Using light to move large quantities of data looks promising, but gaps remain and the adoption timeline will vary by application.
Silicon photonics is attracting growing attention and investment as a companion technology to copper wiring inside of data centers, raising new questions about what comes next and when.
Light has always been the ultimate standard for speed. It requires less energy to move large quantities of data, generates less heat than electricity, and it can work equally well over long or short distances. Moreover, many experts contend it will be harder to hack data delivered with photons than over a copper wire.
For all of these reasons and others, investments are increasing in silicon photonics. A study by UC Santa Barbara showed that between 2009 and 2015, the number of papers more than doubled to roughly 14,000. In 2005, there were about 500 papers on the subject. Unlike in the past, when most research was funded academic or government sources, today’s funding is largely from commercial sources.
There are signs of tangible progress on other fronts, as well. Mentor Graphics and Cadence have tool flows for silicon photonics design. Synopsys, meanwhile, has been expanding its photonics and simulation software line that it acquired when it bought RSoft Design Group in 2012. And GlobalFoundries has ramped up its investment in the process for commercializing it. Now the question is how quickly the infrastructure can be developed, vetted, and ramped to utilize silicon photonics for new applications and in new markets.
The long-term goal is to use photons to carry data across a chip, eventually replacing SerDes, traditional interconnects and maybe even conventional transistors. But that could take a decade or more. The short- and mid-term goals are more focused on the data center and within the network, where data is ballooning from video, various types of imaging (including embedded vision and virtual/augmented reality), and a proliferation of sensors with the Internet of Things. The first mass deployment of light-based communication began in the 1990s, using fiber optics to lay the backbone for the Internet. It has since expanded into data centers, where silicon photonics is used to communicate between racks of servers, and between those servers and storage. The next phase is expected to involve communications between chips within a package.
But there are a number of technical and business-related challenges that must be resolved to really propel this technology forward. Light sources based on materials such as gallium arsenide, indium arsenide (InAs) or indium gallium arsenide (InGaAs) need to be more tightly coupled into the manufacturing process to achieve economies of scale, which has been problematic in the past because these III-V materials are difficult to work with using conventional silicon processes.
In addition, tools are required to effectively deal with waveguide sidewall roughness, spatial separation, and die-level variation in waveguides and devices, according to John Bower, a UC Santa Barbara professor and deputy CEO of the American Institute for Manufacturing of Photonics (AIM), a joint effort between the U.S. government and universities in New York, Massachusetts and California.
“What’s needed now is to increase the lifetime of these lasers,” said Bower. “The goal is 4,000 hours. We’ve seen 2,100 hours, which is not sufficient for anyone’s laser.”
How rapidly those changes will be implemented isn’t clear. Nevertheless, estimates of the size of this market are enough to warrant at least some attention. GlobalFoundries expects this will become a $3 billion market within three years, facilitated by continued improvements in process technology to more easily incorporate light sources into silicon, which has been the chief problem identified by proponents of the technology. And given the data growth rates, silicon photonics is on a very short list of possible solutions.
“IP (Internet protocol) traffic is overwhelming bandwidth,” said Ted Letavic, senior fellow at GlobalFoundries. “Mobile data use will be 30.6 exabytes per month by 2020.”
Letavic believes the first major impact of silicon photonics will be the re-architecting of the data center. Rather than one or two huge data centers, he said the current thinking is to set up many smaller data centers and connect them with silicon photonics. “Optical interconnects will replace copper and be used to augment microwave and millimeter wave. And with 5G, you will need a high-speed interconnect from the data center to small cells. You will need at least 1 gigabit per second to the edge node, and 16 to 25 gigabits per second between the small cell and the base station.”
That scale-out will involve heterogeneous networks, or HetNets, and these will benefit from photonics integration, he said. Then, over time, this technology will migrate into 2.5D, 3D and monolithic packages. “Thermal-optical effects still have to be very accurately models, and substantial improvement is needed on waveguide design. We also need more structured toolsets for reliability.”
GlobalFoundries isn’t alone in seeing silicon photonics as a big opportunity. Hewlett Packard Enterprise, Intel and Juniper Networks, all are aggressively pursuing heterogeneous integration of photonics, said UCSB’s Bower. Cisco has a big investment in this market, as well.
Silicon photonics, in a nutshell, relies on optical waveguides rather than copper to route the light. The real challenge is incorporating the light source into silicon as part of the manufacturing process. The current approach uses quantum wells—basically creating a sandwich of one III-V material, such as gallium arsenide, between layers of a another III-V material, such as indium gallium arsenide. By doing that the electrons are trapped perpendicular to the layer surface, to create a laser light source.
Research is underway to replace quantum wells with quantum dots, which can be finely tuned by changing the size or shape of the dot. As current is applied, the dots emit light at very specific frequencies. Researchers say this will reduce the cost for integrated light sources, lower the power threshold, and increase reliability over time.
Hybrid silicon III-V optical amplifier. Quantum wells are in red. Source: SPIE.
Adaptation vs. reinvention
Despite some fundamental differences in the technology between photonics and silicon-based semiconductors, there also are some overlaps.
“Thirty mask layers are common, and there are more than 400 elements per chip,” said Bower. “What we need is a photonics version of Moore’s Law.”
There are other similarities, as well. “The most sensitive parameters like line width and edge roughness are already being dealt with in silicon,” said Duane Boning, professor of electrical and computer science at MIT. “Sensitivity lags IC applications, but with GlobalFoundries moving from 200mm to 300mm processes, there will be big improvements in photonics. You get finer process control. CMOS technology is still a couple generations ahead, but what we’re seeing in photonics is that it’s gathering enough replication to understand more subtle variations. So we understand thickness of wafers, for example. And we’ve had to deal with so much variation around thermal that we have had to adapt and tune structures.”
He said that some of these approaches may be hard to translate to the silicon photonics world, but they don’t have to be reinvented.
Techniques on the design side are applicable, too. “There is a lot of stuff we learned at advanced nodes, like self-heating, where the solution is very applicable to photonics because it provides a thermal map,” said Gilles Lamant, distinguished engineer at Cadence. “We also have dealt with line-edge roughness on fins, which is very applicable to waveguides. There is a potential to provide other things to photonics, as well, such as design for manufacturing.”
Added Lamant: “The key is making sure that design intent is correctly implemented.”
Process design kits are in the works for silicon photonics, as well. Chris Cone, product marketing manager for Pyxis IC Station at Mentor Graphics, said that to create a design with photonics you take a PDK and build it around two or three specialized devices. “Photonics is a new market, so basically you’re designing around a novel device. But you’re still doing verification for DRC (design rule checking) and LVS (layout versus schematic), lithography simulation, and you still need a full flow as you do in CMOS.”
PhoeniX Software, based in the Netherlands began developing PDKs in 2008 with Imec for silicon photonics, according to Twan Korthorst, PhoeniX’s CEO.
Korthorst noted that other areas are being explored, as well, including biophotonics, where sensing is done with light rather than electronics. “Some light is not confined to the waveguide. It’s influenced by a surrounding ring, which you coat with a chemical. So that chemical may be sensitive to certain bacteria. We’re seeing commercial companies working with this in the biological domain.”
Mentor’s Cone pointed to another biomedical application involving sensors created from an array of micro-capillaries. “You’re able to sort out molecules based upon the molecule source. Then you use optics to interact with them and find potential pathogens in the blood.”
Progress on other fronts, too
While silicon photonics may provide a huge reduction in latency, it’s not the only piece of the puzzle. For data to move at the speed of light, everything else has to be speeded up, as well. That means signals need to be processed and routed much faster than today. At the chip level, this ultimately could become a requirement because of the RC delay in increasingly skinny wires, thinner contacts, and a predicted breakdown in dielectrics at the most advanced nodes.
“There is always a need for higher and higher communication speeds,” said Mike Gianfagna, vice president of marketing at eSilicon. “We’ve seen that with SerDes in the high-end network and compute space. We’ve also seen that in the interface between high-bandwidth memory and the mission-critical stack. Demand for higher-bandwidth is holding steady, which right now makes SerDes the most critical element. It’s like the price of admission these days. It’s been shown to have reliable deployment.”
How long that will be good enough, and when SerDes will run out of steam and provide an opportunity for photonics, is unknown.
“In the networking space there has been talk about high-bandwidth, low-latency technology that would replace SerDes,” said Kurt Shuler, vice president of marketing at Arteris. “For some chips in some niches, photonics may work, and maybe in the future we’ll see it chip-to-chip. We’ve been working with the CCIX (Cache Coherent Interconnect for Accelerators) Consortium, where everything will run on a PCIe PHY using a CCIX controller. In the future, that won’t be on a board. But you will still need physical connectivity, and photons produce less heat.”
Shuler said that at the chip level, the big bottleneck is communication between the SoC and DRAM. “If you look at high-bandwidth memory (HBM) and the Hybrid Memory Cube, everyone is trying to work around that. But DRAM is still slower and cheaper, so given that constraint, how do you make everything run faster? One solution is to add more SRAM. You can build in cache coherency and proxy caches, and it helps if any element can use it rather than just one processor. But with DRAM, it’s not just about latency and bandwidth. It’s also about power.”
And this is where photonics really excels. “If you take a look at the trends on the road map, there is a move to get away from silicon,” said Greg Yeric, an ARM fellow. “That includes materials such as germanium and indium-gallium-arsenide. It’s inconceivable that photonics won’t be part of that, and someday you’re going to see a dedicated photonics chip in a package for the people who really care about performance. We also may see a move toward plasmonics. With nanometer manufacturing capability, you can make the right thickness and pattern of film and convert from a photon to a plasmon.”
But how quickly all of this technology gets rolled up into a vibrant market opportunity may depend on how well other technologies fare. A classic comparison is in the area of Ethernet, which everyone assumed would die off when wireless became pervasive. Quite the opposite occurred.
“Industrial is all moving to Ethernet because of economies of scale,” said Lixin Zhou, senior director of the switching product line at Marvell. “If you look at high-performance computing, the majority of those are based on Ethernet now. It used to be InfiniBand and Fibre Channel. Now it’s Fibre Channel over Ethernet.”
When Ethernet was introduced back in 1980, it had a transfer rate of 2.94 Mbits/second. The latest Ethernet speeds run as high as 100 Gbits/second, and Ethernet speeds inside of existing transceivers are increasing from 25 Gbps/line to 10 Gbps/cable for a four-cable connection, which also provides more granularity in controlling traffic.
Whether that will be enough to offset adoption of faster technologies such as photonics may depend on the application, but the point is that nothing is standing still. So while photonics is definitely on the horizon, the timing needs to be considered in the context of other technologies, the challenges of sticking with Moore’s Law, and a variety of other issues that are only tangentially connected to photonics.
The promise of photonics is enormous. It’s not clear whether speed will be the main driving force, or whether power and thermal effects will propel it forward. It may be one or all of them. But it’s not clear when this will happen or for which markets it will happen first.
This is a technology that is still in its infancy, and change is swirling around almost every aspect of the electronics design through manufacturing flow. Quantum computing, neural networks, artificial intelligence and machine learning all will have an impact on the volume and required speed of moving data, and photonics will be a piece of the puzzle. How big of a market it garners, where that happens—whether it remains as a communications channel between racks of servers and storage or whether it moves further to the chip—and when are all big question marks.
Nevertheless, nothing moves faster than the speed of light, and with the amount of money pouring into research, this market appears to be moving rather quickly.
Photonics Moves Closer To Chip
Government, private funding ramps up as semiconductor industry looks for faster low-power solutions.
Focus Shifting To Photonics
Using light to move data will save power and improve performance; laser built into process technology overcomes huge hurdle.