# On-Chip Silicon Photonics for Data Center Interconnects

An episode of [The Hardware Podcast with Fexingo: Chips, Devices, and Electronics Engineering Conversations](https://invisiblenetwork.ai/podcast/1896799982), 17 Jul 2026, 10 min.

Audio: https://audio.fexingo.com/business/the-hardware-podcast/episode-0117.mp3

Lucas and Luna explore how silicon photonics is moving from lab curiosity to production inside data centers. They focus on a specific case: Intel's recent demonstration of an integrated photonic engine that replaces traditional copper cabling between server racks. The pair break down how this works at the chip level — modulators, waveguides, and photodetectors etched into standard CMOS silicon — and what the performance numbers actually mean: 800 gigabits per second over fiber, with power draw below 5 picojoules per bit. They discuss why this matters for AI clusters, where copper bottlenecks are already a headache. The conversation also touches on the economics: silicon photonics leverages existing fabs, potentially lowering cost per port compared to discrete optics. A concrete look at a technology that quietly changes how data moves. #SiliconPhotonics #DataCenter #Intel #OpticalInterconnect #CMOS #Photonics #ChipDesign #DataCenterNetworking #AIHardware #800G #PowerEfficiency #FiberOptics #Semiconductor #Technology #FexingoBusiness #BusinessPodcast #TheHardwarePodcast #OnChip Keep every episode free: buymeacoffee.com/fexingo

## Transcript

Lucas: If you look at a modern data center — say, one of the big AI training clusters — there is a moment where the copper cable simply stops working. The signal degrades. The power to push data that last meter becomes untenable.
Luna: And that's where people start talking about optical fiber. But traditionally, plugging a fiber into a server rack meant discrete transceivers, separate lasers, a whole extra supply chain.
Lucas: Right. Which is why the chip industry has spent the last decade trying to do something that sounds almost contradictory: etch optical components into silicon. Not just electronics — actual photonics, on the same die, using the same fabrication lines.
Luna: Silicon photonics, in other words. And it's not theoretical anymore. Intel just showed a fully integrated photonic engine that hit 800 gigabits per second over fiber, with power under 5 picojoules per bit.
Lucas: Let's sit with those numbers for a second. 800 gigabits — that's a hundred gigabytes per second. Enough to stream roughly 20,000 simultaneous 4K video streams through a single chip. And the power per bit — 5 picojoules — is about a tenth of what a typical copper SerDes uses today.
Luna: So you're getting higher bandwidth and lower power in the same package. That's the killer combination. But how does it actually work at the chip level?
Lucas: The key components are modulators, waveguides, and photodetectors — all made from silicon or silicon nitride, sitting right next to the CMOS logic. The light source is still a separate laser, usually indium phosphide, but it's coupled into the silicon die through a tiny grating coupler, etched right into the top layer.
Luna: So the laser itself isn't on-chip, but everything else is. That's the pragmatic approach. Indium phosphide lasers are better than silicon lasers, so you hybrid-integrate them.
Lucas: Exactly. The silicon handles the routing and switching of light. The waveguides are basically microscopic glass pipes, about half a micron wide, etched into the silicon dioxide layer. They guide the light from the laser, through the modulator — which encodes data by changing the refractive index using the plasma dispersion effect — and then to the photodetector at the other end.
Luna: And all of this happens on a standard 300-millimeter wafer, using pretty much the same equipment used to make processors. That's the economic story — you don't need a specialized fab.
Lucas: Right. The cost per port drops significantly because you're leveraging existing capacity. Intel's current product, the Silicon Photonics 100G, has been shipping for years. But the new 800G engine takes it to a different league.
Luna: And the timing matters. AI clusters like the ones training large language models are hitting exactly this bottleneck. The compute is there, but the interconnect between GPUs and memory is limited by copper reach and power.
Lucas: NVIDIA's DGX systems, for example, rely on NVLink for gpu to gpu communication, but at the rack level, you still need optical links. And as models grow to a trillion parameters, the amount of data that needs to move between racks explodes.
Luna: So silicon photonics isn't just a nice efficiency gain — it's becoming structural for the next generation of infrastructure. Let's talk about a specific design: the Mach-Zehnder modulator.
Lucas: That's the workhorse. It splits the light into two paths, applies an electric field to one path to change its refractive index, then recombines them. The interference pattern creates the 1s and 0s. It's elegant, and it's been used in fiber optics for decades — just never integrated onto a silicon chip at this scale.
Luna: The challenge is that silicon has a weak electro-optic effect compared to materials like lithium niobate. So to get enough modulation, you need a long device — sometimes millimeters. That eats die area.
Lucas: Intel's trick is to use a resonant ring modulator, which is much smaller — tens of microns instead of millimeters. It uses a ring waveguide that traps light at a specific wavelength; applying voltage shifts the resonance, effectively turning the light on and off. This saves space and power.
Luna: But ring modulators are sensitive to temperature and fabrication variations. That's been the engineering battle — making them robust enough for data center temperatures.
Lucas: And they've largely won it. Intel's 800G engine uses multiple wavelength channels — basically different colors of light — each carrying data. That's wavelength division multiplexing, or WDM, on a chip.
Luna: So a single fiber can carry 8 channels at 100 gig each, or 4 at 200 gig. That gets you to 800 gig. And the next step, 1.6 terabit, is already being discussed.
Lucas: Of course, it's not just Intel. Companies like Cisco, via its acquisition of Luxtera, and Marvell, via Inphi, are also in this space. And startups like Ayar Labs are doing optical interposers that sit right next to the compute die.
Luna: But Intel has the advantage of its own fabs. They can iterate and control the process. That's a real moat.
Lucas: It is. And speaking of what makes this possible — listener support is actually what keeps this show ad-free and allows us to dive into deep technical topics like this. A couple of dollars a month really does make a difference. If you've gotten something out of these conversations, consider buying us a coffee at buy me a coffee dot com slash fexingo.
Luna: Yeah, it genuinely helps. Every contribution goes straight into production time and research for episodes like this.
Lucas: So back to silicon photonics — one angle I find fascinating is the reliability story. Lasers are typically the weakest link in optical links, but when you co-package them with silicon, you can add monitoring photodiodes and feedback circuits that adjust the laser bias to maintain performance over years.
Luna: That feedforward and feedback loop is something you can only do if the electronics and photonics are on the same package. It's a systems-level optimization.
Lucas: Right. And it extends to the thermal side. Silicon photonics operates at higher temperatures than traditional optics, which is critical in a data center where ambient can hit 40 degrees Celsius or more.
Luna: Let's talk about the packaging itself. One of the bottlenecks is coupling light from the fiber into the chip. If you misalign by even a micron, you lose half the light.
Lucas: That's where grating couplers come in. They diffract the light from the fiber into the planar waveguide at a precise angle. Intel uses a process called 'edge coupling' where the fiber is aligned to a facet on the chip edge, but grating couplers allow for surface coupling, which is easier to test.
Luna: And the industry is moving toward co-packaged optics — putting the optical engine right next to the switch ASIC on the same substrate. That eliminates the need for pluggable transceivers entirely.
Lucas: Exactly. The current generation of data center switches, like the Broadcom Tomahawk 5, can handle 51.2 terabits per second of switching capacity. To get that data off the chip, you need optical I/O. Copper just can't do it at that density and power.
Luna: So silicon photonics becomes the enabler for the next generation of switches. And that's why hyperscalers — Google, Amazon, Microsoft — are investing heavily.
Lucas: Google already uses silicon photonics in its Jupiter network. They've been running it for years, and they publish papers on their design choices. It's not a secret anymore.
Luna: One question I get from engineers is: does it replace copper altogether? And the answer is no — inside the box, for short distances under a meter, copper is still fine. But for rack to rack and beyond, optics take over.
Lucas: And the crossover distance keeps shrinking as data rates increase. At 800 gig, the crossover might be one or two meters. So eventually, even within a rack, you might see optical backplanes.
Luna: That would be a huge shift for server design. No more heavy copper cables, less airflow obstruction, better thermal management.
Lucas: And it's already happening in prototype. There are demos of fully optical motherboards using polymer waveguides embedded in the PCB. But that's probably a few years out.
Luna: Let's zoom back to the chip level. One of the neatest innovations is the use of micro-ring resonators for wavelength-selective filtering. You can drop a specific wavelength off the bus waveguide and route it to a particular receiver.
Lucas: That's essentially an on-chip wavelength router. You can build a network-on-chip using light instead of wires, with different wavelengths serving different cores. That's the holy grail — optical interconnects inside the processor itself.
Luna: But we're not there yet. The loss in the waveguides and the crosstalk between wavelengths still limit the scale. It's an active research area.
Lucas: For now, the commercial sweet spot is data center interconnects. And the numbers back it up: the silicon photonics market is projected to hit $5 billion by 2030, driven entirely by AI and cloud.
Luna: That's a nice anchor. It's one of those rare technologies where the physics, the economics, and the market demand all align at the same time.
Lucas: Right. And it's happening on the same silicon that's been refined for fifty years. That, to me, is the story — not a revolution, but an elegant integration.

Shows are listed from Apple Podcasts' charts and directory, and Podcast Index where it is used. Every show and episode belongs to its publisher and plays from the publisher's own feed; this site hosts none of them.