Bernard Kress | Google: How can the microLED display manufacturing ecosystem be repurposed to capture the rapidly growing AI co-packaged optics market?
21:30 - 23:55
Other snippets from this talk
Summary of the clip:
How can the microLED display manufacturing ecosystem be repurposed to capture the rapidly growing AI co-packaged optics market?
Off-the-shelf micro-LEDs developed for AR cannot be used in CPOs as active-matrix backplanes are incompatible with gigahertz data links. But the micro-LED manufacturing ecosystem—300-millimeter wafer-to-wafer hybrid bonding, sub-5-micron GaN emitters and wafer-level optics—can be directly repurposed for high-density CPO architectures. Photonics supports scale-out with pluggable optics and scale-up with co-packaged optics inside racks from GPU to GPU. CPO avoids copper’s power requirement for ultra-low latency and ultra-high bandwidth.
Three CPO architectures are described. A continuous-wave laser with an external laser source and PICs that modulate via Mach-Zehnder or microring resonators. A VCSEL array with direct current modulation, needing no external modulator or PIC. And micro-LED arrays at direct logic level, but not with existing display backplanes. Those backplanes must change for one-gig, two-gig, maybe even four-gig NRZ per LED. Micro-LEDs offer massive parallelism: wide and slow rather than fast and narrow like external laser diodes.
Micro-LEDs show great thermal tolerance and less than 0.5 picojoule per bit, better than VCSEL and external laser diode. Standard CMOS silicon photodetectors work. Blast rate is also negligible because operation is massively parallel. The comparison favors micro-LED arrays for direct logic-level drive, massive parallelism and thermal tolerance, while external laser diodes are fast and narrow. The existing AR display backplane remains the blocker, but the manufacturing ecosystem itself maps onto high-density CPO.
In this short video, you can learn:
* Off-the-shelf AR micro-LED panels cannot serve CPOs because active-matrix backplanes are incompatible with gigahertz data links.
* The micro-LED manufacturing ecosystem—300-millimeter wafer-to-wafer hybrid bonding, sub-5-micron GaN emitters and wafer-level optics—can be repurposed for high-density CPO architectures.
* Micro-LED arrays offer direct logic-level drive, massive parallelism, great thermal tolerance and less than 0.5 picojoule per bit, but need new backplanes for one-, two- or four-gig NRZ per LED.
📋 **Clip Abstract** Micro-LED panels built for AR cannot be dropped into AI co-packaged optics because their active-matrix backplanes do not support gigahertz data links. Yet the underlying manufacturing ecosystem—300-millimeter wafer-to-wafer hybrid bonding, sub-5-micron GaN emitters and wafer-level optics—can be repurposed for high-density CPO, where micro-LED arrays promise direct logic-level drive, massive parallelism, thermal tolerance and less than 0.5 picojoule per bit.
About the speaker:
* Speaker: Bernard Kress
* Company: Google
* Event: Eindhoven 2026
* Location: High Tech Campus, Eindhoven
#WaferToWaferHybridBonding, #Sub5MicronGaNEmitters, #WaferLevelOptics, #PicojoulePerBit, #SiliconPhotonics, #OpticalInterconnect
This is a highlight of the presentation:
Purpose-Driven Pixels: Aligning Display Tech with XR Use Cases
More Highlights from the same talk.
09:34 - 11:50
Why should smart glasses include a display when audio-only models are already a massive commercial success?
Why should smart glasses include a display when audio-only models are already a massive commercial success?
Close to 10 million glasses were sold in 2025, and more than 95% were audio only. People love it. So why include a display? The first reason is to avoid cognitive overload. With audio only, an AI agent can talk while you are talking to someone, and you lose everything. An audio feed plus a display feed lets you process both at the same time.
The second reason is information remanence. Audio comes in and comes out. If there is noise, or you did not get it, too bad. A display can have remanence of one second, maybe two seconds, maybe even three seconds, giving you time to process information from the AI assistant when you need it. Third, what makes display amazing is world locking: not color, resolution or brightness.
World locking is 6DoF, not head tracking and not so much body locked. HoloLens 1, HoloLens 2, Magic Leap 1 and Magic Leap 2 showed this, but with a whole range of cameras. These were headsets, so they had four cameras and two IMUs. Today, smart glasses have one camera and one IMU. AI helps provide decent 6DoF world-locking with one camera and one IMU, making the display really amazing—for example, an AI assistant pointing at something in your reality.
In this short video, you can learn:
* Close to 10 million glasses sold in 2025, and more than 95% were audio only.
* A display avoids cognitive overload by allowing simultaneous audio and visual processing, and it gives information remanence of one to three seconds.
* World locking, not color, resolution or brightness, is what makes display amazing, and AI can provide decent 6DoF with one camera and one IMU on smart glasses.
📋 **Clip Abstract** Audio-only smart glasses sold close to 10 million units in 2025, with more than 95% audio only, so a display must justify itself. The argued reasons are avoiding cognitive overload by processing audio and display feeds together, providing one-to-three-second information remanence, and enabling 6DoF world locking with one camera, one IMU, and AI.
About the speaker:
* Speaker: Bernard Kress
* Company: Google
* Event: Eindhoven 2026
* Location: High Tech Campus, Eindhoven
#AudioOnlyGlasses, #CognitiveOverload, #InformationRemanence, #WorldLocking6DoF, #SmartGlasses, #XRDisplays
15:40 - 17:25
What are the primary manufacturing processes for surface relief grating waveguides, and how do they trade off cost and performance?
What are the primary manufacturing processes for surface relief grating waveguides, and how do they trade off cost and performance?
Today's single SRG waveguide options mainly use imprint on 2.0 glass, with 1.9 or 2.0 index resin, plus an ALD cover for protection. Additive processing uses nanoimprints in the lowest possible low-index resin—not sacrificial—then a fast ALD cover. Fast ALD now grows 45 nanometers in an hour, which crashes production costs compared with a few years ago.
The subtractive process is a direct etch and produces today's best results. But efficiency versus uniformity is always a balance: increasing efficiency decreases uniformity, and increasing uniformity decreases efficiency. Direct etch has the best result, yet it is also the most expensive. So superior performance does not mean glass makers will choose it, because they must squeeze margin out. Hardware is not easy.
Multiple waveguide approaches exist: surface relief, holographic, geometric reflective, glass, plastic. For single SRG waveguides, the main route remains imprint on 2.0 glass with 1.9 or 2.0 index resin, protected by a little ALD cover. An additive route uses nanoimprints in low-index resin and fast ALD. A subtractive route uses direct etch and gives the best results but at the highest cost. Today smart glasses have many display architectures, including reflective visors such as NGO on every side.
In this short video, you can learn:
* Fast ALD can now grow 45 nanometers in an hour, crashing the cost of additive SRG waveguide production.
* Direct-etch subtractive processing delivers the best efficiency-uniformity result but is the most expensive route.
* Increasing SRG waveguide efficiency reduces uniformity, and increasing uniformity reduces efficiency, so glass makers must balance performance against margin.
📋 **Clip Abstract** Surface relief grating waveguides are made by imprinting on 2.0 glass with 1.9 or 2.0 index resin and an ALD cover, by additive nanoimprint in low-index resin with fast ALD, or by subtractive direct etch. Direct etch gives the best efficiency and uniformity but costs the most, while fast ALD growing 45 nanometers per hour has crashed additive costs, leaving glass makers to balance performance against margin.
About the speaker:
* Speaker: Bernard Kress
* Company: Google
* Event: Eindhoven 2026
* Location: High Tech Campus, Eindhoven
#SRGWaveguide, #FastALD, #DirectEtch, #LowIndexResin, #ARWaveguides, #XRDisplays
13:35 - 15:47
How can microLED display designers solve the Etalon match to maximize waveguide efficiency?
How can microLED display designers solve the Etalon match to maximize waveguide efficiency?
Every light engine—LCOS, LDS, laser beam scanning—and microLEDs face the same limit: the Etalon match between light engine and waveguide sets display-system efficiency in nits per watt, not nits per nits or nits per lumen. The transcript treats this match as the defining problem for microLEDs. A full-color monolithic microLED with microlens arrays is not a good idea because multiple cones emerge, so the source remains Lambertian and loses a lot of light. Light dumps become necessary.
Stacked microLEDs were done more than 10 years ago. You can stack microlenses, wave level optics microlenses, but they are stacked: three different cones, at least pointing in the same direction. Pixels then have more or less the size of a pixel interspacing. Put the microlens on or leave it off—it does not really change anything; emission remains Lambertian. That structure cannot create the directional cone needed to match the waveguide.
The best solution is pixels much smaller than the pixel interspacing. There, with a microlens array, you can more or less make the emission cone and match the Etalon view, yielding much more efficient glasses that last longer for all-day use. A waveguide combiner is key to a thin see-through stack. Many types exist: surface relief, holographic, geometric reflective, and others. This combination targets the nits-per-watt bottleneck rather than raw brightness.
In this short video, you can learn:
* The Etalon match between light engine and waveguide determines display-system efficiency in nits per watt, not nits per nits or nits per lumen.
* Full-color monolithic microLEDs with microlens arrays remain Lambertian because multiple emission cones emerge, and stacked microLEDs with stacked microlenses still produce three cones.
* The best route is pixels much smaller than the pixel interspacing with a microlens array to shape the emission cone and match the Etalon view for all-day-use glasses.
📋 **Clip Abstract** MicroLEDs share a core limit with LCOS, LDS and laser beam scanning: the Etalon match between light engine and waveguide sets efficiency in nits per watt. The best solution is pixels smaller than the pixel interspacing, using a microlens array to shape the emission cone and match the Etalon view for more efficient all-day-use glasses.
About the speaker:
* Speaker: Bernard Kress
* Company: Google
* Event: Eindhoven 2026
* Location: High Tech Campus, Eindhoven
#EtalonMatch, #NitsPerWatt, #PixelInterspacing, #LambertianEmission, #MicroLEDDisplays, #ARWaveguides