The Convergence of Photonic and Electronic ICs

At the recent DesignCon in Santa Clara, one of the keynotes was by John Bowers of UC Santa Barbara. I'll start with his summary from his last slide:

There is a silicon photonics revolution happening.

The challenge is that silicon is a terrible material for optics since it is an indirect bandgap material. On the other hand, the current approaches to interconnect are not scaling. Switches take a lot of energy to put data out there, and 100Gb on a copper pair is about the best we can do. Beyond that, we need to use multiple pairs. With photonics, this hasn't happened before since photonics has been so expensive. A single transceiver might cost $500, and you might need hundreds of them to build a big switch, so that is just cost-prohibitive. But increasingly, fiber is being used not just around the datacenter and between datacenters, but within rack.

Since you can't build lasers out of silicon, separate lasers are required, and the light modulation can be done on-chip. Inevitably, that means two (or more) separate components, driving up cost, power, and physical size.

Historically, photonics has been constructed with VCSELs (vertical-cavity surface-emitting laser, pronounced vee-sell), but the wavelength produced is relatively lossy. In a big datacenter a terabit can't be done with VCELs, so they are no longer being used. For datacenters, the most desirable light wavelength is 1310 nm, usually just called thirteen-ten.

A big driver, of course, is AI, but we need much more efficient datacenters since the power consumption is out of control. We can have more efficient components, but also, the more we can connect them, the more efficient they are. Optical interconnect is in the midst of a revolution that John says will result in optical being used not just for distances measured in meters but in millimeters too, and eventually on-chip. The basic challenge is that switch bandwidth is doubling every year, but transistor performance is not increasing that fast, which means multiple lanes. As a result, an increasing share of link power is consumed by the first and last 12".

The form factor is going through several phases:

• Pluggable
• 2.5D co-packaging of electronics and photonics
• 3D co-packaging (wafer on wafer)
• 3D packaging with integrated lasers

High capacity links have increased 32X in the last four years by taking advantage of these packaging technologies. Another change is that photonics used to be manufactured on 3" wafers in old fabs. Now it is done in modern 300mm fabs with the latest equipment. Not quite the latest, it doesn't need 7nm, but 45nm is perfect for 1310 light (roughly 1.5um). So, we have jumped eight generations of manufacturing equipment.

As a result, over just the last few years, the market has switched fast from Indium Phosphide PICs (PIC just stands for photonics IC), to GaAS VCSELs, and to true silicon photonics. By "true" I mean that the photonics wafers and the electronic wafers are packaged together. If you want to see the last time I wrote about photonics, see my posts How to Design Photonics If You Don't Have a PhD: iPronics and Ayar Labs and DesignCon: The Future of Fiber Optic Communications. They were written just two years ago, but a lot seems to have changed since then.

With efficiency has come scale. Units were literally shipping in hundreds just a few years ago, and now they ship in tens of millions (Intel have said they are doing 6M units per year).

Another significant gain comes from the increased performance. If you have a 1.6Tbps transceiver, it turns out to be small and you don't need many of them, so you can put enough around a chip to build a 51Tbps switch.

The dream is to build lasers directly on the silicon by epitaxially growing III/V materials on the silicon substrate. But there are various mechanical issues. You tend to end up with a lot of defects, the devices do not last very long, and they degrade rapidly. People have been trying to do this for 40 years but recently there have been some breakthroughs. Until recently, the best was to last 400 hours, which is useless for commercial purposes. But with quantum dots (QD), this goes up to millions of hours.

Perhaps surprisingly, lasers built on silicon are better than native substrate lasers because lost with Si, SiO₂, and Si₃N₄ is so much lower than InP or GaAs waveguides.

The status:

Lasers epitaxially grown on Si will be quantum dots. We are not quite there yet but this is the 6th generation we are now working on. Quantum dots are insensitive to defects, insensitive to reflections. They are really small devices. And quantum dots work really well at 1310 for datacenter applications.

And we end where we came in:

There is a silicon photonics revolution happening.

Sign up for Sunday Brunch, the weekly Breakfast Bytes email.

.