With packaging costs dominating the cost of optical transceivers, will silicon photonics ever provide the cheap, fast optical links that data centre managers crave? Rebecca Pool finds out
At a time when internet giants are demanding faster and cheaper optical networking products for their data centres, silicon photonics has yet to deliver. At this year’s Optical Fiber Conference (OFC) in Los Angeles, Facebook network architect, Yuval Bachar, again called for optical links priced at $1 per gigabit running over singlemode fibre. Yet commercial development of silicon photonics products isn’t progressing fast enough to keep up with the industry’s desires.
‘Silicon photonics simply hasn’t delivered on the data centre promise,’ said Chris Cole, director of transceiver engineering at Finisar. ‘Many interesting things are being done with silicon photonics for long-haul communications but for the data centre, we have a completely different set of issues.’
From modulators to ring resonators, the building blocks of silicon photonics integrated circuits (ICs) are falling into place. Optical vendors such as Luxtera and Mellanox are already selling parallel singlemode transceivers based on silicon photonics technology to data centre operators. However, devices that can transmit multiple wavelengths over singlemode fibre are still on the ‘to-do’ list.
‘Where silicon photonics has failed to deliver is WDM, which is really what people want,’ asserted Cole. ‘Parallel singlemode fibre is interesting and some of the high-end data centres like it as they have very structured cabling and can control their cabling costs. But in general, people don’t want to see this... if you want to go at a higher speed, WDM is required.’
Data centre operators are interested in WDM over singlemode fibre because it promises to be cheaper in the long run. WDM transceivers send and receive multiple wavelengths over one fibre pair, whereas parallel implementations would need four fibre pairs to get to 100G. This makes the overall cost of fibre roughly four times more expensive, and as data centres grow in size, these costs become prohibitive.
Facebook’s Bachar says that future is approaching fast. The social media company is seeing at least a doubling of its data centre connectivity requirements every 8 to 10 months, he says. ‘For us it makes sense to do the transition to 100G in a singlemode pluggable module. If we can do this in a singlemode duplex module, then future [speed] transitions are going to be much easier,’ he observed.
Promising results are coming out of research laboratories. For example, earlier this year, Belgium-based research organisation, imec, alongside Tyndall National Institute, Ireland, and the Belgium-based Universities of Leuven and Ghent, demonstrated a 4x20Gbps WDM hybrid CMOS silicon photonics transceiver. While this is a clear signal that cheaper, high-density, singlemode optical fibre links are on the way, the product is still under development with no planned date for commercial release. In terms of quickly and cheaply assembling and packaging a finished product, the industry still waits.
Thinking inside the box
For singlemode data centre applications, designing WDM transceivers in silicon photonics brings a host of new challenges. Photonics packaging must provide optical input and output coupling as well as electronic connections and thermal management. For the photonic chip manufacturer, each area has its challenges, but it is the optical coupling of the chip to a singlemode fibre that is proving particularly troublesome.
For example, coupling a singlemode fibre input or output to a photonic chip requires the optical glass fibre to be very accurately aligned to the silicon waveguide coupler on the photonic chip. And herein lies the problem.
Mainstream microelectronic chip packaging relies on automated high-throughput assembly tools that have a placement accuracy of ±10µm. However, singlemode optics demand 1–2µm alignment tolerances. This mismatch means photonic device manufacturers cannot use existing high-throughput production line tools. Instead, they have to adapt assembly processes and manually find the best optical coupling position for the fibre and chip, which takes time and doesn’t come cheap.
Indeed, as Tymon Barwicz, lead researcher of silicon nanophotonic packaging at IBM, highlights, final package costs can be an order of magnitude more expensive than the actual chip itself. ‘Standard alignment tools fail so manufacturers have to resort to very expensive active alignment methods, which is where the problems start,’ he said.
‘In the 1990s, the focus was on very expensive telecoms devices, and manufacturers didn’t really care about the expense,’ he commented. ‘But right now we are stuck with this legacy and are trying to bring the prices much lower so packaging doesn’t completely dominate the price of the device.’
With this in mind, industry players such as Luxtera, Mellanox, Intel and IBM, as well as research organisations including CEA-LETI, France, and the Tyndall Institute, Ireland, have been busy developing fibre-coupling techniques to connect photonic chips to singlemode fibre.
Two main strategies exist: grating couplers that use diffraction to direct light upwards into a fibre above the chip surface, or edge coupling, using an inverted taper to couple light between a horizontal fibre and the edge of the circuit.
Arguably, grating couplers have seen the most success to date. Commercial offerings are available from Luxtera that implement parallel singlemode fibre coupling via this method, but it has limitations when it comes to WDM, as Professor Peter O’Brien, who leads photonics packaging research at the Tyndall Institute, explains.
‘Grating couplers are an attractive solution to the fibre-coupling challenge but they can have a relatively low coupling efficiency,’ he said. ‘The grating structure also gives this coupler a pronounced spectral dependency so its coupling efficiency is strictly wavelength dependent. So if you want to multiplex with a WDM fibre, the grating coupler must be designed for a certain bandwidth… the grating coupler may not be the best solution.’
Clearly the edge coupler is an alternative, and promises more efficient coupling and a broader spectral range, but fabricating structures that can match the relatively large optical mode of the fibre to the smaller mode of the silicon waveguide on the photonic IC is hardly trivial. And, as myriad organisations pursue different edge-coupler designs, the costs associated with such tight coupling tolerances are never far from researchers’ minds.
Indeed, be it a vertical grating coupler, or horizontal edge-coupling, today’s packaging costs are still prohibitively high as manual assembly is often required. ‘A big cost here is around the time taken to align the components,’ said O’Brien. ‘In these “active” alignment processes, gross alignment takes place but then you still do a final step of fine active alignment. Ideally you would want a fully automated or passive alignment system, to reduce the time taken by engineers and technicians, and this is what we are now starting to investigate.’
To this end, O’Brien and colleagues at the Tyndall Institute are developing passive processes for coupling photonics ICs to singlemode fibre via what they call an ‘optical interposer’ – a piece of silicon sandwiched between the circuit board and the photonic ICs, which contains optical waveguides to guide light to new locations with a more suitable geometry.
Tyndall has also recently introduced a new planar packaging process that uses a standard optical fibre, with an angle-polished end facet to reflect light from the fibre core down to a grating coupler on the silicon photonic IC (and vice versa for output couplers). The polished fibre is placed in a V-groove submount, with an automated vision system used for the final stages of alignment.
For both grating-coupler and edge coupling, these processes should reduce or even eliminate the need for precision manual alignment of the fibre and photonic circuit, a critical step in cutting packaging costs and moving towards high-volume assembly processes.
The Tyndall Institute works with Finetech, a Germany-based supplier of semiconductor packaging equipment, but O’Brien believes the industry in general needs to work more closely with equipment providers to design tools for high-volume photonic IC packaging processes.
‘Packaging people should be working more closely with the equipment suppliers, developing roadmaps so they can look at future requirements from a packaging perspective,’ he said. ‘That’s what we’re starting to do... and this is what happens in the semiconductor industry.’
The Tyndall Institute is hardly alone in its endeavours to develop cheaper, passive alignment processes for photonics IC packaging; other key players include IBM, CEA-LETI and more. As a result standardisation has become a pressing issue.
‘A lot of companies are working on solutions; IBM and Luxtera, for example, are solving their own customised approaches,’ said O’Brien. ‘However, there are currently no general standards; no standard packages, no standard fibre interfaces, and this needs to happen a lot more.’
But, as the researcher points out, change is afoot. The European Photonics Industry Consortium (EPIC) has started looking at standards for packaging while the Tyndall Institute has also established a set of design rules for silicon photonic packaging processes, with input coming from industry partners.
‘For example, there’s a standard pitch for fibres and electrical connections but a lot of people that design chips don’t realise this,’ he said. ‘Standardisation of packaging design is a real issue but we’re working to address this.’
As researchers pursue passive designs, Barwicz and colleagues at IBM have also been developing fibre-coupling interfaces that allow connections to photonic chips to be assembled using standard high-volume, low-cost microelectronic packaging equipment. One promising approach uses a ‘mechanically compliant polymer interface’, essentially a polymer ribbon with integrated optical waveguides.
A standard singlemode fibre can be butt-coupled – a technique widely-used in fibre connectors – to the mode-matched polymer waveguide. This polymer waveguide is carefully designed to transfer the light from the fibres to adiabatic waveguides inside the polymer ribbon which are then laterally coupled to waveguides on the photonic die.
And, crucially, IBM also lithographically defines self-alignment structures in the coupling region of the polymer ribbon and silicon. Here, alignment ridges are defined on the polymer, which mate with grooves on the silicon substrate. When the ribbon and substrate are pressed together, the placement improves from a typical ±10µm accuracy before compression, to 1 to 2µm accuracy in the finished component.
‘All of our photonics chip packaging relies on self-alignment, which we believe offers the cheapest and highest throughput method to do this,’ said Barwicz. ‘In practice, placement can be very crude, but because of these nested structures, [the components] re-align by themselves. In this way we are bringing the cost efficiency of microelectronics packaging into silicon photonics packaging.’
But, as Barwicz admits, his company’s photonic packaging technology is still under development. ‘We are in the last development stages and have a lot of interested parties, but products are not yet on the market,’ he said. ‘This could be one, two years, or beyond, depending on the qualification level required by the given application. [Products] will be tested really, really hard and this takes time and volume.’
IBM recently demonstrated a reference design for a fully integrated photonics chip carrying four 25Gb/s optical channels. According to the company, within a full transceiver design, these four channels could be wavelength multiplexed on-chip to provide 100Gb/s bandwidth over duplex singlemode fibre.
The company says the manufacturing process for the resulting 100G transceiver is now fully described by a process design kit: a collection of foundry-specific data and script files used with electronic design automation tools to create a design. Yet packaged products could still be several years away, IBM admits.
So, will this technology – and indeed the up and coming developments from the likes of the Tyndall Institute, CEA-LETI and more – provide data centre managers with the cost-effective optical interconnects they crave? Only time will tell, but Chris Cole from Finisar believes that the industry is, at last, heading in the right direction.
‘People haven’t figured out the right way to do this so that’s why we have a wealth of approaches,’ he said. ‘But the good news is the industry is working on the right problems. Generally when you identify the problems, eventually those problems get solved. It’s not always in ways that we anticipate because that’s the nature of innovation. But I think we will now see real progress and we will see the products.’
Rebecca Pool is a freelance science and technology journalist based in Boston, UK