A new kind of optical interconnect for data centers
In data centers today, there is a constant need for speed. And it's not only inside computer data servers, where one might think it would be. Computing power, capacity, and speed has been increasing at a fast clip for years now in response to market demand, and many artificial intelligence and machine learning applications, such as ChatGPT, Google Search, and Alexa can attest to this fact. However, data center servers, also known as "the cloud," are only as fast and as powerful as the infrastructure that supports them. And the next generation of AI, machine learning, and other computationally intensive applications will demand data transfer rates that are expected to increase exponentially in the coming years.
A key part of data center infrastructure are optical interconnects, which function as a bridge between electrical and photonic signals and connect data server clusters to fiber optic cables for data transmission at the speed of light. As the central processing units, or CPUs, within these data servers become more powerful and computational demands for speed continue to increase, optical interconnects must also advance to keep up. But pushing optical interconnects beyond what they already can do is a difficult challenge. Speed and data capacity are pressing concerns, but so are the size and complexity of the interconnects as well as their energy efficiency. Addressing these factors while engineers are reaching the ceiling for the number of transistors that can be placed on a semiconductor chip (an upper limit known as the end of Moore's Law) will take creativity and innovation.
Now, University of Washington Department of Electrical & Computer Engineering (UW ECE) Assistant Professor Sajjad Moazeni has received a three-year grant from the National Science Foundation to envision and develop a new kind of optical interconnect for data centers—one that will be compact, energy efficient, and speedy enough to support the next generation of AI, machine learning and other compute-intensive applications run in the cloud.
"All sorts of applications, from ChatGPT all the way to the processors used to run simulations for physics, chemistry, and medicine rely on powerful computers and the infrastructure located in data centers,'" Moazeni said. "We need to scale up optical interconnects as we build newer applications and newer, more powerful processors. This requires a lot of research and a paradigm shift."
Coherent co-packaged optical interconnects
To help shift the paradigm for optical interconnects, Moazeni is taking a technique used in long-distance fiber optic telecommunication lines called "coherent optical transmission" and applying it to the optical interconnects used in data centers. Most interconnects used in data centers today are optical; however, these connectors only transmit data on the amplitude of light signals streaming through fiber optic cables. By contrast, coherent optical transmission imprints data on both the amplitude and the phase of the light, which doubles the amount of information the optical interconnect can send.
One barrier to using this technique inside data centers is that coherent optical interconnects take more energy while transmitting data because more signal processing is needed to recover the phase and amplitude and decouple them from each other. Moazeni is addressing this challenge by embedding the optical interconnect into a silicon photonic chip (an optical and electronic circuit on a silicon microchip) that is then "co-packaged" alongside the CPU. Placing the electro-optical conversion as close as possible to the CPU has several benefits, including making the optical interconnect fast, compact, and energy efficient. Moazeni said he expects that using coherent optical transmission and co-packaging the optical interconnect with the CPU will increase total data transfer speeds by 10 times and in a much smaller form factor. It will also make this new kind of interconnect almost 10 times more energy efficient as compared to standard coherent optical interconnects in use today.
"What we are trying to do in this research is merge two different domains of interconnects," Moazeni said. " We are taking the coherent concept and principle from telecommunication and making it suitable for shorter links inside the data center."
Education, commercialization and next steps
Moazeni is leveraging this work to provide educational opportunities for students at UW ECE. Graduate students and undergraduate students will contribute to the research under Moazeni's supervision as he develops this coherent co-packaged optical interconnect. And because blending silicon photonic chips with optical devices built into electronics is a new area, Moazeni is building the topic into courses that he teaches. He is also putting together related tasks appropriate for high school student interns. He said he hopes that through coursework and hands-on projects, students will become excited about this technology, and it will encourage them to dig deeper into studying silicon photonic chip design.
Over the next three years, Moazeni and his research team will build a prototype of the coherent co-packaged optical interconnect chip with assistance from GlobalFoundries, which is providing silicon photonic technologies and the fabrication process for the chip through a university partnership program. Moazeni also plans to commercialize this interconnect. He anticipates that it could be in data centers five to 10 years from now, depending on the market.
"We are combining two future trends together," Moazeni said. "One is co-packaged optics, the other is bringing coherent optical interconnects into the data center. And it's also important to note that we're building this device. It is not just computer simulations or abstract papers. This research is going to result in actual chips in the lab being tested, verified, and characterized."
Learn more about this research on the NSF website and more about UW ECE Assistant Professor Sajjad Moazeni on his UW ECE bio page.
More information:
www.nsf.gov/awardsearch/showAw … storicalAwards=false
Provided by University of Washington - Department of Electrical & Computer Engineering