Electrical signalling has beaten off the challenge from optical communications a number of times in electronic systems, but connecting by light may be about to stage a comeback.
The electronics industry is good at spawning 'technologies of the future' that seem destined to remain forever in the future. Optical signalling joined the list several decades ago. Although it forms the backbone of the Internet, photons have not displaced electrons for shorter distances; attempts to do so have often been met with reversals.
Bob Blake, European marketing manager at programmable-logic supplier Altera, recalls the first attempt at introducing optical technology to large computers and telecom switches. Earlier in his career, he worked at Southampton-based BICC-Vero. The company was a major supplier of printed circuit-boards (PCBs) and backplanes for large modular computers and telecom switches. In the 1990s, it seemed that the electrical signalling used across those boards was running out of 'runway'.
The so-called 'skin effect' was the problem. High-frequency electrical signals tend to move to the surface of a wire, effectively constraining them to a much narrower path than the physical width of a PCB trace would suggest. The narrower the path, the higher the resistance to the flow of electrons. The net effect of this parasitic resistance is that the bits are smeared into each other, making it progressively harder for the receiver to work out what data was meant to be sent.
Even at the relatively short distances inside a computer – usually a lot less than 3m from point to point within even a rack-based mainframe – the maximum data rate per connection looked to be limited to hundreds of megabits-per-second. Electrical backplanes had sprouted thousands of connections in an attempt to overcome the limit, but were running out of room to fit them. Although connector technology evolved to pack more pins into a smaller space, the physics ofinterference took hold. To stop adjacent lines from disrupting each other every time a bit was transmitted, the connector manufacturers inserted more and more ground pins.
The optical option
The alternative was to shove the data through a smaller number of high-speed connections. As electrical signalling had apparently topped out, optical seemed to be the way to go, as photons are troubled by a different set of physical considerations to those of electrons. 'We talked a lot about optical backplanes then,' recalls Blake; but very few customers moved away from electrical.
Unfortunately, the components needed to convert from electrical to optical signals were expensive, and the connectors tricky to make. Electrons may suffer from problems such as the skin effect; but as long as they have a conductive path to follow, they will follow it. Optical connectors require precise alignment and tuning, which all adds to the cost.
The answer to higher speeds came through modifying the electrical signal. Boosting the energy of the high-frequency portions of the electrical signal made it easier for the receiver to spot the transitions between bits. Progressively more complex forms of equalisation made it possible to undo the effects of high-frequency smearing. The result was a boost to the datarate, pushing far beyond 1Gbps.
Speeds have pushed to 6Gbps and beyond. The limitations of FR4 – the fibreglass used for most PCBs – are beginning to show again, however. Blake says the realistic top speed of FR4 using current pre-emphasis and equalisation techniques is around 8Gbps. More advanced substrate materials such as Megtron 6 can support higher speeds but 'that is five times the cost of FR4', according to Blake. The on-chip transceivers that Altera has designed for its FPGAs can work, in principle, beyond 20Gbps.
'[The problem is that] at 28Gbps we really have run out of steam with electrical,' reports Blake. Altera's response has been to work on a way to bring optical interfaces into its FPGA devices. The company aims to have demonstration chips ready by the end of 2011. There is still room for improvement in electrical signal processing. A 2008 paper in IEEE Transactions on advanced packaging by Rambus engineers Wendemagegnehu Beyene and Amir Amirkhany concluded that it should be possible to push data rates to 20Gbps and beyond with better equalisation and a move to the kind of multilevel signalling used by modems. A year later, Dong Kam and colleagues from IBM's TJ Watson laboratory agreed that it should be possible to get to 25Gbps albeit at distances of less than a metre; but there was a cost: power. Above 10Gbps the energy needed to support all the signal processing climbs dramatically.
Kam and colleagues estimated that it would take 22mJ/Gb on a 10Gbps interface, about 70 per cent higher than an equivalent optical link; but the power at 20Gbps would increase to 40mJ/Gb, more than double its optical equivalent; and the optical link would be far less constrained in terms of cable length.
Despite the increase in energy and complexity, companies are still hesitant to move to optical signalling. In a demonstration for CES 2010, Jason Ziller, director of Thunderbolt technology marketing at Intel, claimed: 'Electrical cables are reaching their limits in terms of how fast you can go and how long those cables can be. Consumers want smaller, thinner devices. But the proliferation of connectors, especially large electrical connectors is actually impeding that trend.'
Intel's answer was a move to optical cabling. Light Peak, as it then was named, had a projected launch data rate of 10Gbps. But a little over a year later, Light Peak had lost the light and moved back to electrical. Improved signal processing made it possible to deliver 10Gbps over a copper cable and cost less than an equivalent optical module. To deliver 100Gbps, Thunderbolt will almost certainly have to shift to optical transmission. But for the moment it will have to wait for that migration.
Although it is likely to be more expensive, especially in the early days, optical provides a route to get to 12Gbps and higher without putting a large dent in a data centre's electricity bill. Optical offers the possibility to scale to more than 100Gbps per channel between chips and across backplanes in communications switches using the same techniques as those used by international communications links, albeit on a much smaller scale.
One big advantage of using photons instead of electrons is that they do not interfere with each other. Using different wavelengths of light, you can send multiple signals in parallel down the same length of fibre or waveguide. This technique, called wavelength division multiplexing (WDM), is used for the fastest parts of the Internet's backbone. There is no reason why it cannot be scaled down to the chip level – the laser diodes need be no larger than a few micrometres across.
The bad news is that silicon is far from being the world's greatest photonic material. The semiconductor is good for passive control of optical signals so has been used by manufacturers such as Oclaro for chip-scale waveguides. But the lack of a direct bandgap makes it a poor choice for building lasers. Intel and others have pursued research into the possibility of making lasers out of silicon.
The key to these devices is the Raman effect. In the mid-infrared region, where the material is mostly transparent, silicon is 10,000 times better than glass at demonstrating the Raman effect. Normally, when photons are scattered after striking the atoms of a material, they keep their original wavelength. But that is not always the case. In the late 1920s, Chandrasekhara Venkata Raman and Kariamanickam Srinivasa Krishnan found that sometimes the collisions result in the photons transferring energy to or from the atoms in the material. This transfer of energy leads to a change in wavelength, producing photons of a different colour. It is possible to concentrate and amplify this'light, resulting in a type of laser that does not rely on population inversion.
Prototypes on show
Researchers from the University of California in Los Angeles (UCLA) demonstrated the first silicon laser based on the Raman effect in 2004. Intel followed with a continuous-wave design in 2005 that the developers thought would suit sensing applications.
There is a drawback: the basic Raman laser needs an external laser to act as a pump, making it impossible, at least today, to have a practical all-silicon laser. Multiple Raman lasers could be driven from a single, off-chip laser diode but lasers based on indirect bandgap materials such as indium gallium arsenide would offer higher efficiencies.
For inter-chip communications, Intel has looked at inserting III-V materials into a silicon substrate – at the wafer level to reduce cost – using WDM to increase the datarates achievable without massively increasing the cost of the optics.
Last year, the chip giant demonstrated a 50Gbps inter-chip link as a spin-off from its Light Peak work. There are other ways to achieve a similar end. Increased interest in chip-stacking (E&T vol 6, issue 9, p82, http://bit.ly/ET-3DIC) among manufacturers promises to bring down the cost of putting laser-bearing chips on top of a CMOS logic part.
'We believe that, in the medium term, we should start to integrate optical interfaces into the [chip] package,' says Blake. 'The laser industry is starting to become more focused on consumer applications and the cost of doing this is coming down; but we do not expect this to be used for every application.'
Eric Beyne, director of packaging research at Belgian R&D institute IMEC, reckons the trend towards optical will be helped by a separation of I/O and core logic inside the chip package. It is getting more expensive to build large electrical I/O transceivers on advanced logic processes.
So it will ultimately make sense, in his view, to have them on separate chips that are then stacked or placed side-by-side on a silicon substrate. Says Beyne, 'Ultimately this I/O device will be an optical link to other modules in the system'.
Questions remain about how quickly optical can become a solution for metre-scale interconnects where technology quickly see-saws from one contending technology to the other. Optical's prospects are looking up but could easily be delayed by another breakthrough in the efficiency of electrical signal processing. *