Nvidia updates data center roadmap with Rosa CPU and stacked Feynman GPUs — optical NVLink, Groq LPUs with NVFP4, and NVLink also on deck - Tom's Hardware

Skyler: Hey, welcome back to Line Rate — the show about datacenter fabric and A I high-speed interconnect. I'm Skyler Monroe, and if you care about the pipes, the silicon, and the standards that hold together modern A I infrastructure, you are in exactly the right place. Quick thanks to our sponsors: Ada, which 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. Alright — today we've got a packed show. Nvidia just dropped a roadmap update with optical NVLink and stacked GPUs, Google revealed a fabric architecture called Virgo that's throwing out the three-tier playbook, Lightmatter is pushing its photonic interconnect platform into high-volume manufacturing territory, and a photonics startup called Xscape just raised 37 million dollars and launched a new laser aimed squarely at A I datacenters. Let's get into it. Skyler: So Nvidia had a big roadmap drop, and there's a lot to unpack here — we're talking several generations of hardware announced essentially in one breath. Let's walk through what matters for the interconnect and fabric crowd. The headline GPU news is Feynman, which is the successor to Blackwell. Feynman is going to use stacked dies — so we're talking chiplet-style vertical integration for the GPU itself, which is a significant departure from how Nvidia has traditionally built monolithic or horizontally tiled GPU packages. That alone is interesting from a thermal and signaling standpoint, but what really caught my attention is what they're doing to NVLink. Skyler: NVLink is Nvidia's proprietary GPU-to-GPU interconnect, and it has been the secret sauce behind their NVSwitch-based scale-up fabrics. Right now NVLink runs over copper — very short-reach, very high-bandwidth, lives inside the node or the rack. The update here is optical NVLink. Nvidia is moving NVLink onto optical transport, which fundamentally changes the reach profile. If you can run NVLink over fiber at scale, you're no longer constrained to a tight rack topology for scale-up. You can start talking about scale-up across longer distances — potentially across rows, maybe across pods — without collapsing to a scale-out Ethernet or InfiniBand layer for GPU-to-GPU traffic. That is a genuinely big deal. Skyler: The other piece is the Rosa CPU, which is their next data center CPU on the roadmap. Now Rosa is interesting in context because Nvidia has been building out their Grace-based CPU story — Grace Hopper, Grace Blackwell — and Rosa continues that trajectory. For fabric folks, the CPU matters because it defines the host-side interface to the A I fabric: PCIe generation, memory bandwidth, how you feed the GPUs. But honestly, the optical NVLink announcement is the one that changes topological assumptions for system architects. Skyler: And then there's the Groq angle in here, which is a little unexpected. Groq LPUs with NVFP4 support is being floated — NVFP4 is Nvidia's 4-bit floating point format, and if third-party accelerators are adopting it, that's Nvidia using numerical formats as a soft lock-in mechanism, which is a playbook worth watching. The broader implication of this roadmap drop is that Nvidia is signaling they will control the optical layer inside the rack and potentially beyond it. That puts them in more direct tension with the open optical interconnect ecosystem — the CPO players, the co-packaged optics consortiums, the UALink crowd. Nvidia is essentially saying: we'll do the optics ourselves, on our terms, under the NVLink umbrella. That's a competitive statement as much as it's a product announcement. Skyler: Alright, let's talk about Google's Virgo. This one came out of Google's infrastructure team and it's a really substantive architectural announcement. Virgo is a scale-out fabric for A I accelerators, and the core claim is that it eliminates the traditional three-layer network topology — you know, access, aggregation, core — and replaces it with a flat, unified fabric. Now that framing will sound familiar to anyone who's been following the Clos network debates in hyperscaler circles, but let me tell you why Virgo is interesting beyond the marketing. Skyler: Traditional datacenter networks for A I training have this awkward tension: you want a high-bandwidth, low-latency scale-up domain for your tightly coupled GPU collectives — your AllReduce traffic, your gradient synchronization — and then you need a separate scale-out domain to span thousands of GPUs across racks and pods. Those two networks have historically had different latency profiles, different congestion behaviors, different management planes. You end up with either a hybrid fabric where you're managing two overlapping topologies, or you make compromises in one dimension to serve the other. Skyler: What Google says Virgo does is unify those domains. A single fabric layer that handles both the tight collective traffic and the broader scale-out communication without a hierarchical tier break in between. They're saying this is built around their TPU pods, which makes sense — Google's TPU interconnect has always been a custom affair, and Virgo appears to be the fabric architecture that ties together the latest generation of TPU scale-out at the cluster level. The claim is that by removing that tier break, you reduce latency variability, simplify congestion control, and ultimately get better utilization out of your accelerator fleet because you're spending less time in the network. Skyler: Now there are real engineering tradeoffs here that are worth being skeptical about. Flat fabrics sound great until you have to build the switching silicon to support them at scale. The reason three-tier networks exist is partly a practical one — you can build each layer with commodity merchant silicon at reasonable cost and power. A truly flat fabric at hyperscaler scale either requires very high-radix switches, very long fiber runs, or some clever optical topology that lets you avoid the electrical switching bottleneck. Google hasn't fully published the hardware substrate here, so we don't know exactly what Virgo is running on at the physical layer. Skyler: What we do know is that this is Google drawing a very clear line: we don't need your three-tier spine-leaf Ethernet fabric for A I. We built something purpose-built. That's a shot across the bow at the networking vendors who have been selling into hyperscaler A I buildouts, and it's a signal to the rest of the industry that the era of adapting enterprise network topologies to A I workloads is over. The hyperscalers are going custom, and the merchant silicon ecosystem better be ready to serve that — or watch its addressable market shrink. Skyler: Next up — Lightmatter. So Lightmatter has been on my radar for a while as one of the more technically credible players in the photonic interconnect space, and this week they made two announcements that are worth covering together. First, they unveiled the Passage L20, which is their latest photonic interconnect product. Second, they announced something called vClick Optics, which is their offering for high-volume co-packaged optics manufacturing. Skyler: Let me take the Passage L20 first. Passage is Lightmatter's interconnect platform that uses photonics to build chip-to-chip and board-to-board communication fabrics. The L20 is positioned as a scale-out interconnect — so we're talking about connecting accelerators across nodes, not just within a package. The pitch is lower latency, lower power, and higher bandwidth density than electrical or pluggable optical alternatives at the distances where things get hard for copper. We don't have the full spec sheet in front of us, but the L20 naming suggests this is aimed at the 20-nanometer class integration — whether that's the process node for the photonic components or a bandwidth density figure, I'd want to dig into that more. Skyler: The more strategically interesting announcement is vClick Optics. Co-packaged optics — CPO — is the approach of integrating optical components directly into the switch or ASIC package rather than using pluggable modules on the faceplate. CPO has been talked about for years as the inevitable next step in data center optics, but the manufacturing story has been genuinely hard. You're trying to take what has historically been a precision, low-volume optical assembly process and turn it into something that runs at the volume and yield that hyperscaler deployments demand. Skyler: vClick Optics is Lightmatter's answer to that manufacturing challenge. They're essentially saying: we've figured out a way to do high-volume CPO assembly, and we're offering that as a service or platform for others building CPO products. If that's real — and I want to emphasize, we'll need to see the yield data and customer traction to know — it could be a significant unlock. Because right now, a lot of the CPO roadmaps from the big switch vendors and hyperscalers are bottlenecked not by the photonic designs but by the ability to manufacture them at scale with acceptable yields. Skyler: Lightmatter is also doing something strategically clever here: they're not just selling interconnect products, they're positioning themselves as the manufacturing backbone for the broader CPO ecosystem. That's a different business model than a pure chip company, and it's a bet that the value in the next generation of optical interconnect lives in the assembly and integration layer, not just the photonic design. It'll be interesting to see if the big optical module players — your Coherents, your IIs-VIs, the Inphi-era folks — view this as competition or as a potential foundry partnership. Skyler: Quick one on Thailand — Data Center Dynamics reported that Thailand is moving toward finalizing new data center regulations by mid-October, and the key threshold is that facilities over 100 megawatts will be classified as hyperscale. Now, 100 megawatts is a real number — we're in a world where individual A I training campuses are pushing past that, so this isn't an abstract line. For anyone planning capacity in Southeast Asia, where Thailand has been a growing hub for regional cloud infrastructure, this classification matters because hyperscale designation typically comes with different permitting requirements, grid access rules, and sometimes environmental review obligations. It's not a technical story per se, but for the datacenter architects and the investors in this audience who are thinking about where to build next — Southeast Asia regulatory clarity is actually something you want to track. Watch for the mid-October announcement. Skyler: Last story — Xscape Photonics. They just announced 37 million dollars in new funding and launched an eight-wavelength laser product aimed at A I datacenter networks. Let's break this down. Xscape is a photonics startup, and the funding round — 37 million dollars — is a solid Series A or B-scale number for a deep photonics company. These aren't cheap businesses to build because you're dealing with compound semiconductor fabrication, precision packaging, and qualification cycles that take time. Skyler: The product launch is an eight-wavelength laser, and that's the technically interesting piece. In optical datacenter interconnects, your laser source is foundational — it determines your wavelength plan, your channel spacing, your compatibility with wavelength-division multiplexing architectures, and ultimately your aggregate bandwidth per fiber. An eight-wavelength laser is targeting the WDM applications where you want to push multiple channels down a single fiber to multiply your effective bandwidth without multiplying your fiber count. Skyler: Now, eight wavelengths is not the highest count you'll see in the market — telecom WDM systems run hundreds of channels — but for datacenter applications, the optimization is different. You want low cost, low power, high yield, and compatibility with the silicon photonics platforms that the big switch and ASIC vendors are designing around. Eight wavelengths in a package that can be co-packaged or surface-mounted into a CPO module is a very practical, very datacenter-native target. Skyler: The A I datacenter context here is important. As GPU cluster bandwidth requirements have exploded — we're talking hundreds of terabits per second of fabric bandwidth for a large training cluster — the optical components ecosystem is under pressure to deliver more bandwidth per unit cost and power. Laser sources are not a solved problem at those scales. The yield, the thermal stability, the wavelength accuracy — these are active engineering challenges. A startup raising 37 million dollars and shipping a product says they have at least a credible technical thesis, and the market timing is right. Skyler: Xscape joins a list of photonics component startups that are getting funded on the back of A I infrastructure demand — Ayar Labs, Celestial A I, and others are all working different parts of the photonic interconnect stack. What's interesting about the laser-focused angle is that it's going after a supply chain bottleneck that is real and often underappreciated. The silicon photonics community has made incredible progress on modulators and detectors, but the laser source — especially a high-quality, multi-wavelength, co-packageable laser — remains one of the harder integration problems. 37 million dollars to go after that problem in a market where hyperscalers are spending billions on A I compute? That's a reasonable bet. Skyler: Alright, that's a wrap on today's Line Rate. We covered Nvidia's optical NVLink and stacked GPU roadmap, Google's Virgo flat fabric that's throwing out the three-tier playbook, Lightmatter's Passage L20 and their vClick CPO manufacturing platform, Thailand's incoming hyperscale data center regulations, and Xscape Photonics raising 37 million dollars to go after the A I laser market. As always, Line Rate is brought to you by Ada — helping organizations integrate A I into their real workflows — and Zen Semiconductor. If you found this useful, subscribe wherever you get your podcasts, share it with someone who's arguing about scale-up versus scale-out, and check the show notes for links to everything we covered today. See you next time.