In the field of AI training, the communication interconnects between GPUs and other chips are crucial for data transmission, but the bandwidth limitations of these interconnects restrict the performance of AI training. A survey in 2022 showed that AI developers typically only utilize 25% of GPU capacity.
Vivek Raghunathan, CEO and co-founder of Xscape Photonics, proposed a potential solution: adopting new interconnect technologies with higher bandwidth, with the core technology being silicon photonics, a silicon-based material that transmits data by manipulating light.
Xscape Photonics is located in the heart of Silicon Valley in Santa Clara. Its technology originated from a laboratory at Columbia University, where three professors invented a technology they believed could transmit massive amounts of data using light. In 2022, after recruiting Raghunathan and Yoshitomo Okawachi (a laser engineer and long-time colleague of Gaeta), they brought Xscape from the lab to the market. Raghunathan had previously helped Broadcom establish a silicon photonics team and served as a product manager for silicon photonics at Intel.
Image Source: Generated by AI, provided by Midjourney
Traditional interconnect technologies consist of metal wires that transmit data via electrical signals. These metal interconnects require significant power, generate a lot of heat, and are bandwidth-limited by their medium's conductivity. In data centers, electrical data between fiber optic link components must be converted to optical signals and back, introducing latency.
In contrast, Xscape's silicon photonics technology consumes very little power and generates negligible heat. Raghunathan noted that optical communication was mainly used for long-haul fiber systems in the past, but recent technological advancements have made photonic chip integration possible, bringing optical interfaces from the electronic plane into the optical plane within the chip.
Xscape's first product is a programmable laser designed to power fiber interconnects in data centers, particularly links between GPUs, AI chips, and memory hardware. Raghunathan claims that the laser can transmit multiple data streams over the same link using different colors (i.e., wavelengths) of light without interference.
Xscape faces the same challenges as most hardware startups: mass manufacturing and selling products. Compared to photonic competitors like Ayar Labs and Celestial AI, Xscape's laser can be manufactured using the same facilities as microelectronic devices for phones and laptops.
Xscape is actively engaging with 10 potential customers ranging from suppliers to hyperscale data center operators and has received funding from Cisco and Nvidia, whose venture capital arms participated in Xscape's recent $44 million Series A round. These investments are not strategic, meaning these companies are not current customers. However, Raghunathan pointed out that Cisco is one of the world's largest sellers of optical network components.
The latest round of financing was led by IAG Capital Partners, bringing the company's total funding to $57 million. Raghunathan said the funds will be used to expand Xscape's 24-person team and scale up manufacturing of its lasers and related photonic technologies.
Xscape undoubtedly faces a daunting task. In addition to competing with Ayar and Celestial, the company is also competing with Intel in the multi-billion-dollar silicon photonics market. Intel claims to have shipped over 8 billion photonic chips and 3.2 million on-chip lasers since 2016.