Nokia hails end-to-end Ultra Ethernet test in AI data center drive - SDxCentral

Skyler: Hey, welcome to Line Rate — the show tracking the chips, standards, and companies wiring together the modern A I datacenter. I'm Skyler Monroe, and if it moves bits between accelerators, racks, or datacenters, we cover it here. Big thanks to our sponsors today: Ada, who helps organizations actually integrate A I into their real workflows, and Zen Semiconductor, an umbrella company with a portfolio of businesses across the silicon industry, including Ada for A I deployment at scale. Alright — we've got a packed show today. Nokia just ran an end-to-end Ultra Ethernet test in a live A I datacenter environment. Broadcom is quietly assembling what might be the most audacious infrastructure land grab in the industry. And co-packaged optics is going from "interesting research topic" to "this is actually happening" — with multiple analyst firms dropping market projections. Let's get into it. Skyler: So let's start with Ultra Ethernet, because this is one of those stories where the headline sounds exciting and the details are actually worth unpacking carefully. Nokia announced a successful end-to-end Ultra Ethernet Consortium — U E C — test in what they're describing as an A I datacenter environment. And look, any time you see "end-to-end" and "Ultra Ethernet" in the same sentence, that's worth paying attention to, because until recently most of what we'd seen from the U E C was spec documents, working group outputs, and demo setups that were more lab bench than production environment. Skyler: So what is the U E C actually trying to do? The pitch is straightforward — take Ethernet, which everyone already knows and has infrastructure for, and make it genuinely competitive with InfiniBand for A I training workloads. The problem the U E C is solving is that standard Ethernet has some fundamental behaviors that are just bad for tightly-coupled A I collective operations. Things like packet reordering, congestion spreading in ways that create head-of-line blocking, and transport layer semantics that weren't designed with all-reduce and all-to-all patterns in mind. U E C addresses this with a set of transport enhancements — better congestion control, reordering tolerance at the receiver, and more explicit end-to-end reliability mechanisms. Skyler: Nokia's role here is interesting. They're a networking equipment vendor, and the fact that they're running an end-to-end test — meaning from NIC through switch fabric to the other NIC — demonstrates that at least one major vendor is putting real engineering resources into validating this stack in something resembling a real environment. That matters because U E C has been criticized, not unfairly, for being a consortium that moves slowly and whose spec still has open items. Seeing end-to-end validation is a maturity signal. Skyler: Now, I want to be honest about what we don't know yet. The performance numbers from this test weren't fully disclosed in the coverage I've seen, and "A I datacenter drive" is doing a lot of work in that headline. We don't know if this was a production-scale cluster, how many endpoints were involved, or how the latency and throughput compared head-to-head with InfiniBand or even RoCE v2 setups. Those details matter enormously for whether U E C is genuinely competitive or still a "promising alternative" that needs another year of baking. Skyler: The competitive picture here is real though. InfiniBand is NVIDIA's home turf through the Mellanox acquisition, and the rest of the industry has very strong incentives to have a credible Ethernet-based alternative. Broadcom, Arista, Cisco, and now clearly Nokia are all invested in making U E C work. And importantly — the hyperscalers want this. If you're Google or Meta or Microsoft, you want to be able to build A I fabric without being locked into a single vendor's end-to-end stack. U E C, if it delivers, is that path. So Nokia's test is a meaningful data point, even if we need more detail before calling it a full validation. Skyler: Alright, story two — Broadcom. The Register ran a piece framing this as Broadcom quietly plotting a takeover of the A I infrastructure market, and honestly, "quietly" is doing a lot of heavy lifting in that headline, because if you've been watching Broadcom for the last 18 months, this has not been quiet at all. Skyler: Let me lay out what Broadcom is actually doing, because it's a multi-dimensional play. First, the custom silicon business. Broadcom has been the go-to partner for hyperscalers who want to build their own A I accelerators — the XPUs — rather than buying NVIDIA GPUs. Google's TPUs have Broadcom DNA. Meta's MTIA has Broadcom involvement. There are at least two or three other major hyperscaler A I chip programs that are either confirmed or strongly rumored to involve Broadcom's ASIC design services. This is not a coincidence — Broadcom built deep custom silicon capabilities over years of acquisitions and they're now harvesting that investment. Skyler: Second, the networking side. Broadcom makes the Tomahawk and Jericho switch silicon that runs a huge percentage of datacenter switching infrastructure globally. As A I clusters scale up and the demand for high-bandwidth, low-latency switching explodes, Broadcom's switching ASICs are right in the middle of that. The Tomahawk 5 — at 51.2 terabits per second — is the switch chip that serious scale-out A I networks are being built on right now. And Broadcom is already talking about what comes next. Skyler: Third — and this is the part that I think is underappreciated — Broadcom's SerDes and optical DSP IP. As you go to higher speeds — 112 gig PAM4, 224 gig — the SerDes becomes the hard part, and Broadcom has some of the best SerDes IP in the industry. That matters for both the switch chips and for the co-packaged optics transition, which we're going to talk about in a few minutes. Skyler: The Register's framing of "takeover" is a bit dramatic, but the underlying logic is sound. Broadcom's strategy is to be indispensable at every layer — compute silicon, switching silicon, and the SerDes and DSP IP that makes optical interconnects work. If that plays out, they're not just a component supplier, they're a structural dependency for A I infrastructure at hyperscale. The counterargument is that this makes Broadcom a target for vertical integration — hyperscalers may want to take some of this in-house. But that's a five-year conversation, and right now Broadcom is executing. Skyler: One thing worth noting — Broadcom's CEO Hock Tan is one of the most disciplined capital allocators in the semiconductor industry. When he makes an acquisition or a strategic bet, it tends to be calculated. The A I infrastructure play isn't opportunistic — this looks like a multi-year strategy that's been in motion for a while, and it's starting to become visible to everyone now that the A I build-out is consuming hundreds of billions of dollars in capex annually. Skyler: Okay, let's shift gears to co-packaged optics, because there are actually three stories here that I want to weave together — a TrendForce report on CPO and NVIDIA's compute architecture, a Yole Group analysis on CPO and A I infrastructure, and a MarketsandMarkets report on photonic packaging. I'm going to treat these as one conversation because they're all pointing at the same inflection point. Skyler: The TrendForce piece is particularly interesting because it connects NVIDIA's architecture roadmap explicitly to co-packaged optics adoption. Here's the core argument: as NVIDIA scales up its compute architecture — and we're talking about the transition from NVLink 4 to NVLink 5, and the scale of clusters that Blackwell Ultra and whatever comes after it demands — the bandwidth requirements between compute nodes start to strain what pluggable optics can deliver within the power and density constraints of the rack. Co-packaged optics, or CPO, solves this by integrating the optical engine much closer to the switch or NIC ASIC — reducing the electrical distance the signal has to travel before it hits a photon, which cuts power and increases density. Skyler: TrendForce's projection is that CPO penetration in A I datacenters will rise steadily. Now, "rise steadily" is analyst-speak for "we think this is real but we're hedging." What I'd say is that the technical case for CPO is essentially closed — the physics are not in dispute. The open questions are manufacturing yield, reliability at volume, and the ecosystem around repair and replacement, because when your optics are co-packaged with your switch ASIC, you can't just hot-swap a bad transceiver the way you can with a QSFP-DD. Skyler: The Yole Group piece takes a broader view, looking at CPO across the full A I infrastructure stack — not just the scale-up interconnect between GPUs, but also scale-out fabric between racks and pods. Yole's point is that the power budget pressure is real at every level. A large A I training cluster is consuming tens of megawatts just for compute, and the interconnect — the electrical and optical signaling — is a meaningful fraction of that. CPO's efficiency advantage compounds as you scale. A few watts saved per port times hundreds of thousands of ports is real money and real cooling infrastructure. Skyler: The MarketsandMarkets report puts a dollar figure on the photonic packaging market — and these reports always have to be taken with some salt on the specific numbers, but the directional signal is consistent with what Yole and TrendForce are saying. This is a market that's transitioning from early-adopter phase to volume phase over the next three to five years, and the companies that are establishing manufacturing and design-in wins right now are going to have a structural advantage. Skyler: Who are the players to watch in CPO? Intel has been beating the drum on this for years with their silicon photonics work out of what was Luxtera. Broadcom — and here's that name again — has CPO programs tied to their switch silicon. Marvell has been aggressive here too, particularly for switch-side CPO. And then you have the pure-play photonics companies — Coherent, Lumentum, II-VI heritage companies — who are racing to establish themselves as the optical engine suppliers for co-packaged designs. Skyler: The NVIDIA angle is particularly fascinating. NVIDIA's architecture creates the demand signal, but NVIDIA itself — through NVLink and NVSwitch — has historically kept that interconnect proprietary and copper or active electrical cable based at shorter ranges. If the next generation of NVIDIA architecture requires optical at the switch level, that's an enormous pull signal for CPO at volume. And that's what TrendForce is implying. Worth watching closely as Blackwell Ultra deployments ramp and as whatever comes after Blackwell starts showing up in roadmap leaks. Skyler: Alright, let's land this plane. What's the through-line across all of today's stories? It's that A I infrastructure is scaling faster than any previous technology wave, and every layer of the stack is under pressure to evolve simultaneously. Ultra Ethernet is trying to make scale-out networking competitive and open. Broadcom is positioning to be the foundational silicon layer across compute, switching, and optical DSP. And co-packaged optics is moving from research to deployment because the physics of pluggable optics simply don't scale to where we're going. Skyler: These aren't independent trends — they're coupled. U E C needs high-performance switch ASICs, those ASICs need CPO to hit the bandwidth density targets, and the custom silicon programs need companies like Broadcom to make the SerDes work. It's a system, and understanding how these pieces interact is how you anticipate where the industry goes next rather than just reacting to headlines. Skyler: Thanks for tuning in to Line Rate. This has been a great episode to dig into — lots of moving pieces, and honestly we could spend a whole episode on any one of these stories. If you want to go deeper, the show notes have links to all five source pieces we covered today. Line Rate, brought to you by Ada — who helps companies integrate A I into their real-world workflows — and Zen Semiconductor. If you found this useful, subscribe wherever you get your podcasts, leave a review if you have a minute, and share it with someone in your org who should be paying attention to this stuff. We'll be back soon with more. Stay sharp.