Silicon photonics is the technology that manufactures hundreds of components required for optical communications with CMOS processes and has been used for many years to produce coherent optical modules for metro and long-distance communications. The increasing demand for bandwidth brought about by artificial intelligence is now opening the door for silicon photonics to make its way into data centers to improve its economics and functionality.
What's inside an optical module?
key semiconductors such as digital signal processors (DSPs), transimpedance amplifiers (TIAs), and drivers used to produce optical modules have steadily improved in terms of performance and efficiency due to Moore's Law and other factors.
This is not the case with optics. Modulators, multiplexers, lenses, waveguides, and other devices used to manage optical pulses have historically been provided as discrete components.
"Optics use almost parts," said Loi Nguyen, executive vice president and general manager of Cloud Optics at Marvell. "It's hard to scale."
Lasers are particularly challenging, and module developers are forced to choose between a variety of technologies. Electro-absorption modulation (EML) lasers are currently the only commercially viable option that can meet the 200G/s speeds required to support AI models. EML is typically used for longer links and is the laser of choice for 1.6T transceivers. Fab capacity for EML lasers is not only limited, but also very expensive. Together, these factors make it difficult to scale at the speed required by AI.
Enter the field of silicon photonics
Unlike traditional optical modules, silicon photonics uses common wavelength (CW) lasers, which are cheaper and easier to manufacture.
"The CW is like a light bulb...... It just glows constantly. It's easier to make, it's available from a variety of sources, and it's inexpensive," Nguyen said. "All the high-speed magic used to modulate data happens in silicon photonics.
Silicon photonic devices can also be produced in 200 or 300 mm fabs. For a decade, Marvell has demonstrated that silicon photonics can be mass-produced for use in coherent modules, something few companies have been able to do.
Why silicon photonics now?
For example, if a discrete module has eight 200G channels in one chip, it needs four EML lasers to operate at 1.6T. With silicon photonics, everything is integrated, and four channels can share a single laser, which means that the module only needs two cheaper CW lasers to operate. Integrated silicon photonic modules are also more reliable and scalable.
Silicon photonic modules require fewer and cheaper lasers to provide the same bandwidth as discrete photonic modules
"Lower costs, fewer lasers, and higher levels of integration mean higher reliability and better scalability," he adds. "When there's a huge pull in the market, consolidation always wins."
The future of silicon photonics is here
Marvell today announced a live demonstration of a 6.4T 3D silicon photonic engine with 32 channels, each running at 200G of electro-optical power. The engine integrates hundreds of components into the chip, including putting the TIA and the driver on the same device. This first-of-its-kind engine features a modular design that can be scaled from 1.6T to 6.4T and beyond. Given the integration, the engine lowers the cost per bit and creates more scalable options to meet the growing bandwidth demands compared to discrete solutions.
The 3D silicon photonics engine will serve as a building block for expanding optics and has multiple use cases in the field of optical interconnects. The first application will be a pluggable optical module, which will increase the number of channels that can fit into a module from 8 to 16, 32 or even 64. In the future, silicon photonics will enable co-packaging optical solutions. Eventually, photonics should find a home in chiplets.
Artificial intelligence will continue to drive the world, and connected technologies must scale to meet demand. By bringing silicon photonics to the data center, Marvell can continue to deliver the bandwidth it needs while improving scalability and reducing costs.
