中文 English

Making Light More Reliable

Silicon photonics is a promising technology, but it may take a while.

popularity

The buzz around photonics in packages and between packages is growing. Now the question is whether it will work as expected, and where it will be useful.

Replacing electrical with optical signals has been on the technology horizon for some time. Light moves faster through fiber than electrons through copper. How much faster depends upon the diameter of the wires, the substrate and interconnect materials, and the distance that signals need to travel. So while it may take longer for an electronic signal to move from one side of a large chip to another, the difference may be negligible if that signal only needs to move across a direct-bond connection within a package.

Where optical really excel is over large distances, such as between server racks and racks of solid-state drives that may be on another floor or even in another building. Many of the trunk lines for data and voice communications are fiber optic, and they have been in use for decades.

The upsurge in interest for photonics at the chip level is due to diminishing returns for scaling and the persistent and growing impact of power — dynamic and leakage — on performance. Light is extremely fast, relatively secure (light is more difficult to hack than electrical signals), and enough is known about its properties to manage it effectively.

Moreover, light itself uses very little power. It’s the conversion from electrical to optical and back to electrical that consumes most of the energy. As of today, there is no way to store data optically. It needs to be converted to electricity.

Until very recently, most of the optical couplers looked like large boxes. They were flimsy, poorly attached, and made silicon photonics look like something out of the early days of electronics. But opto couplers have improved significantly over the past year. Researchers at the University of Twente in The Netherlands have designed a much more appealing connector, making it look like an advance in electronics rather than a step back to the days when every electrical engineer needed a plug for a soldering iron.

That’s a good start. But for these devices to have mass commercial impact, they need to be reliable. Photons can get banged around by line-edge roughness in waveguides the same as electrical signals, and the processes for manufacturing waveguides in large volume don’t exist today. This is why many companies that have been working with photonics have been using silicon interposers to guide optical signals between chips. It’s easier to bond a silicon interposer to a silicon chip because there is no concern about differences in coefficients of thermal expansion, which creates stress in advanced packages.


Fig. 1: 300mm silicon photonics wafer.
Source: Wikimedia Commons  Ehsanshahoseini

Optical has other issues, as well. Reliability, for example, depends upon the ambient temperature around the optical signals. So if a chip nearby is on all the time and heating up the silicon photonic laser or the waveguide, it can skew the signal beyond what an optical filter is designed for. The result could be data loss or an incomplete transmission. Today, there is no standard way to test for all of that because there is no way to know how or where those devices will be used. Results would vary greatly if the same device was used in a car or a cell phone.

Finally, the light sources tend to degrade over time. Aging in silicon depends upon a variety of factors, ranging from heat to impurities in the silicon or latent manufacturing defects, leaving many question marks about how well devices will behave over time. But aging in III-V materials, which are typically used for the lasers, is nowhere near as well documented as silicon, and test and inspection coverage is an iffy prospect with any new technology.

Companies that have been working with silicon photonics tend to focus on higher-margin, extremely high speed applications such as networking chips. But for this technology to really take off, chipmakers and systems companies need a much broader set of tools, a better way to track the performance of these devices, as well as much stronger industry support and investment.



Leave a Reply


(Note: This name will be displayed publicly)