Rob Coppinger wanders the periodic table for fast-as-light computing
Nothing is faster than light and computational speeds are progressing to the point where the metal circuits that link microprocessors are holding back what can be done. To solve this problem photonics has the answer: it is referred to as silicon photonics. Recently the subject of a European Union project called Helios, researchers want photonic microelectronics to be as ubiquitous as sunlight.
‘The first 50 gigabit per second (gbps) modulator in silicon ever, the first 40gbps modulator in silicon with a large extension ratio and good hybrid laser work by other groups – we had quite a number of technical firsts in Helios and in particular on the modulator, which we’ve been involved in,’ says Professor Graham Reed. Reed is based at the University of Southampton’s faculty of physical and applied sciences. Because computers operate on silicon and its mass production is a mature technology, the challenge has been to get a material that can be compatible with silicon and will lase – because silicon will not. ‘You have to do better than metal in terms of power consumption,’ adds Reed. Industry also wants these fast-as-light computers to be low-power. Reed explains: ‘Green photonics has to be as power efficient as possible. Oracle wants to do better than femto joules per bit. Others will argue for sub-1 pico joule per bit.’
This lower power, along with lower costs, is a major justification for the idea of silicon photonics and the main driving force for it has been the large semiconductor companies. Intel, IBM, Oracle; these large corporations want to include photonics on chip circuits to interconnect different technologies.
But the chips themselves can’t be lasers, as Reed explains: ‘Silicon is not a very good material for light emission.’ The problem with silicon is that it has an indirect band gap that makes fabrication of a laser unlikely and it has a structure that means it has a linear electro-optic coefficient. While silicon is not good for lasers, Reed points out that some have ‘found ways to make it just good enough for an application but not good enough for long haul [telecommunication] applications’.
But attempts to make true lasers from silicon have not worked. Reed explains that, to solve this, elements from specific parts of the periodic table are used – the III V elements. These are from the third and fifth groups of periodic element tables. Indium, from the fifth group, and phosphide from the third group are often used. Silicon is a group four material.
Reed says: ‘What people have done is introduce some III V elements and either bonded it or just glued it on to the structure. They used that III V as the gain part of the system that is producing a laser.’ As such silicon photonic on chip lasers are hybrid lasers. ‘In group 3-5 you can make a very efficient solid state laser and when you have a laser on a silicon chip when you have a laser on a chip it tends to be a hybrid 3-5 laser,’ adds Reed. But there are problems with III V elements. They are dopants for silicon, so they effect the electronic properties of silicon. ‘You don’t want to put III V elements in a silicon foundry because you will contaminate it,’ says Reed. To avoid this, the III V elements are added to the silicon at the very end or the start of its production process.
The solution according to Reed is to use Germanium. ‘MIT is trying to make a Germanium laser and that is a compatible material. Germanium is a group four material and is more compatible with silicon.’ Another advantage of Germanium is that it has a relatively long wavelength bandgap of 1.6 microns.
Researchers have succeeded in making light emitting bridges of germanium that could enable microprocessor components to communicate using light. A manufacturing technique to render the semiconductor germanium laser-compatible through high tensile strain has been developed by researchers at Zurich’s Paul Scherrer Institute and the Politecnico di Milano. The scientists demonstrated that they can use their method to alter the optical properties of germanium, which is typically unsuitable for lasers. Martin Süess is a doctoral student at the Laboratory for Nanometallurgy at Eidgenössische Technische Hochschule (ETH) Zürich. He said: ‘With a strain of three per cent, the material emits around 25 times more photons than in a relaxed state.’ Microchips that can communicate by light instead of copper integrated circuits can be the foundation for much faster computers. In order to bring the germanium into a laser-compatible stretched form, the researchers use the slight tension generated in germanium when it evaporates on silicon. Exposed germanium strips that remain attached to the silicon at both ends experience strain between those two ends. That strain becomes so intense that it becomes laser-compatible. Whether it is Germanium or a III V element, adding lasers to microchips increases the level of heat.
‘Those microchip cores are ruling at very high temperatures; if you put a laser on that chip, that laser will demand power and that chip will get hotter,’ explains Reed. It’s a good idea if you have highly dense interconnect to operate with silicon because it is temperature dependent – that means modulators or multiplexers tend to work by means of the refractive index and that will drift if the temperature changes.’ Microchips are already running hot and, if the temperature is hot enough, that can cause problems of drift. Cooling is energy intensive for the likes of data centres and one option Reed points to is letting the chips run very hot. ‘You could run everything hot and stabilise it at a high temperature, which can be cheaper and easier than cooling.’
A firm with silicon photonic products in the market already is Californian company Kotura, where Arlon Martin is vice president of marketing. He says: ‘Our strategy is to build in a CMOS wafer in a large foundry devices that are integrated that perform all the optical functions for 100 gigabit transmission and to do that on a chip that is no bigger than 4 to 5mm on its side.’
Martin points out that, as well as the manufacturing challenge, Kotura has to deliver something that has low power consumption and is cost-effective for cloud computing providers. ‘This growth in cloud computing and data centres, it is really that networking fabric that connects the hundreds and thousands of servers together that we use when we log into Facebook or LinkedIn,’ explains Martin.
For silicon photonics, the sales volumes for the data centre will be vast for low-cost 100Gb lines to the server. Existing data centres have a great many server blades with 1Gb ports, and these blades are in racks of 40 at a time. That means 40 electrical cables to the top of the rack and its switch and a need to integrate those 40 1Gb blades, but typically they will connect with a 10Gb port. The problem now is that the server blades have 10Gb ports so that means the switches at the top of the rack, and the clusters in the middle, all need 100Gb ports. These horrendous bottlenecks are what is driving the move to photonics. Another factor is the need to try to keep the switches and ports power consumption as low as possible and for a reasonable cost.
The next big growth market for Kotura is supercomputing. Martin says: ‘For us, that is the second phase and basically there is not much difference between data centre architectures and high-performance computers, except for the fact that they squish everything together.’ A key difference is where the optical communications are. They are moved from the front panel of the server, where the transceivers typically are, to the rack at the mid-board. This is where the interconnect is, right in the middle of the machine next to the CPU or its hub chips. ‘Really we see the embedded or board-mounted modules as sort of the next generation and those market numbers are very large.’ As well as widespread demand due to the many servers the internet consists of, the transmissions rates are expected to rise to 100Gb and then to 400Gb and eventually terabits. Martin sees terabit transmission rates in the 2017 timeframe.
He says: ‘We have demonstration chips, receiver chips, that support 40 channel Wavelength Division Multiplexing at 40Gb each so 40 times 40 is 1.6 terabits for a single chip, and it’s about 1cm on the side – so pretty small, but the electronics are not quite ready for that. The optical stuff will get there a little ahead of the electronics, but we’re going to start seeing architectures like that within four or five years.’
For now, the market for Kotura is data centres. In data centres the data transfer distances can be up to 2km or as little as two metres, from one end of a rack to another. But Martin explains, that, typically ‘500m is a good distance for these things. These are not connected to server blades at the top of the rack, just a couple of meters between them, but once you start connecting these things you need a budget that will support 2km from one to another.’
Luxtera is interested in the shorter interconnect application for silicon photonics. ‘Today most copper interconnect in networking gear and computer gear is running at 10Gbps, so copper is able to cope with that. Today copper can do a reasonable job,’ says Chris Bergey, Luxtera’s vice president for marketing. ‘Optical is only used for connects that are longer in length, but as you see interconnects go to 25Gbps with 100Gbps Ethernet; even at 40Gbps Ethernet. People want 20 to 30Gbps to reduce the pin count on connects. You can do copper, but you start to spend quite a bit on material with that, and you’re burning a lot of power.’
Founded out of the California Institute of Technology, Luxtera has sold more than 500,000 silicon photonic devices since 2008 including high-performance computing devices for interconnects.
‘And now we’re doing four by 14Gb and that’s the history. There is a lot of interest in silicon photonics,’ adds Bergey.
Other applications Luxtera wants to exploit are optical modules and the company has partnered with ST Microelectronics. ‘They are licensing our technology. That is one European area we’ve been engaged with,’ says Bergey.
Light is on the march in computing, and photonics can only play a larger role in future as computer design removes the material bottlenecks one by one – starting with copper.