Nikhil Balram | Mojo Vision: Should interconnect go narrow and fast, or wide and slow with 53,000 LEDs at 2 Gbps each?
16:10 - 18:10
Other snippets from this talk
Summary of the clip:
Should interconnect go narrow and fast, or wide and slow with 53,000 LEDs at 2 Gbps each?
The interconnect debate has borrowed communications thinking, but compute took a different route. CPU speeds plateaued around five gigahertz, and GPUs were even lower; instead of narrow and fast, they went massive, very wide and parallel. That raises the question: if interconnect goes wide and slow, how wide can it go—100 lanes, 1,000 lanes, 10,000 lanes? A traditional CPO chiplet shows the narrow-and-fast extreme: sixteen indium phosphide lasers outside the package, modulators, and waveguides moving light in and out.
The alternative is a micro-LED array: 1.3-micron LEDs at 1.8-micron pitch, giving 53,000 LEDs in 0.15 square millimetres. Each LED can toggle fast enough for two gigabits per second. This is wide and slow, running in the gigahertz range—the natural speed of compute—and working on CMOS. The theoretical bandwidth reaches 0.7 petabits per square millimetre, claimed to be 700 times what the narrow-and-fast approach can achieve, though it remains an asymptote.
The contrast is architectural. Left side: sixteen external indium phosphide lasers at very high speed, modulators and waveguides inside a chiplet packaged with a GPU—a very complex system. Right side: a display-style array where every LED is a lane. The approach trades per-lane speed for lane count, matching compute's parallel scaling instead of communications' narrow-and-fast scaling. It asks how wide wide-and-slow can become, and whether thousands of slow lanes can beat a few fast ones.
In this short video, you can learn:
* CPUs plateaued near five gigahertz and GPUs lower, so compute scaled wide and parallel rather than narrow and fast.
* A traditional CPO chiplet uses sixteen indium phosphide lasers outside the package plus modulators and waveguides, making a complex high-speed system.
* A 1.3-micron LED array at 1.8-micron pitch packs 53,000 LEDs into 0.15 mm², each toggling at 2 Gbps for 0.7 petabits per mm² theoretical bandwidth.
📋 **Clip Abstract** Compute escaped the narrow-and-fast path by scaling wide and parallel after CPU speeds plateaued near five gigahertz. A micro-LED alternative puts 1.3-micron emitters at 1.8-micron pitch—53,000 LEDs in 0.15 mm², each capable of 2 Gbps—toward a theoretical 0.7 petabits per mm², versus a traditional CPO chiplet with sixteen indium phosphide lasers, modulators and waveguides.
About the speaker:
* Speaker: Nikhil Balram
* Company: Mojo Vision
* Event: Eindhoven 2026
* Location: High Tech Campus, Eindhoven
#MicroLEDArray, #IndiumPhosphideLasers, #CPOChiplet, #PetabitsPerMillimeterSquared, #OpticalInterconnect, #MicroLED
This is a highlight of the presentation:
From Pixels to Petabits: A Wafers-In, Wafers-Out Micro-LED Platform for AI
More Highlights from the same talk.
06:08 - 07:38
How does one wafer-scale CMOS, micro-LED and lens stack serve both microdisplays and optical interconnect chiplets?
How does one wafer-scale CMOS, micro-LED and lens stack serve both microdisplays and optical interconnect chiplets?
The display architecture that Mojo Vision started building 11 years ago places micro-LEDs on a designed CMOS back plane, hybrid-bonding tiny LEDs on top, converting some to red and green, then adding lenses to couple light into optics. Because the goal was scaling, the flow was built as a wafer-in, wafer-out process. Optical interconnect, described as a tightly coupled system problem, turned out to map very nicely onto this same stack.
On the optical chiplet side the CMOS changes to a very advanced node with an electrical interface to other parts of the system. Micro-LEDs are still required, though their tuning differs while much of the structure stays the same. Lenses then couple into fibers. Because data must travel the other way, silicon photodetectors are needed; with blue, foundries already offer very mature, robust silicon photodetectors.
The same manufacturing flow was extended to support both thrusts. On the display side there is a front plane and a back plane, which must be brought into an optical system. Careful adjustments are required, yet the basic principles carry over. That reuse is the serendipity: an architecture designed for microdisplays now supplies the CMOS, micro-LED, lens and photodetector building blocks for wafer-scale optical interconnect chiplets.
In this short video, you can learn:
* A wafer-in, wafer-out process built for microdisplay manufacturing was extended to cover optical interconnect chiplets as well.
* The interconnect chiplet keeps micro-LEDs and lenses but shifts the CMOS to a very advanced node and adds silicon photodetectors for return-path data.
* Blue micro-LEDs pair with mature, robust silicon photodetectors already available from foundries.
📋 **Clip Abstract** Mojo Vision's 11-year-old microdisplay architecture — CMOS back plane, hybrid-bonded micro-LEDs, lenses — maps onto optical interconnect chiplets. The same wafer-in, wafer-out manufacturing flow was extended to both, with an advanced-node CMOS, retuned micro-LEDs, fiber coupling and silicon photodetectors for the return path.
About the speaker:
* Speaker: Nikhil Balram
* Company: Mojo Vision
* Event: Eindhoven 2026
* Location: High Tech Campus, Eindhoven
#WaferInWaferOut, #HybridBondedMicroLED, #SiliconPhotodetectors, #BlueMicroLED, #OpticalInterconnect, #MicroLEDDisplays
12:40 - 15:08
Can quantum dots designed for milliwatt TV flux survive 10 watts per square centimetre in AR pixels?
Can quantum dots designed for milliwatt TV flux survive 10 watts per square centimetre in AR pixels?
Television quantum dots are built for milliwatt-per-square-centimetre flux, but AR pixels need many watts per square centimetre—a factor of 1,000. Three years ago, a 0.2-micron red quantum-dot thin film was pumped with blue light at four watts per square centimetre and 25°C for 500 hours. Off-the-shelf commercial dots, tested in the same chambers, showed a degrading line. That gap drove years of work on highly reliable quantum dots.
Customer spec is typically 1,000 hours at 10 watts per square centimetre and 50°C. Because waiting 41 days per experiment is impractical, testing accelerates to 65°C and 25 watts per square centimetre, with equations extrapolating back. Sophisticated internal testing capabilities and rapid DOEs support this, developed closely with customers. The target is passing customer requirements, not merely collecting thin-film data.
The third change moved from thin films to pixelated quantum dots. In the product configuration—VGA pixels—testing presented at SPIE found L80, the 80% level, between 1,100 and 1,200 hours at 50°C. That passes customer requirements. A mock-up shows a back plane and flash, but no actual chip. The result is measured in the pixel configuration AR displays need, not just in test films.
In this short video, you can learn:
* Commercial TV quantum dots degrade under AR-level flux, which can be 1,000 times higher than their milliwatt-per-square-centimetre design point.
* Accelerated testing at 65°C and 25 W/cm², with extrapolation equations, avoids waiting 41 days for 1,000-hour results.
* Pixelated VGA quantum dots reached L80 between 1,100 and 1,200 hours at 50°C, passing the typical 1,000-hour, 10 W/cm², 50°C customer spec.
📋 **Clip Abstract** Quantum-dot reliability is the key barrier for AR pixels, where flux demands can be 1,000 times beyond TV quantum dots. By moving from thin films to pixelated VGA quantum dots and using accelerated 65°C, 25 W/cm² testing, L80 between 1,100 and 1,200 hours at 50°C was achieved, passing customer requirements.
About the speaker:
* Speaker: Nikhil Balram
* Company: Mojo Vision
* Event: Eindhoven 2026
* Location: High Tech Campus, Eindhoven
#QuantumDotReliability, #AcceleratedLifetimeTesting, #PixelatedQuantumDots, #L80Lifetime, #MicroLED, #ARDisplays
09:27 - 10:53
Why is a 69-gram, bulky AR display still built on 2012-era LCOS, and what would replace it?
Why is a 69-gram, bulky AR display still built on 2012-era LCOS, and what would replace it?
The engineered display works, but the system remains bulky: the temple is really bulky and it weighs about 69 grams. That makes it heavy, bulky and expensive. Astonishingly, it uses LCOS. Google Glass came out in 2012 and used LCOS, so 14 years later the same old, mature approach persists. It is very reliable, but mature. For the attractive form factor shown previously, this legacy engine is the bottleneck.
To reach that desirable form factor, the conference view is that a single-panel RGB microLED is required. The listed advantages are obvious, and a single-panel microLED enables a tiny optical engine. The one designed here is less than 0.2 cc, so it can support the needed form factor. That is the proposed replacement path for the bulky LCOS system, moving from a mature, reliable component to a much smaller RGB microLED engine.
Commercial scale is framed by Meta's positive projection from SPI earlier this year: 100 million units. The stated goal was to achieve 100 million units, with a price target of $100 per eye. That combination is very attractive. The open issue is execution: how do you go about doing this? The clip ends at that question, after establishing single-panel RGB microLED and a sub-0.2 cc engine as the route to the desired form factor.
In this short video, you can learn:
* The current display is about 69 grams, bulky in the temple and expensive, despite using reliable LCOS.
* A single-panel RGB microLED is presented as necessary for the attractive form factor, enabling an optical engine under 0.2 cc.
* Meta projected 100 million units with a $100 per-eye price target, leaving the execution question unresolved.
📋 **Clip Abstract** The clip contrasts a bulky, 69-gram display using mature LCOS with the required single-panel RGB microLED path. It cites a less-than-0.2 cc optical engine and Meta's 100-million-unit, $100-per-eye projection, then asks how to execute it.
About the speaker:
* Speaker: Nikhil Balram
* Company: Mojo Vision
* Event: Eindhoven 2026
* Location: High Tech Campus, Eindhoven
#SinglePanelRGBMicroLED, #LCOSDisplay, #Sub02ccOpticalEngine, #69GramDisplay, #ARDisplays, #MicroLEDDisplays