Disruptive technology or just a silicon blip?
Silicon photonics-based products are coming to market after more than a decade of development work. But the optical component industry continues to debate the significance of the technology.
Optical component makers that have long used established materials such as planar waveguides, indium phosphide and gallium arsenide acknowledge silicon photonics’ potential but say they have yet to see telling design advantages using the technology. In contrast, adherents believe silicon photonics not only delivers more compact, lower-cost products but changes the way systems can be architected. Moreover, the importance of the technology will only grow as signal speeds increase and optics moves closer to the chips.
‘It is a very exciting time,’ says Mario Paniccia, Intel fellow and general manager of the company’s silicon photonics operation. ‘I think we are at a cusp of a new technology wave.’
But Dale Murray, principal analyst at market research firm, LightCounting, asks: ‘Is this technology intended to do things that we can’t otherwise do? Or is it going to take what we can do today and do it so much more economically?’ He tends to the former: silicon photonics will come into its own, but for applications that are still to emerge. Certainly, interest in the technology is growing – with equipment makers acquiring silicon photonic start-ups, optical transceiver designs based on silicon photonics coming to market, and first system architectures using the technology being unveiled.
Luxtera launched its first 40-gigabit active optical cable product in 2007. The silicon photonics company subsequently sold its active optical cable business to Molex to concentrate on optical transceiver and optical engine designs. Luxtera still provides optics for active optical cables and for 40-gigabit and 56-gigabit QSFP optical modules. Luxtera and Kotura (acquired in 2013 for $82 million by switch vendor Mellanox Technologies) are both developing 100-gigabit (4x25 Gig) QSFPs, likely to appear from 2014.
Cisco Systems has launched a custom 100-gigabit pluggable module, the CPAK, based on silicon photonic technology gained with the $272 million acquisition of Lightwire in 2012. More recently, Huawei acquired Caliopa, a four-year-old silicon photonics Belgium-based start-up, for an undisclosed fee. The Chinese system vendor said it would bolster its investment in European R&D, and highlighted silicon photonics.
Interest in silicon photonics is not confined to datacom and telecom vendors. European chip company STMicroelectronics licensed Luxtera’s silicon photonics technology in 2012 and is preparing a 300mm (12-inch), 65nm process wafer fabrication line to deliver its first products in 2014. ‘We think we are the only ones doing the processing in a 12-inch line,’ says Flavio Benetti, general manager of mixed processes division at STMicroelectronics.
One advantage of silicon photonics is its use of fabrication techniques developed for the chip industry. ‘The maturity of the technology and the precision and accuracy by which things are done in a CMOS fab is beyond any other technology in this world,’ says Roel Baets, a professor at the Photonics Research Group at Ghent University. ‘It has become amazingly mature, and silicon photonics is exploiting that maturity.’
Silicon has a high refractive index compared to silica or air, says Baets: ‘That high index contrast buys you the fact that you can miniaturise many functions so that the chips take little space.’
Another indication of the potential of silicon photonics is the broad range of applications it is being applied to; from short-reach interconnect to long-distance optical transmission. ‘Silicon photonics is going to be ubiquitous across the network: in the cloud, the data centre, the wide area network, the IP network, all the way to the radio access network environment,’ says Gunter Reiss, vice president, business development and strategy, product area IP and broadband, business unit networks at Ericsson.
Ericsson is one of several companies that invested in silicon photonics start-up Skorpios Technologies in 2011. The company is also conducting its own silicon photonics R&D at sites in Italy and Sweden.
‘The lots of applications lead to lots of approaches, and underpinning that are multiple technologies, so that the base of silicon photonics is diverse in itself,’ says Eric Hall, vice president of business development at Aurrion. ‘Nothing is common,’ agrees Murray. ‘It is still a nascent and fragmented ecosystem.’
The original vision for silicon photonics was to implement photonics and the driver electronics on the one chip. IBM’s high-density silicon photonics optical engine that operates at 25 gigabit-per-second (Gbps) per channel, made using a 90nm CMOS process, is such a monolithic design.
But the general trend is to implement the functionality on separate chips due to the better economics. That is because photonics requires broad, 130μm lithography features whereas high speed drive electronics – 25Gbps and greater – typically use advanced 40nm and 28nm CMOS process nodes which are more expensive. Implementing photonics alongside the electronics using such an advanced process wastes expensive die area due to the relatively large features of the photonic circuits. The fundamental shortfall of silicon photonics is that silicon does not lase. As a result a laser source in the form of III-V material must be coupled for the photonic design.
One approach is to couple the external laser through fibre or by using flip-chip technology to attach a laser to the photonic die. This is the approached used by Cisco, IBM, Mellanox, Luxtera and STMicroelectronics.
Alternatively, III-V material is bonded to the silicon wafer and using photolithography and etching, both the laser and silicon photonic circuitry are constructed. This approach is known as heterogeneous integration (see box, below).
Volume manufacturing is an important driver for silicon photonic device makers. Chris Bergey, vice president of marketing at Luxtera, highlights the applications that can use silicon photonics but says a key consideration is choosing a sufficiently large market. ‘Is the opportunity big enough that you want to go invest in that area?’ says Bergey.
Luxtera is focussed on data centre interconnect, building on its active optical cable business for which the company estimates it has supplied 600,000 photonic chips.
Volumes are also an issue for STMicro, which needs to fill its 300mm wafer production line. ‘We need to go for product; we need to go for volumes. This is where we are addressing our attention,’ says Benetti. The chip maker has yet to detail its product plans but says it too is targeting short-reach interconnect.
Mellanox Technologies makes Ethernet and InfiniBand switches. The company says its highest speed switch runs at 56-gigabit and its next-generation platform will use 100-gigabit ports with both copper and fibre-based interconnect.
Copper cables and VCSEL-based optics will remain the dominant interconnect options but beyond 100-gigabit, Mellanox questions the viability of VCSELs. ‘We will need something that will deliver those kinds of speeds reliably, and this is where silicon photonics enters the picture,’ says Gilad Shainer, vice president of marketing.
‘Mellanox is a company that tries to develop and own all the technology we need; owning the technology gives us the ability to add value and depend on ourselves,’ says Shainer, explaining the company’s acquisition of Kotura.
Intel’s investment in silicon photonics over the last decade is to help spur sales of its microprocessors. The chip giant is a member of Facebook’s Open Compute Project, based on a disaggregated system design that separates storage, computing and networking. The disaggregation can be within a rack or across rows of equipment.
Intel’s Rack Scale Architecture (RSA) is a disaggregated design. One RSA implementation for Facebook uses three 100-gigabit silicon photonics modules per tray. Each 100-gigabit module comprises four transmit and four receive fibres, each at 25Gbps. Each tray uses a Corning-developed MXC connector which can accommodate up to 64 ClearCurve multi-mode fibres, for a total link bandwidth of 1.6 terabits (see box, below).
Mellanox’s SX1036 switch can accommodate 36, 56-gigabit QSFP ports across the front panel. ‘This is the densest switch on the market,’ says Arlon Martin, senior director of marketing at Mellanox. ‘We cannot fit more on the front panel.’
To expand switch capacity, Mellanox is both developing a 100 gigabit QSFP as well as embedded modules that will sit on the board, closer to the switch silicon. Mellanox and others are investigating whether a QSFP would support 50Gbps channels to deliver a 200-gigabit module. Mellanox is also exploring other pluggable form factors such as the recently announced 400-gigabit CDFP that promises to double the capacity on a switch’s front panel.
Embedded optics promises greater switch interface densities as well as lower power consumption. ‘If you want to do something other than a factor of two improvement, more like a factor of 10, you will need connectors on the front panel as opposed to pluggable transceivers,’ says Martin.
Luxtera points out that putting the optics on the board halves the power consumption by avoiding having to run 25Gbps electrical signals across the board to the front panel. ‘Beyond that, it is building ASICs where you come out of the ASIC optically,’ says Bergey. ‘That is what we are doing with the mid-board stuff: proving out the technology such that people can build those types of devices.’ Luxtera is already working on co-packaging optics with the ASIC. ‘We will have most of the pieces proven out by early 2015, with a 2016 production timeframe,’ says Bergey.
IBM has developed a 25Gbps-per-channel silicon photonics optical engine technology for its platforms, for chip-to-chip and backplanes, less for data centre interconnect. But the company is open to selling the engine to optical module players.
Datacom may have the lure of volumes but silicon photonics is also being applied to telecom. Both Ericsson and Teraxion are exploring silicon photonics for telecom as well as datacom.
Ericsson’s Reiss highlights such applications as ASIC-to-ASIC backplane connectivity for IP routers and next-generation cellular base stations, transceivers for data centre top-of-rack switches, modules on routers for IP-over-DWDM, and software-defined networking switches.
Using silicon photonics for 100-gigabit coherent transmission modules promises to reduce power consumption to a quarter, and cost to a fifth compared to existing designs, says Reiss. Skorpios has already detailed a tunable laser design for long-distance transmission earlier this year. ‘The tunable laser chip was one of the most interesting parts and why we decided to invest in Skorpios,’ says Reiss. ‘We saw the disruptiveness, and the material savings, size and power dissipation.’ Ericsson expects silicon photonic-based modules to appear on IP router line cards in the coming two to three years.
Teraxion’s initial silicon photonics focus has been 100-gigabit coherent receivers. The firm has been developing indium phosphide technology it acquired for transceivers in the data centres.
Implementing its coherent receiver design in silicon, the chip cost is cheaper than using indium phosphide. ‘We are impartial, we have both platforms now,’ says Carl Paquet, director, product line manager at Teraxion. But the major cost saving comes not from the chip but its packaging: Teraxion does not need a gold box as the silicon receiver need not be sealed hermetically. ‘Product cost comes mostly from the packaging,’ he says.
The Optical Internetworking Forum is developing a micro-integrated coherent receiver (ICR) specification that will fit within a CFP2 pluggable module. ‘Most likely our next product will meet this micro-ICR specification,’ says Paquet. Companies are exploring ways to improve the economics of using silicon photonics for the lower-volume telecom markets.
In a post-deadline paper given at OFC 2013, Aurrion detailed the making of various transmitters on a silicon wafer. These include tunable lasers for telecom that cover the C- and L-bands, and uncooled laser arrays for datacom. The ability to manufacture multiple designs on one wafer benefits the economics of silicon photonics for telecom by manufacturing them alongside higher-volume datacom sources, says Hall.
Meanwhile, the industry continues to grapple with the challenges associated with an emerging technology. One main challenge that many highlight is solving the light source problem.
‘Obviously everybody would hope that, one day, one can simply integrate light sources using wafer-scale manufacturing but that is one of the big challenges in the years to come,’ says Baets.
‘Coupling the light in the fibre attachment – these are processes that still have a high degree of improvement,’ says STMicro’s Benetti. ‘We are at a good stage of speed and precision in the placement of the fibres but there is still much to do.’
Single-mode versus multi-mode fibre
As the size of mega data centres continues to grow, so does the optical reach needed to connect switching equipment. VCSEL-based optical modules only have a reach of 100m over OM4 multi-mode fibre yet 500m and greater spans are emerging.
‘Data centre operators view fibre as part of the infrastructure and a long-term asset,’ says Luxtera’s Bergey. ‘They are putting in so much fibre and they really like the idea of single mode; it gives them an ability to be future-proofed.’
Single-mode transceivers using distributed feedback (DFB) lasers or silicon photonics can easily achieve 500m to 1000m.
‘Silicon photonics makes a great single-mode transceiver and you have a market now looking for bandwidth,’ says Bergey. ‘That is our focus.’
In contrast, Intel has developed a multi-mode-based silicon photonics approach which it argues is more economical. ‘When you look at connecting single-mode fibre to the rack, or the board, that connectivity at the connector level and at the module level is very expensive because it requires precise alignment,’ says Paniccia. ‘Silicon photonics is inherently single mode, and we have worked hard to make it “multi-modable” so that we can deliver a total solution: silicon photonics, packaging, assembly and cabling that is cost-effective.’
A reach of 300m is achieved using Corning’s ClearCurve multi-mode fibre, and Intel has demonstrated it working over 820m. ‘People do want single mode fibre and longer reaches but the average span in a data centre is 100m, maybe 75m,’ says Paniccia. ‘The long lengths that go up and down the data centre, 500m to 1,000m, are not the volume links.’
But analysts question how much traction a new fibre will receive. ‘When data centre people have a choice between OM4 multi-mode versus single mode and they ask themselves: “How do I future-proof my data centre?” and a new fibre comes along, that is not making their decisions easier,’ says LightCounting’s Murray.