Photonics, which involves the generation, detection, and manipulation of light in the form of photons through emission, transmission, modulation, signal processing, and more, has been around for decades. In telecommunications, this technology has enabled high-speed data transmission over optical fibers, revolutionizing how we communicate by allowing vast amounts of information to travel long distances quickly and efficiently.
However, as the demands of modern computing, particularly in AI and high-performance computing clusters, continue to skyrocket, traditional photonics are being pushed to their limits.
“The demand for large language model training has never slowed down,” says Dr. Vivek Raghunathan, CEO and Co-founder of Xscape Photonics. “The problem has shifted from device-level performance to actively designing data centers with faster interconnections to improve performance. Currently, a limiting factor for performance is how fast GPUs (AI accelerators) can communicate with each other and with the memory, compute, and networking stacks.”
With a PhD in photonics from the Massachusetts Institute of Technology and a background in advancing AI accelerators to memory interconnect technology and bringing light-based technologies to market, Vivek was uniquely positioned to tackle this emerging challenge head-on.
Inspired by the work of three Columbia University professors, he founded Xscape Photonics together with them.
Recognizing the potential of Xscape Photonics’ innovative technology, Cisco Investments is thrilled to add Xscape Photonics to its portfolio, supporting their mission to revolutionize data transmission.
Exploring Multi-Wavelength Photonics Technology
Xscape Photonics is building photonic integrated chips (“PICs”) that enable ultrahigh-bandwidth connections within data centers and high-performance computing systems. What sets Xscape Photonics apart is their multi-wavelength programmable laser technology. In short-distance data connections, typically, one laser is used for each light color, or wavelength. However, Xscape Photonics can generate hundreds of different colors of light from a single laser, significantly increasing efficiency and reducing complexity.
The foundation of Xscape Photonics' technology is rooted in the pioneering research conducted at Columbia University. Xscape Photonics co-founders Professor Alex Gaeta, Professor Keren Bergman, and Professor Michal Lipson developed this multi-wavelength laser platform to tackle the difficulty in efficiently moving massive amounts of data across AI accelerators and in and out of memory stacks. Their solution effectively addresses the bottlenecks that limit the performance of AI and high-performance computing clusters.
"The core of our differentiation relies on having a scalable and cost-optimized multi-color laser photonics platform that encodes each color with a data stream and integrates it into a single fiber," Vivek says.
Leveraging Color to Increase Efficiency
By generating multiple data streams onto different colors of light with a single laser, Xscape Photonics technology significantly decreases overall cost, power, and footprint.
"The focus was on figuring out how we can solve the problems that the hyperscalers are seeing in terms of addressing the supply capacity and also the bandwidth density problems within the data center with a programmable multi-color laser technology," Vivek noted.
Additionally, as data centers and AI clusters grow, the ability to scale bandwidth without a proportional increase in physical infrastructure becomes crucial. Xscape Photonics' technology seamlessly expands to help meet these demands, future-proofing data centers and enabling them to handle the increasing needs of modern computing efficiently and cost-effectively. This makes data centers as compact and efficient as individual chips.
Scaling for the Future
In the future, Vivek sees data centers operating as seamlessly integrated systems, much like a single, large-scale computer chip. This would require the customers to reimagine the fabric of connectivity across AI accelerators and the compute, memory and networking stacks. He thinks customers could think about disaggregated pools of AI Accelerators and memory stacks for efficient resource utilization, while providing them the ability to mix and match resources for different customer workload needs.
"Our goal is to enable our customers to build a sustainable data center fabric that can scale the performance of AI while keeping power consumption and costs within reasonable amounts," Vivek says, reflecting his forward-thinking approach and dedication to pushing the boundaries of what's possible in data transmission and data center design.
Cisco Investments is excited to welcome Xscape Photonics to its portfolio of companies, and we look forward to helping them accelerate their innovation to market.