GPUs and other chips used to train AI communicate with each other within a data center through an “interconnect.” However, these interconnections have limited bandwidth, which limits the performance of AI training. A 2022 study found that AI developers typically struggle to use more than 25% of their GPU's capacity.
One solution could be new interconnects with much higher bandwidth, according to Vivek Raghunathan, CEO and co-founder of startup Xscape Photonics. The secret sauce, he says, is silicon photonics, silicon-based materials that manipulate light to transmit data.
“Xscape has created a platform that connects disparate computing elements in a sustainable way while providing the best possible performance,” Raghunathan said in an interview with TechCrunch. “The core of this platform expansion relies on energy-efficient and cost-effective systems that do not yet exist in the industry.”
Xscape is based in Santa Clara, in the heart of Silicon Valley, but its roots lie in the labs of Columbia University, where three professors (Alexander Gaeta, Keren Bergman, and Michal Lipson) are working on terabytes of data. He invented a technology that he thought could be used to transmit. Overwrite.
The trio spun off Xscape in 2022 after recruiting Raghunathan and Yoshitomo Okawauchi, a laser engineer and Gaeta's longtime colleague. Mr. Raghunathan joins the company via Broadcom, where he helped establish the silicon photonics team and served as the company's silicon photonics product manager at Intel.
Traditional interconnects consist of metal wires that transmit data in the form of electrical signals.
Metal-based interconnects require large amounts of power and generate large amounts of heat. The conductivity of the medium limits the bandwidth. Data centers with fiber-optic links between components also have to convert the interconnect's electrical data to optical data and back to optical data, which introduces delays.
In contrast, silicon photonics like Xscape consume minimal power and generate negligible heat.
“Historically, we used optical communications primarily for long-distance fiber optic systems,” Raghunathan says. “However, recent advances have enabled the integration of optics-on-a-chip in the form of silicon photonics, bringing the optical interface from the electronic to the optical plane all the way inside the chip.”
Xscape's first product is a programmable laser that powers fiber-optic interconnects in data centers, specifically the links between GPUs, AI chips, and memory hardware. Raghunathan argues that lasers can use different colors of light (or wavelengths) to send multiple data streams along the same link without interference.
“Crowded electrical systems tend to have crosstalk, interference, and other problems,” he says. “But within the optical domain, data can be modulated with different colors, wavelengths, or channels, all copropagating within the same wire or fiber and not interfering with each other.”
Assuming the technology works as advertised, Xscape faces the same challenge as most hardware startups: manufacturing and selling products at scale. In a potential rival to photonics rivals like Ayar Labs and Celestial AI, Xscape's lasers can be made using the same equipment used to make microelectronics in cell phones and laptops. .
First generation lasers can only emit between 4 and 16 colors. However, Xscape is already planning an improved version that can output up to 128.
Xscape said it is “actively working” with 10 customers on potential deployments, ranging from vendors to hyperscalers, and with Cisco, whose venture arm invested in a recent $44 million Series A round. It says it has secured funding from Nvidia. This investment is not strategic. In other words, businesses are not currently customers. But Raghunathan points out that Cisco is one of the world's largest sellers of optical networking components.
“This reflects Cisco and Nvidia's confidence in the value we bring to this ecosystem,” Raghunathan said.
The latest round of funding was led by IAG Capital Partners and brings the company's total raised to $57 million. Raghunathan said the proceeds will be used to grow Xscape's 24-person team and expand its manufacturing of laser and related photonics technologies.
“This funding will enable Xscape to push the boundaries of our platform and integrate it with simulation, high-performance computing, and AI software to enable customers across all industries to take innovation to new heights. ” said Raghunathan.
Xscape certainly does a good job of that. Besides Ayar and Celestial, the company competes with Intel in the multibillion-dollar silicon photonics market. Intel claims to have shipped more than 8 billion photonics chips and 3.2 million on-chip lasers since 2016.