Using light to move data will save power and improve performance; laser built into process technology overcomes huge hurdle.
Silicon photonics finally appears ready for prime time, after years of unfulfilled expectations and a vision that stretches back at least a couple decades.
The biggest challenge has been the ability to build a light source directly into the silicon process, rather than trying to add one onto a chip after manufacturing. Intel today said it has achieved that milestone, setting the stage for building economies of scale into the process. That may take several more years, but it nonetheless represents an important step for this technology.
“We have solved the problem of integrating the laser into the process,” said Alexis Bjorlin, general manager of the Connectivity Group at Intel. “We invested in a methodology to bond light-emitting III-V indium phosphide to silicon so that the lasers are defined in silicon. This is the Holy Grail of silicon photonics.”
The first implementations of this technology will be between systems within a data center, where silicon photonics already is in widespread use. This is a relatively price-insensitive but fast-growing market, supported by improvements in performance and energy that are amortized across thousands of servers. But with a laser built into the semiconductor manufacturing process, the economics of photonics will change significantly. Bjorlin said the goal is to move photonics inside of multi-chip packages sometime in the next four years, with on-chip photonics following after that.
“The play for photonics is immense bandwidth with the scale of silicon and silicon manufacturability,” she said. “Right now we can drive a 3X per bit power reduction. So you have higher-rate switches, and you get an improvement in power consumption. The core differentiator there is the laser integrated on silicon.”
This also expands Intel’s opportunity well beyond just the chip to the interface, according to Jim McGregor, principal analyst at Tirias Research. But he cautioned that nothing will happen overnight, because it takes time for new process to roll through the cost curve that has defined the semiconductor industry.
“For about the next five years, you’ll see it confined to high-performance systems,” McGregor said. “After that, it will be used in more places. As we start seeing massive servers and enterprise neural networks, with machine learning and artificial intelligence, that will require next-generation processors. We’re already starting to see some of those next-generation processors being developed. Silicon photonics is likely to be one of the enablers.”
Seeing the light
There are a number of advantages to using silicon photonics. For one thing, light pulsing through a waveguide on a chip generates much less heat than electrons moving through copper wire—particularly a very skinny copper wire where resistance, as well as the power necessary to drive those electrons, generate heat. Light is also faster that electrons, less prone to cybersecurity issues, and there are fewer physical effects. Photons do not interact the way electrons do, which allows them to be bunched together much more tightly and to cross paths without affecting signal integrity.
“This is a new way to multiplex signals,” said Chris Cone, product marketing manager at Mentor Graphics. “But it also will require new architectures. Topologically, you need to connect everything with wave guides. You etch, grow, and then anneal to smooth the sidewalls to allow light to bounce through them. One thing that helps, though, is that it’s possible to cross waveguides. It will not short out like electrical signals because photons don’t interact the way electrons do.”
This is just the beginning, too. Cone said there are a number of emerging markets where this kind of technology will be useful, such as in the biomedical field where light can be used to interact with potential pathogens in body fluids run through a chip. “And you don’t need to use the most advanced process to make this work. It could be done at 350nm rather than 14nm.”
In the short-term, the real target is cloud farms. “As we disaggregate processing from the memory, you need a high-bandwidth connection between the memory and where data is processed,” said Gilles Lamant, distinguished engineer at Cadence. “In the past, we saw this with network appliances connected to fiber. Now, it’s being used for cloud farm environments.”
Intel isn’t the only company working to replace the copper interconnects in data center switches with one based on photonics, but it is the first to publicly announce a commercial, silicon process-based solution for the light source. It is unlikely to be the only player in this market, however. The question now is how big a jump the company has, and how much market share it can corral before others deliver a solution. At the very least, though, Intel has established a market direction, as it has repeatedly done in the past.
“The play for optics is immense bandwidth with the scale of silicon and silicon manufacturing,” said Bjorlin. “The big hurdle was about coupling issues, and that is solved with it being integrated in a silicon wafer using a standard 300mm process with proven repeatability.”
She added that copper will still play a role, but the massive explosion in bandwidth in everything from data centers to autonomous vehicles and airplanes will require fiber. “There will be 50 billion connected devices, and they will all drive data into data centers.”
Photonics Moves Closer To Chip
Government, private funding ramps up as semiconductor industry looks for faster low-power solutions.
Advanced Packaging Is Real. Now What?
With all the industry leaders now on board, big changes are expected.