Enterprises are moving from artificial intelligence pilots to large-scale production, putting intense pressure on data center networks. Supporting graphics processing unit-intensive training and scalable inferencing requires raw bandwidth, intelligent automation and secure, predictable infrastructure.

To meet those demands, Juniper Networks Inc. is combining its Ethernet research and validated blueprints with Hewlett Packard Enterprise Co.’s expertise in rack-scale servers and supercomputing. Together, they aim to deliver pre-tested systems that create a stronger foundation for AI-ready networking, according to Praful Lalchandani (pictured, middle), vice president of product – data center platforms and AI solutions – at Juniper.

“AI training is a widely distributed computing problem, but so also is inferencing,” he said. “Most likely for enterprises, they’re not going to be building large language models. Most likely, they’re going to be consumers of those models. My point is that training or inferencing is highly distributed, and that means the network needs to be highly performant.”

Lalchandani, along with Jon Green (right), senior sales engineer at Juniper, and Bharath Ramesh (left), head of product, AI factory solutions, at HPE, spoke with theCUBE’s Bob Laliberte at The Networking for AI Summit event, during an exclusive broadcast on theCUBE, SiliconANGLE Media’s livestreaming studio. They discussed how this integrated approach addresses performance, intelligent automation, security and ease of deployment, giving enterprises the confidence to scale their AI ambitions. (* Disclosure below.)

Security and scalability by design for AI and intelligent automation

Performance is critical, but security can’t be an afterthought. For enterprises, the importance of integrating protection from day one can’t be overstated. Whether AI services are consumed as software-as-a-service or infrastructure-as-a-service, multi-tenancy introduces risk, according to Green. The safest path is on-premises infrastructure, where enterprises maintain full control.

“AI is different than standard enterprise systems in a really important way, which is observability of the data itself,” he said. “You think about an AI model that’s been trained, that now I’m going to do inferencing through because of things like tokenization. The data — the programming of this — is going to be a really large vector space. It’s going to be a bunch of numbers, and there’s no real way to know what those numbers actually mean.”

AI introduces new security risks, particularly around data integrity. Unlike traditional applications, where anomalies can be traced to code or records, AI outputs stem from opaque vector spaces, making adversarial attacks and data poisoning especially dangerous, according to Green.

“The most important thing that we can do is to actually get control of that infrastructure ourselves,” Green said. “Behind that, the standard enterprise security controls are still effective at doing what they do, but I want to see what’s happening inside that cluster.”

For inference at scale, HPE and Juniper are pioneering “AI-optimized Ethernet.” Simply throwing bandwidth at the problem is not enough — Ethernet was designed as a best-effort protocol, not a congestion-free one. However, through techniques such as remote direct memory access-aware load balancing and Data Center Quantized Congestion Notification, Juniper has boosted Ethernet’s predictability and performance, making it viable for intelligent automation workloads once dominated by InfiniBand, Ramesh added.

“We have distilled that down into reference architectures that take you through not just the hardware building blocks, but the platform building blocks, the software and service building blocks that you need to build and operate this at a large scale,” he said. “You can start at 80% of the solution and iterate from there, instead of starting from 0% every time. Because ultimately, we’re believers that the quicker we get customers onboarded and successful with large-scale AI, the more successful they’re going to be and then the more successful we’re going to be as a result.”

Here’s the complete video interview, part of SiliconANGLE’s and theCUBE’s coverage of The Networking for AI Summit event:

(* Disclosure: TheCUBE is a paid media partner for The Networking for AI Summit event. Neither Juniper Networks Inc., the sponsor of theCUBE’s event coverage, nor other sponsors have editorial control over content on theCUBE or SiliconANGLE.)

Photo: SiliconANGLE