Lithography has reached human engineering limits, yet cannot lock down microring chip wavelengths
Silicon photonics once deemed microrings unproducible due to wavelength drift. This company built a laser annealing tuning device that achieved 99% wavelength yield, circumventing dependence on cutting-edge lithography.
The argument · tap a timestamp to hear it
Compute growth far exceeds copper wire capacity
Over the past decade, global compute grew roughly 100 million times, nearly 10-fold annually, yet chip area grows only with the square of edge length, so the perimeter carrying data grows naturally slower. When bandwidth demand becomes 10x, 100x, or even 1000x greater, copper completely fails, and interconnect transitions for the first time from a computing byproduct to the core node determining AI cluster efficiency and power efficiency.
Microrings are two orders of magnitude smaller than MZI, enabling wide-and-slow design
Traditional Mach-Zehnder (MZI) modulators are roughly 3 millimeters; microring modulators around 15 micrometers in diameter—more than two orders of magnitude difference. MZI, constrained by edge layout, can only stack bandwidth by raising per-channel speeds; microrings, small enough to blanket an entire chip like sesame seeds, use many low-speed parallel channels instead, achieving higher total bandwidth density and lower power consumption.
Performance means nothing if it cannot scale to production
Microrings excel in the lab—pair them with a tunable laser and you get excellent transmitters. But at production scale, purchased laser wavelengths mismatch each microring's response wavelength, making mass manufacturing impossible. The advisor judged it clearly: the device performs well, you can obviously see that, but it cannot scale to production—the core industry concern is that microring wavelengths naturally drift, preventing all-in commitment to the approach.
— Zhang WeiweiLithography precision has reached its ceiling, yet cannot lock phase stability
Conventional approach: better lithography tightens line width to suppress wavelength drift. But even at 0.5 or 1 nanometer precision—the human engineering limit—such jitter stays perceptible to precision optics and causes phase shift. The solution: stop pursuing perfect lithography. Instead, separate line width control and wavelength correction into two steps: standard lithography first, then a dedicated device to tune each microring's phase.
— Zhang WeiweiTuning device achieves 99% wavelength yield
The device, named Zhiwei 1000E (智威1000E), uses tens of thousands of pulsed laser scans to trace amorphous silicon regions on the chip with picometer precision, pulling each microring's wavelength to target. On this single metric—microring wavelength—it achieved 99% yield. This means microring manufacturing no longer requires EUV or cutting-edge lithography, enabling mass production with mature process nodes.
— Zhang WeiweiCustom lasers at $3,000 per unit drove in-house light source development
Each microring needs one wavelength; eight-channel microring arrays theoretically require eight separately tuned narrow-linewidth DFB lasers, priced around $3,000 each. Eight lasers lined up generate cross-heating and cannot be solved by stacking power; splitting them damages performance. The team self-developed comb lasers, letting one device produce 8 or even 16 wavelengths directly, drastically reducing light source cost and complexity.
Electrical chips translate between GPUs and photonics
GPUs' built-in SerDes targets 50-ohm copper impedance at roughly 400-800 millivolts, but microrings are nonlinear, requiring 1.6-2V or higher driving voltage, specific reverse bias, and extreme temperature sensitivity for real-time wavelength locking—none native to GPUs. The team designed an electrical chip between GPU and photonic chip, translating electrical signals into a form photonics can read, completely transparent to the GPU.
OIO is not a better optical module but a new computing paradigm
The team explicitly rejected becoming the next Zhongji Xuechuang (中际旭创), since optical modules target communications, while scale-up interconnect comes from compute—maximizing utilization, minimizing latency at any cost. OIO will define a new paradigm for computing optical interconnect. Against Nvidia's closed-loop full-stack ecosystem (the team calls this the 'Apple ecosystem'), they chose the reverse: an Android ecosystem, opening the technology to all GPU makers worldwide.
In their own words · checked verbatim
I think if optical interconnect truly gets into GPUs, compute may no longer be the central node of an AI cluster—interconnect could become the central node of the entire cluster.
我觉得随着光互联如果真的做到了GPU里面去 很有可能计算已经不是整个AI集群的一个中心节点了 互联可能会变成整个AI集群的一个中心节点
The performance is clearly good, you can obviously see that, but it cannot scale to production.
他说这个东西性能虽然很好 你明显能follow 但是它没法被量产
Zhang Weiwei22:16
I said at that point: no matter how good a lithography tool you use, no matter what you do, you've reached the human engineering limit. But optical systems, wavelengths—these are naturally sensitive nodes. Your slightest variation is still perceptible to them.
我当时就是说那你在用最好光刻机不管你怎么做 你达到人类工程极限了 但是我光学光波长光学机器 它是天然的一个敏感气节 我就是能感受到你的微弱
Zhang Weiwei24:18
These narrowband light sources, single-channel DFB lasers, are priced at roughly $3,000 per unit.
就是说大概就是说精种光源 拔不通道的DF激光器零化源 它的价格大概是3000美金一颗卖给你
When Tesla was doing electric cars 20 years ago, no one else was doing motors, motor controllers, or batteries—you had to do it all yourself. You had to be integrated.
特斯拉当年在坐电动车的时候 20年前 没有人去做电机电控电池 那你必须都得做了 你要被整机
We're not trying to make a better optical module. OIO is meant to define a new paradigm for computing optical interconnect.
我们OIO并不是要做一个更好的光模块 而是OIO要定义一个计算光互联的新的范式
But we're willing to build an Android ecosystem. I want to open our technology to all such companies.
但是我们愿意做安卓生态 我希望把我们open给所有的这样的公司
Figures
| Compute growth over past decade | ~100 million times (roughly 10-fold annually) | 5:02 |
| MZI modulator size | ~3 millimeters | 15:09 |
| Microring modulator diameter | ~15 micrometers | 15:09 |
| Microring wavelength yield (after tuning) | 99% | 29:25 |
| Custom DFB laser unit price | ~$3,000 per unit | 33:33 |
| Company valuation | Just over $1 billion (unicorn) | 3:02 |
| Chip tape-out rounds | Photonic chip: 8 rounds; electrical chip: 5 rounds | 45:46 |
Glossary
- CPO
- Packages optical devices with switching chips together, primarily for switch-side optical interconnect.
- OIO
- Connects light directly to computing chip (GPU/CPU) pins, a new form of interconnect technology.
- UCIE
- Protocol standard for chiplet-to-chiplet interconnect, originally designed for distances spanning centimeters to millimeters.
- microring modulator
- Ring-shaped optical modulator only tens of micrometers in diameter, operating through wavelength resonance.
- MZI
- Traditional optical modulator achieving modulation through light path phase interference, millimeter scale.
- WDM
- Technique allowing multiple signals at different wavelengths to share one fiber simultaneously.
How to listen
Investors tracking AI compute infrastructure, semiconductor and optical communication engineers, and practitioners working to understand CPO/OIO technology roadmap debates.
Around 4:40-5:40 contains transcript noise, skip directly.