# How Chips Are Using On-Chip Photonic Transceivers for Ultra-Fast Data

An episode of [The Hardware Podcast with Fexingo: Chips, Devices, and Electronics Engineering Conversations](https://invisiblenetwork.ai/podcast/1896799982), 9 Jul 2026, 9:20.

Audio: https://audio.fexingo.com/business/the-hardware-podcast/episode-0101.mp3

Episode 101 of The Hardware Podcast explores the cutting-edge integration of photonic transceivers directly onto silicon chips, bypassing traditional copper interconnects. Lucas and Luna break down the physics of silicon photonics, the manufacturing challenges at TSMC and Intel, and real-world applications from data centers to autonomous vehicles. They discuss the role of companies like Ayar Labs in pushing optical I/O into production, and how this shift could reshape bandwidth, latency, and power efficiency in computing. A concrete look at a technology poised to change how chips talk to each other. #Photonics #SiliconPhotonics #OnChip #Transceivers #OpticalInterconnect #DataCenter #AI #MachineLearning #TSMC #Intel #AyarLabs #Bandwidth #Latency #PowerEfficiency #Technology #Hardware #FexingoBusiness #BusinessPodcast Keep every episode free: buymeacoffee.com/fexingo

## Transcript

Lucas: You know that moment in a data center when a rack of servers is screaming at full throttle, but the bottleneck isn't the processor — it's the little copper wire carrying data between chips? That's been the silent ceiling on performance for years.
Luna: And we've been bumping up against it more and more with AI workloads, where you've got hundreds of GPUs trying to talk to each other simultaneously.
Lucas: Exactly. So the hardware response — and it's been brewing for a decade — is to replace those copper traces with light. Not fiber optic cables running into the rack, but actual photonic transceivers fabricated right onto the silicon die. On-chip lasers, modulators, and photodetectors.
Luna: Before we dive deeper — and this ties in because it takes infrastructure to fund this kind of R&D — a quick honest thing: this show stays ad-free thanks to a small group of listeners who chip in monthly over at buy me a coffee dot com slash fexingo. It's what lets us keep covering these deep hardware topics without chasing clicks.
Lucas: Yeah, it's a model that works because of you all. And honestly, it's why we can spend an entire episode on something like integrated photonics instead of rushing through it. So, thank you.
Lucas: Alright, so on-chip photonic transceivers — the idea is that instead of sending electrical signals across a PCB trace that gets lossy at high frequencies, you convert that electrical signal into light, route it through a tiny waveguide etched into the silicon, and convert it back at the destination.
Luna: And the light can carry data at much higher rates — we're talking hundreds of gigabits per second per channel — with practically no signal degradation over the distances inside a chip package.
Lucas: Right. The key enabler here is silicon photonics. Silicon is transparent at certain infrared wavelengths, around 1.3 to 1.55 micrometers, which is the same band used in telecom fiber. So you can fabricate waveguides, modulators, and detectors using standard CMOS processes.
Luna: But the laser itself is the tricky part. Silicon doesn't emit light efficiently — it's an indirect bandgap semiconductor. So how do you get the light source on-chip?
Lucas: That's the billion-dollar question. The industry has two main approaches. One is to bond a iii v laser — made from materials like indium phosphide — onto the silicon die. That's what companies like Intel have been working on with their integrated photonics research. The other is to use a separate laser source that's coupled into the chip via a fiber or grating coupler.
Luna: And then there's the Ayar Labs approach, which I find really elegant. They separate the laser entirely — it's a remote light source — and they only integrate the modulator and detector on the chip. That avoids the thermal and reliability issues of putting a laser on a hot processor.
Lucas: That's a smart trade-off. Ayar Labs has been making waves with their 'TeraPHY' optical I/O chiplet. It's not a full transceiver on the same die as the CPU, but it's a dedicated chiplet that sits next to the processor in the same package and handles all the optical conversion. They recently announced a partnership with Intel and are sampling to customers.
Luna: And the bandwidth numbers are staggering. They're claiming a single TeraPHY chiplet can deliver 4 terabits per second of bidirectional bandwidth. That's orders of magnitude above what a typical copper SerDes can do.
Lucas: And power efficiency is the other big win. Copper interconnects at those speeds consume a lot of power — sometimes a picojoule per bit or more. Optical links can drop that to under a picojoule per bit, which in a data center translates to megawatts of savings.
Luna: So where does this actually ship first? I'm guessing hyperscale data centers — the cloud providers like Amazon, Google, Microsoft — are the obvious early adopters.
Lucas: Absolutely. And specifically in AI clusters. When you're training a large language model, you have hundreds of GPUs that need to synchronize gradients every few milliseconds. The interconnect becomes the bottleneck. Optical I/O can dramatically reduce that all-reduce time.
Luna: And it's not just training. Inference too — especially for real-time applications like autonomous driving, where latency matters. If the car's perception chip has to talk to the planning chip, optical links could shave off microseconds.
Lucas: Right. And automotive is actually a potential market for silicon photonics, though the reliability requirements are tougher. You need the photonic components to survive under the hood — temperature swings, vibration, dust. But there are companies like Sicoya in Germany working on robust photonic transceivers for exactly that.
Luna: Let's talk about the manufacturing side. TSMC recently announced their own silicon photonics platform, and they're integrating it into their advanced packaging flow. What does that mean for the industry?
Lucas: It's huge. TSMC is the world's largest foundry. If they standardize a photonics process that can be combined with their 3nm or 2nm logic, then every chip designer in the world suddenly has access to optical I/O without having to reinvent the wheel. They've demonstrated a compact transceiver with a 50 Gbps per lane data rate, and they're scaling up.
Luna: And Intel has their own integrated photonics push with their 'ODIN' project. They recently showed a 112 Gbps per lane optical link using a hybrid silicon laser. So it's not just startups — it's the incumbents.
Lucas: Exactly. And the economics are starting to make sense. A few years ago, the cost of integrating photonics was prohibitive — you needed exotic materials, extra mask layers, custom packaging. But as process nodes advance, the incremental cost of adding a few photonic layers is shrinking.
Luna: One thing I find interesting is the wavelength-division multiplexing angle. You can send multiple wavelengths of light down the same waveguide, each carrying independent data. That multiplies the bandwidth without needing more physical lanes.
Lucas: Yeah, that's the killer feature. In a copper trace, you're limited to one signal per wire. But with WDM, you can have 4, 8, even 16 channels on a single waveguide, each at 50 or 100 Gbps. Suddenly a single optical link can deliver a terabit.
Luna: And that directly addresses the 'memory wall' problem — the gap between processor speed and memory bandwidth. If you can put photonic transceivers right next to the memory controller, you can feed the CPU or GPU data much faster.
Lucas: There's a fascinating paper from a team at Columbia University and IBM that demonstrated a complete on-chip photonic transceiver in a 45nm CMOS process. They integrated the laser, modulator, and detector all on one die. It's not production ready yet — the laser efficiency still needs work — but it shows the path.
Luna: And what about the packaging? Because the fiber alignment is notoriously difficult. You have to couple light from a fiber that's 9 microns in core diameter into a waveguide that might be 0.5 microns. That's sub-micron alignment tolerance.
Lucas: That's one of the biggest manufacturing hurdles. But solutions are emerging. Some use grating couplers that diffract light vertically into the chip, which relaxes alignment. Others use lensed fibers or even direct edge coupling with automated alignment. Ayar Labs has a 'SuperNova' light source that couples through a fiber array with active alignment.
Luna: So it's not just a science project — there are real roadmaps to volume production. When do you think we'll see the first mass-market chip with on-chip photonics?
Lucas: I'd say within three to five years for data center applications. TSMC's platform is already in early customer qualification. For consumer devices — laptops, phones — it'll take longer, maybe a decade. The cost and power budget need to come down further.
Luna: But the trend is clear. We're hitting physical limits with copper. The data demands of AI and cloud computing are accelerating the shift. And on-chip photonics is the most promising path to keep the data moving.
Lucas: Exactly. It's one of those technologies that's been 'five years away' for the last decade, but now the pieces are finally falling into place. The foundries are invested, the startups are shipping, and the customers are ready. I think 2027 is going to be a landmark year for optical I/O.
Luna: I hope so. It's a great reminder that the biggest innovations in computing aren't always about faster transistors — sometimes they're about how those transistors talk to each other.
Lucas: Well said, Luna. And on that note, we'll leave it there for today. Thanks for listening.

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.