Apstra founder launches Aria to tackle AI networking performance

There are a growing number of networking startups working to tackle the challenges of AI networking. Aria Networks is hoping that its differentiated approach based on rich network telemetry will actually solve real-world problems for network operators.

Mansour Karam, founder and CEO of Aria Networks, is no stranger to the world of creating networking startups that disrupt the existing market. Karam spent nearly two decades building his networking credentials. He joined Arista in 2006 during the company’s early days and later spent several years at Big Switch Networks during the initial SDN wave. Karam also founded intent-driven networking vendor Apstra to reinvent the management and operational layer for networks; Apstra was acquired by Juniper Networks in 2020.

While SDN and intent-driven networking represented disruptive innovations in their time, AI is reshaping how networking architectures are designed and deployed today. AI networking represents more than an incremental shift. The backend Ethernet networks connecting GPU clusters have different requirements than traditional cloud infrastructure. 

“If the network isn’t performing optimally, then you can’t utilize your GPUs optimally,” Karam told Network World. “The GPUs are extremely expensive, and they’re the ones that are generating your revenue.”

From switch-centric to path-centric architecture

Market opportunity and AI requirements are the key reasons why Karam decided to start Aria Networks. 

He noted that overall data center networking has been growing at single-digit percentages for the past decade or two. AI networking is changing that trajectory dramatically. “When you have explosive growth like that, gaps for customers feel underserved, and really opportunities like this provide an opportunity for new entrants,” Karam said.

Aria’s technical approach differs from incumbent vendors in its focus on end-to-end path optimization rather than individual switch performance. Karam argues that traditional networking vendors think of themselves primarily as switch companies, with software efforts concentrated on switch operating systems rather than cluster-wide operational models.

“It’s no longer just about the switch itself. It’s really about the end-to-end path,” Karam explained. “When you look at these jobs being scheduled, it’s about the paths the traffic are going to take through the network, end-to-end that really matter.”

Telemetry at microsecond resolution

The company is targeting the backend Ethernet network that connects GPUs in AI clusters. It’s building with merchant silicon from Broadcom and using the open-source SONiC network operating system.

Aria’s core differentiation centers on extracting and acting on network telemetry that already exists in modern switching silicon but remains largely untapped outside of hyperscale environments. “In order to deliver on this performance, you need the data, you need telemetry, and this telemetry today exists,” Karam explained. “If you look at these ASICs from chips like Broadcom, they have tons of telemetry at the microsecond resolution.”

The challenge, according to Karam, is figuring out how to effectively extract, store, process and act on the telemetry data at scale, which is something that Aria is working on delivering as part of its platform.

Deterministic versus probabilistic network optimization

Aria is not only building networking gear for AI networks but also using AI to help improve networking.

Karam draws a clear distinction between rule-based deterministic approaches that have been used in the past and AI-driven probabilistic methods for network optimization. “When we built my previous company with Apstra, we didn’t have AI, and so we did everything very deterministically. Everything was rule based,” Karam said. 

That deterministic approach worked well in controlled environments but had limitations. Probabilistic AI-driven methods offer advantages for intuitive performance detection and dynamic reaction in complex scenarios. “When you’re trying to be more intuitive about detecting performance issues or reacting on the fly, then probabilistic approaches, AI, bring a unique advantage,” Karam explained.

But Karam emphasized that simply adding AI capabilities to existing architectures won’t deliver meaningful results. He’s critical of vendors that layer AI chatbots on top of legacy systems.

“A lot of what you see from a lot of vendors, when they say ‘we bring AI,’ what they’re going to do is kind of slap an AI chatbot on top of existing architecture,” Karam said. “But with AI, what we’ve seen again and again across every domain, is that in order for AI to become effective, you really need to specialize it for this domain, meaning building an architecture from the ground up that is optimized for AI.”

More to come as the company peels back the onion

Aria is taking a staged approach to revealing technical details about its platform as it continues to develop technology. The company has disclosed some of the foundational technologies it’s building on, including SONiC and microsecond telemetry, but is only incrementally revealing more details as it continues to build out.

“The power of AI is tremendous,” Karam said. “And having AI as a tool to bring to bear, combined with the data to really go and solve these problems and deliver on this performance optimization, the opportunity there is immense.”