Key takeaways
- Arago raises $26 million to develop its light-based AI chip, Jef, aimed at reducing energy consumption.
- The startup seeks to lower AI inference energy usage by 90% compared to traditional GPUs through innovative technology.
- Jef combines analogue and digital electronics with silicon photonics for high efficiency and performance.
- Founded by experts in physics and AI, Arago offers a unique perspective in the photonic computing sector.
- CEO Nicolas Muller emphasizes challenges in creating a reliable, high-performance photonic computing chip for AI ecosystems.
French chip startup Arago has secured $26 million in seed capital to create its light-based AI accelerator, known as Jef.
The company’s goal is to lower AI inference energy consumption by a factor of 10 compared to current GPUs by integrating analogue and digital electronics, silicon photonics, and free-space optics, as reported by the EE Times.
Unlike most photonic computing startups, which experts in optics or photonics typically establish, Arago was founded by individuals specialising in physics and AI, a fact that Arago CEO Nicolas Muller believes provides the team with a unique viewpoint.
Arago has developed a photonic AI accelerator that is designed for high energy efficiency and throughput. By utilising light as the basis for computation, the chip optimises performance per watt and dollar, enabling AI workloads to run with considerably reduced energy and cost requirements.
Based on Arago’s exclusive multi-physics computing core, high-volume AI tasks characterised by 𝑛^3 complexity can be accelerated and performed with extremely low power consumption.
Arago refers to its inference chip, Jef, as a “multiphysics processor” since it integrates features of both analogue and digital electronics alongside silicon photonics and free-space optics.
Muller believes that Arago faces significant challenges, including the development of a photonic computing chip that operates reliably and is sufficiently high-performing and user-friendly to be integrated into AI technology ecosystems.
“We started from the top, meaning: what does it take to infer a model very efficiently?” Muller told EE Times. “Then we turned that into a data sheet for a product.”
“Historically, companies that were tackling this space were born around 5 to 10 years ago, where the technical and scientific landscape was extremely different,” Muller said. “You had pure optical technologies and different technologies that are based on free space optics. What we built is a technology that overcomes all of those challenges by only using optics at very specific places.”
The company states on its website that contemporary AI models are increasingly limited by the compute and memory capacities of traditional silicon-based accelerators. Improving AI efficiency and deployment necessitates a shift away from conventional transistor technology.
