Will silicon photonics replace copper cabling in mainstream datacentres?

Silicon photonics can increase bandwidth in servers, improve data transfer speeds and reduce complexity. But are datacentres ready for it?

Traditional copper cabling is stifling datacentre evolution and high-performance computing (HPC) because of its slow data transfer capacity. Silicon photonics – which uses optical fibre for data transfers – increases bandwidth in servers and racks, improves data transfer speeds and reduces datacentre complexity. But are datacentres ready for it yet?

Silicon photonics is an evolving technology in which data is transferred among computer chips by optical rays (laser light), which can carry far more data in less time than electrical conductors can. The optical fibre is directly built into semiconductor chips to give IT “computing at the speed of light”.

Standard copper-based Ethernet networking is inadequate for HPC applications, datacentres or for managing growing data volumes efficiently, and IT is struggling to provide faster systems with more effective bandwidth and, hopefully, at lower cost to end users.

So why is silicon photonics (SiPh) important? In a TEDTalk documentary on YouTube, Intel fellow Mario Paniccia (pictured) says: “If you had the capability of transferring data at 50Gps between two devices, you are talking about transferring an HD movie in less than a second.”

As SiPh develops further, it can realise data transfer at 1Tbps in a cost-effective way. “That means you can transfer or download a whole season of HD TV from one device to another in less than a second,” says Paniccia. “It will allow us to keep up with Moore’s Law and we will not be limited by internal network speeds.”

Moore’s Law has governed the computer industry for almost half a century. “The problem with Moore’s Law is that it is reaching a limit,” IBM fellow Bernie Meyerson tells Computer Weekly in a video interview. “You do realise that at some point you could have transistors that are one atom thick. When you cut them in half, that’s nuclear fission.”

But how can silicon photonics be cost-effective? Jeff Demain, business development director for Intel's silicon photonics solutions group, says: “When you look at today’s fastest supercomputer, the Tianhe-2, it takes 17MW to power it, plus an additional 7MW for cooling. The amount of copper cabling contributes to that power because the large number of cables in the machine block cool air flow.”

SiPh progress by chipmakers and semiconductor firms

Chipmakers Intel, IBM, NEC and Cisco regard silicon photonics as a promising technology that can change how datacentre systems exchange data and make rack equipment leaner, and they are investing in the technology.

If you had the capability of transferring data at 50Gps between two devices, you are talking about transferring an HD movie in less than a second

Mario Paniccia, Intel

Intel is set to launch silicon photonics chips later this year as it moves the technology from research to the production phase. Demain says there is huge bandwidth growth across multiple markets, including datacentres, HPC, telcos and consumer electronics.

Such data explosion requires innovative technologies, he says. “Intel silicon photonics offers opportunity for cost-efficiency, increasing as volume grows,” says Demain.

In 2011, Intel co-funded a silicon photonics research and production centre at the University of Washington.

Demain says the technology will allow new capabilities and flexibility to address the growing needs of datacentres.  

Cisco is also betting on the technology. It acquired a SiPh startup, Lightwire, in 2012 to deliver high-speed networks with the “next generation of optical connectivity” that can enable datacentre users to meet the growing demands of video, data, voice, mobility and cloud services.

At the 2013 Open Server Summit, Cisco engineer Joel Goergen said IT systems can be changed using photonics to develop a new generation of datacentre architecture that will be more power-friendly and cooling-efficient.

IBM, too, is investing in the technology. Its “silicon nanophotonics” uses light instead of electrical signals to transfer data, allowing large volumes of data to be moved rapidly between computer chips in servers, large datacentres and supercomputers via pulses of light. 

In 2012, IBM announced it had constructed a 1Tbps parallel interconnect using external vertical-cavity surface-emitting lasers (VCSELs) and photo-detectors communicating through 48 holes drilled through a 90nm silicon complementary metal-oxide semiconductor (CMOS) chip.

Meanwhile, in Europe, STMicroelectronics signed an agreement with Luxtera to develop a dedicated silicon photonics process at its wafer laboratory in France.

The need for higher-speed interconnects within IT components and between them in racks and across racks is a pressing one

Clive Longbottom, Quocirca

Silicon photonics is also capable of feeding a number of parallel optical data streams into a single fibre, which will reduce datacentre cabling problems.

But although suppliers may be keen to commercialise silicon photonics, is the industry really ready for it?

Ovum analyst Daryl Inniss says IT companies have been investing in SiPh research for more than 10 years, with little revenue in return.

Datacentre experts, too, are cautious about the industry's readiness for SiPh. “We will have to see where this [silicon photonics] goes,” says Clive Longbottom, datacentre analyst at Quocirca. “Sure, copper is not the long-term future for getting data from one place to another, but this has been said before about ‘ordinary’ fibre, yet copper not only maintains its hold, but speeds have exceeded what anyone had expected copper could do.” 

Intel says its SiPh products are made at the wafer scale, not with tweezers, while traditional optics are tweezer-assembled from individual parts.

“The need for higher-speed interconnects within IT components and between them in racks and across racks is a pressing one,” says Longbottom. “Coming up with a common form of connection, such as silicon photonics, could be really useful in collapsing connection chaos.”

Copper cabling's limitations

Ovum's Inniss says copper cabling is limited partly by distance restrictions, and the traditional optical approaches currently available have size, cost and power issues. 

Silicon photonics devices far exceed the capabilities of copper cabling by offering data rates of about 100Gbps and so are are better suited for HPC applications and datacentres.

Excitement has been created around products and devices based on SiPh, partly as the telecom and data communication segments are moving to 100Gps interconnects, says Inniss. 

But even if Intel starts to ship such connections by the end of this year, experts feel the penetration into datacentres will be relatively slow.

One reason for this could be that the existing on-board connectivity in datacentres is not “retro-fittable” and a datacentre would have to rip off its existing architecture to accommodate silicon photonics.

“Only interconnects between systems will be capable of using plug-in cards,” says Longbottom. 

The key advantage of silicon photonics is that optical and electronic functions can be integrated and fabricated onto the same chip

Jim Tully, Gartner

A Gartner research paper describes silicon photonics as a transformational technology offering many advantages, including large data rates, longer-distance connections, low latency and volume economics.

“The key advantage of silicon photonics is that optical and electronic functions can be integrated and fabricated onto the same chip,” writes Gartner analyst Jim Tully in his research paper. This potentially leads to lower costs and physically smaller solutions.

“We sense an increase in research effort and funding over the past two or three years following a period of relative decline around the time of the economic downturn,” Tully adds. “We expect the technology will see increasing use in experimental supercomputing platforms over the next few years.”

IBM expects optical silicon will further its “exascale computing” efforts, says Tully.

In the video interview, IBM’s Meyerson says silicon chips will not suddenly be replaced by a new technology – because nothing else currently exists that can manufacture chips on an industrial scale .

So when will SiPh enter the mainstream datacentre? It will happen – but it will take a long time, says Longbottom.

On its maturity charts, Gartner scores silicon photonics as “embryonic” with a market penetration of less than 1% of the target audience – racks in datacentres and high-speed computers.

“The technology has been seen on the horizon for many years, yet visibility outside the industry is quite low,” says Tully. “We expect that, after a number of additional years of development, it will transition to productive use relatively quietly, rather than undergo a significant hype and then crash cycle.”

One issue that may limit SiPh’s entry into mainstream datacentres may be the lack of standardisation. For example, Intel is the only developer working at 1,310nm, while IBM is experimenting on 90nm.

Inniss adds: “Intel will have to develop steps to standardise this interface and/or support expanding the transmitter and receiver ecosystem to multiple suppliers.”

But there are other challenges, too. One of the major hurdles faced by SiPh developers is the fact that laser devices, which generate the IR beams that carry the data, are power-hungry. Another limitation is that silicon material has not proved effective for lasing because of thermal dissipation within the material.

Experts estimate that silicon photonics may play a bigger role in enterprises that have the luxury of completely replacing their datacentre infrastructure, or those that want to add an extra section or replace a large part of their existing datacentre facility.

“For those that have massively data-dependent systems and require the lowest possible data latencies (and the deepest pockets), we may see faster adoption here,” says Longbottom. 

“However, overall, it will start off as a technological answer looking for the right problem.”

Read more on Clustering for high availability and HPC