High-performance computing is brimming with several developments: from the artificial intelligence infusion to innovations in exascale and quantum computing.

Broadcom Inc., a stalwart in ethernet networking, is at the forefront of pushing the boundaries in AI and machine learning. The company has introduced the concept of “ultra ethernet,” a term that encapsulates the integration of intelligence into networking infrastructure, according to Hemal Shah (pictured, left), distinguished engineer and system, software and standards architect at Broadcom.

“At Broadcom, we have been doing ethernet networking for decades now,” he said. “AI/ML is pushing the envelope, bringing really large-scale clustering, asking for high bandwidth. And what we have is, both in terms of our [network interface cards] and switches, we have our solution today. We are innovating with other things around ultra ethernet, where we are putting more intelligence in network infrastructure in the NIC to make AI/ML solution from networking standpoint the best.”

Shah and James Wynia (right), director of product management for networking, Enterprise Solutions Group, at Dell Technologies Inc., spoke with theCUBE industry analysts Lisa Martin and David Nicholson at SC23, during an exclusive broadcast on theCUBE, SiliconANGLE Media’s livestreaming studio. They discussed the partnership between Dell and Broadcom, as well as the latest advancements, current challenges and future landscape of AI/ML networking. (* Disclosure below.)

Enabling hyperscale capability

Broadcom is a large-scale supplier of enterprise networking and connectivity hardware. Many of the company’s offerings, particularly the AI-focused ones, are underpinned by Dell solutions. The collaboration is pivotal, allowing both companies to stay at the forefront of hyperscaler networking solutions. The Tomahawk line and Trident line are primary examples, demonstrating Dell’s commitment to providing top-tier solutions for both hyperscalers and enterprises, according to Wynia.

“We’ve been partnering with [Broadcom] for literally decades, and it’s always exciting to see what they’re brewing up,” Wynia said. “We’ve been supporting the Tomahawk line, the top of the food chain for hyperscaler networking solutions and the Trident line, the top of the line for enterprise solutions. We have switches in those lines, and we continue to work with them.”

There’s also Broadcom’s Jericho3-AI chip, which is designed to connect supercomputers and drive complex network flows. Alongside the Tomahawk 5, it illustrates the capability for multistage switching fabric in AI/ML applications, according to Shah.

The ultra ethernet concept is nuanced and separate from the conventional form, with the addition of features such as multi-pathing and selective retransmissions. The importance of these enhancements lies in addressing congestion points and optimizing data flow in large-scale AI/ML environments, Shah added.

“With ultra ethernet, what we are doing is keeping the same ethernet ecosystem infrastructure but adding things like multi-pathing, adaptive routing and congestion control mechanism with scale to a large number of nodes,” he explained. “Also, addressing some of the RoCE transport deficiencies means RDMA goes to the next level by having selective retransmissions.”

Graphics processing units also play a critical role in the new AI-focused networking paradigm, further emphasizing the need for reliable line-rate connectivity, a feat achieved through ultra ethernet, according to Wynia.

“That’s where these large language model solutions come in where they send all this data, like ChatGPT sending the whole internet of data through in a matter of days, [and] being able to learn specific [information] to medical, traffic or air control [data],” he said.

Here’s the complete video interview, part of SiliconANGLE’s and theCUBE’s coverage of SC23:

(* Disclosure: TheCUBE is a paid media partner for SC23. Neither Dell Technologies Inc., the main sponsor of theCUBE’s event coverage, nor other sponsors have editorial control over content on theCUBE or SiliconANGLE.)

Photo: SiliconANGLE