rockport is also working on building additional ‘lighthouse customers’ across multiple vertical markets.

Asked about competitive head-to-head latency comparisons, Sultzbaugh indicated that on the surface, the metrics can be misleading. Specs on nanoseconds of latency are usually cited with no data transfer, in which case it's just theoretical capability, he said. "What's important is how you utilize that bandwidth and how you permit predictable performance across the fabric."

"On our roadmap today," he continued, "we’re supplying 300 gigabits per second per node. You have much higher bandwidth switches today, but they get congested because they’re built in terms of the spine leaf architecture. And that's fundamentally the issue. It's that you have to go through multiple switch hops, and you aggregate those switch hops. And so even though they’re capable of having very low latency, because you have this tail of latency that exists because packets get stuck in the network, the whole infrastructure is waiting. And so you move from capability of nanoseconds of latency to milliseconds of latency at an application level. We need to find a way to really measure in a benchmark that [indicates] what you’re really going to get in terms of performance. And that's our challenge."

"I think a lot of the customer base today already understands it and knows it," he added. "And then we have to educate the rest of them."

Rockport faces a competitive landscape dominated by InfiniBand – supplied by Nvidia's Mellanox division – and Ethernet, which is how HPE's Slingshot networking can technically be categorized, as well as the proprietary OmniPath fabric, now under the control of startup Cornelis Networks.

Mark Nossokoff, senior analyst with Hyperion Research, sees potential in the new switchless fabric offering. "I think there is definitely room in the networking space for some innovation in different approaches," he told HPCwire. "InfiniBand and Ethernet are marching along on their evolutionary cycles, then you have HPE come in with Slingshot, which is Ethernet-like. Now comes the switchless approach, which is unique and innovative, and I think it may very well have some merit [by] addressing the ability to have shortened predictable latencies and minimizing the congestion that occurs on heavily utilized networks."

Congestion issues are most pervasive on systems with lots of different types of heterogeneous jobs, said Nossokoff. "I believe the Rockport solution is aimed at large systems that have large numbers of users, and large numbers of heterogeneous types of jobs that are run on them. And by heterogenous, that's about the size of the jobs, the length of how long the jobs take to get done, and, and the utilization of the system.The greater the utilization of the system, the more you’re going to see the benefit of the Rockport solution," he said.

Shipping now, Rockport's networking technology comprises three elements (described below):