Beyond the Data Center: High-Performance Networks for AI - Cisco Blogs

Skyler: Hey, welcome to Line Rate — the show about datacenter fabric and A I high-speed interconnect. I'm Skyler Monroe, and if you live and breathe the chips and standards moving data from appliance to appliance, rack to rack, and datacenter to datacenter, this is your show. Big thanks to our sponsors Ada, who helps organizations integrate A I into their real-world workflows, and Zen Semiconductor, an umbrella company with a portfolio of businesses across the silicon industry. Couldn't do this without them. Alright, today we're getting into some meaty stuff — Cisco's take on where high-performance networking goes beyond the datacenter walls, what Futuriom sees as the top trends connecting A I clusters all the way to the edge, Broadcom's framing of scale-out networking for A I infrastructure, Nokia actually running end-to-end Ultra Ethernet tests across a live switch family, and a great IEEE Spectrum piece on optical networking and what lasers are doing for A I datacenters. Let's get into it. Skyler: Alright, first up — Cisco dropped a blog post titled "Beyond the Data Center: High-Performance Networks for A I," and look, I know blog posts from big vendors can sometimes feel like marketing wrapped in a technical bow, but there's actually some real substance worth unpacking here. The core argument Cisco is making is that the network requirements for A I workloads don't stop at the datacenter edge — and that's a framing shift worth paying attention to. For a long time, the conversation around A I fabric has been hyper-focused on what happens inside a single cluster. How do you move tensors between GPUs fast enough? How do you minimize all-reduce latency? And that's all valid and critically important. But the reality is that as A I systems scale — and they are scaling aggressively — you start having training jobs that span multiple datacenter facilities, inference pipelines that need to pull from distributed model shards, and data pipelines that have to move enormous datasets from wherever they live to wherever the compute lives. That's not a within-cluster problem anymore. That's a wide-area networking problem with high-performance constraints that most WAN infrastructure wasn't designed to handle. What Cisco is really getting at is that the fabric conversation needs to expand. The latency budgets are tighter, the bandwidth demands are higher, and the traffic patterns are completely different from traditional enterprise or even hyperscaler WAN traffic. A I workloads are bursty in very specific ways — you get these massive synchronization events during training, collective communication patterns that look nothing like your average web traffic. And when you try to push that over infrastructure that was designed for human-scale applications, things get ugly fast. Now Cisco has obvious commercial interest in making this argument — they sell networking gear from the campus edge all the way into the core — but the underlying point is technically sound. The A I networking problem is metastasizing beyond the rack and the pod and the datacenter. And whoever figures out how to deliver consistent, low-latency, high-bandwidth connectivity at that wider geographic scale is going to be sitting on a very interesting business. This is also where standards like UEC start to matter outside of their original cluster context — because you need interoperability not just within a vendor's stack but across the boundaries of different facilities and different operators. So keep an eye on how Cisco positions their next-gen WAN and optical portfolio in relation to A I. This blog feels like early narrative-setting for something bigger coming down the pipe. Skyler: Next, Futuriom published a piece on the top trends in networking for A I — "Connectivity from A I Clusters to the Edge" — and it's a good high-level survey if you want to understand where analyst thinking is coalescing right now. The framing overlaps a bit with what Cisco was doing, but Futuriom comes at it from more of a market trends angle rather than a vendor pitch, so it's worth looking at separately. A few things stand out. First, there's a lot of attention on the distinction between scale-up and scale-out networking, and the recognition that you need different approaches for each. Scale-up is your NVLink territory — tightly coupled, proprietary, extremely high bandwidth, designed for GPU-to-GPU communication within a node or a small cluster. Scale-out is where Ethernet and InfiniBand compete — connecting pods, connecting racks, eventually connecting facilities. And the interesting tension right now is that the line between those two domains is getting blurry. As GPU clusters get larger and NVLink domains hit their physical limits, you need scale-out fabric that can operate with latency and bandwidth characteristics that start to approach scale-up territory. That's a hard problem. Second, Futuriom flags the edge connectivity piece — and this is interesting because it's not just about inference at the edge, it's about the whole data pipeline. Training data has to come from somewhere. Model updates have to propagate somewhere. And as enterprises start deploying A I in more distributed ways, the network path from the training cluster to the inference endpoint becomes a first-class infrastructure concern. Third — and this is where U A Link and UEC both get mentioned — there's growing acknowledgment that the proprietary interconnect lock-in from dominant players is creating real buying risk for hyperscalers and enterprise customers. Open standards aren't winning on pure technical merit yet, but they're winning on the political economy of procurement. Nobody wants to be locked into a single vendor's fabric at the scale these deployments are reaching. The report is optimistic on Ethernet's long-term trajectory for scale-out, and I think that's directionally right, though the timeline always seems to slip a bit. Still a solid read for the bigger picture. Skyler: Okay, Broadcom. Their page on "Scale-out Networking for A I Clusters" is less a news story and more of a positioning statement, but it's Broadcom, so we pay attention. They're the Tomahawk and Jericho shop — they make the merchant silicon that's inside a huge fraction of the switches deployed in hyperscaler and cloud provider networks. So when Broadcom talks about how scale-out networking for A I should work, they're not just describing the market, they're describing what their chips are built to do. The key technical claim they're pushing is around the ability to handle the all-reduce and collective communication patterns that dominate A I training traffic. These aren't just high-bandwidth flows — they're highly synchronized, barrier-heavy, latency-sensitive patterns that can absolutely fall apart if your switching fabric introduces any meaningful jitter or queuing delay. Broadcom's pitch is that their merchant silicon, combined with the right RDMA and congestion control stack, can deliver the performance you need at a price point that proprietary solutions can't match. And there's something to that. The Tomahawk 5, for example, is sitting at 51.2 terabits per second of switching capacity — that is a staggering number. The challenge has always been whether the software stack around it — the congestion control, the load balancing, the collective communication offload — can keep pace with what the hardware is capable of. And that's where Broadcom has been investing. They've been pushing hard on their RDMA implementation, on adaptive routing, on the kinds of features that matter specifically for A I traffic as opposed to general-purpose cloud traffic. The competitive context here is InfiniBand and specifically Nvidia's Quantum-X800 switches. Nvidia has the advantage of tight co-design between the GPU, the NIC — ConnectX — and the switch. Broadcom's counter is ecosystem breadth and merchant availability — you can get their silicon in gear from Arista, from Cisco, from a dozen white-box vendors. The bet is that open wins over time. And given the trajectory of hyperscaler procurement strategy, that's not an unreasonable bet. Skyler: Now here's the story that actually made me sit up straight this week. Nokia announced that they successfully completed an end-to-end Ultra Ethernet test across their datacenter switch family. Let me give you a second to appreciate what that means, because "end-to-end U E C test" is doing a lot of work in that headline. The Ultra Ethernet Consortium — which launched in 2023 with a pretty aggressive set of founding members including A M D, Arista, Broadcom, Cisco, HPE, Intel, Meta, and Microsoft — has been working on a spec that would extend Ethernet's capabilities specifically for A I and HPC workloads. The core problem they're solving is that standard Ethernet, even at high bandwidth, has characteristics that make it painful for the tight, synchronous collective communication patterns of A I training. Things like head-of-line blocking, suboptimal congestion response, lack of native support for the transport semantics that collective operations need. U E C is supposed to fix that by defining new transport-layer behaviors, better multipath utilization, and receiver-driven load balancing that standard IP-over-Ethernet doesn't give you. The spec is still evolving — version 1.0 was published but there's significant ongoing work around the transport layer — so when Nokia says they ran an end-to-end test, the question is: end-to-end of what, exactly? And they're being pretty specific — they're saying they ran UEC traffic across their full switch family, which includes the 7250 and 7750 platforms, meaning they validated that the behavior propagates correctly across multiple hops, across different device types in their portfolio, under conditions relevant to A I cluster traffic. That's meaningful. It's not just a loopback test on a single device. You need to prove that the congestion response, the load balancing decisions, the transport behaviors all work correctly when you have actual switch-to-switch paths with real traffic. Nokia isn't the flashiest name in the datacenter switching conversation — that space is dominated by Arista, Cisco, and increasingly by Nvidia — but they are serious infrastructure vendor with real deployments in carrier and hyperscaler environments. Them doing this work signals that UEC is progressing from a paper spec toward something you might actually configure in a production switch. The competitive angle here is interesting too. Nokia demonstrating end-to-end UEC capability gives them a credible story in the A I datacenter market that they haven't historically had. Watch for more interoperability demos from UEC members over the next couple of quarters — that's usually how these standards transitions start to build momentum. We're not at broad interop testing across vendors yet, but Nokia planting a flag is a real signal. Skyler: And finally, IEEE Spectrum ran a piece on optical networking in A I datacenters — "A I Data Centers Get a Laser Boost With Optical Networking" — and this is a topic that I think doesn't get enough airtime in the fabric conversation, because it's kind of foundational to everything else we've been talking about. Here's the thing — all of the bandwidth numbers we throw around when we talk about switching ASICs and interconnect standards, those are only meaningful if you can actually move the optical signals that carry that bandwidth. And the optics side of this story is undergoing as dramatic a transformation as the silicon side. The piece from IEEE Spectrum focuses on a few key areas. First, the move toward co-packaged optics, or CPO. Traditional datacenter networking uses pluggable optical transceivers — you've got your QSFP-DD modules, your OSFP modules — they're great, they're interchangeable, but they consume a lot of power and they introduce signal loss that gets worse as you push to higher data rates. CPO takes the optical components and packages them directly with the switching ASIC, which dramatically reduces the electrical path length and the associated power budget for the serializer-deserializer — the SerDes — circuitry. At 51.2 terabits per second and beyond, CPO starts to become not just attractive but arguably necessary from a power standpoint. The second thread is around silicon photonics. Rather than using traditional III-V materials like indium phosphide for the laser and modulator components, silicon photonics tries to implement as much of the optical functionality as possible in standard CMOS silicon. The appeal is cost and integration — you can potentially co-integrate photonic components with electronic logic using processes that the semiconductor industry already knows how to scale. The challenge is that silicon is not great at generating light, so you still need external laser sources, and the integration story is still evolving. But progress has been real. Third, the piece touches on what all of this means for the power envelope of A I datacenters. A single large training cluster can pull tens of megawatts. A meaningful fraction of that is optical networking and the SerDes driving it. Getting optics right isn't just a performance story — it's an energy story, and energy is increasingly the binding constraint on how fast you can scale A I infrastructure. The co-packaged optics ecosystem is still early — there are interoperability questions, there are thermal management challenges, there's the whole question of replaceability in the field when something fails — but the direction of travel is clear. If you're thinking about what switches look like in three to five years, CPO is going to be a big part of that picture. Skyler: Alright, that's a wrap on today's episode of Line Rate. We covered a lot of ground — from Cisco's argument that A I networking is a beyond-the-datacenter problem, to Futuriom's trend survey on scale-out and edge connectivity, to Broadcom's merchant silicon pitch for A I fabric, Nokia's real end-to-end UEC validation across their switch family, and IEEE Spectrum's deep dive on what optics and co-packaged photonics mean for the future of A I datacenters. Thanks again to our sponsors — Line Rate is brought to you by Ada, who helps organizations integrate A I into their workflows, and Zen Semiconductor, an umbrella company with a portfolio of companies across the silicon industry. If you got value out of this episode, please subscribe wherever you get your podcasts and share it with someone who needs to be in this conversation. All the source links and additional context are in the show notes. See you next time.