For more than a decade, the GPU has dominated the landscape of intensive calculation. Thousands of models of artificial intelligence, from the simplest classifier to the generative multimodal systems, were trained on these massively parallel architectures. However, as the demand for power explodes, the physical limits of these chips become more and more restrictive. It is in this context that the Multiphysical calculationat the crossroads of electronics and photonics.
The principle is to use several physical supports simultaneously to process information, and no longer only electrons. The light, transported by photons, then becomes a full -fledged calculation vector. The objective is to gain in energy efficiency, calculation density, and industrial scalability capacity.
From transmission to treatment
Photonics has already conquered the field of interconnections. The optical fibers manage the transmission of data between servers, racks, or datacenters. But what is played out today is the Passage of light as a channel in light as treatment media.
This is where the paradigm of multiphysics comes in. Rather than electronic and photonic opposing, it is a question of integrate in a hybrid architecture. This approach makes it possible to benefit from the best of both worlds, namely the industrial maturity of CMOS and the effectiveness of optical calculation.
Escape the planned ia chips
The major digital players invest hundreds of billions of dollars in specialized asicsoptimized for precise models or frameworks. But these architectures evolve faster than the cycles of material development. As a result, fleas are obsolete even before their marketing.
Multiphysical calculation, by focusing on the generic (and universal) treatment of matrices, aims to bypass this structural instability. It thus becomes possible to execute various models (transformers, MOE, LLMS) with a stable material backend, compatible with Software standards (Pytorch, Tensorflow) And with traditional manufacturing chains.
It is precisely this niche that the French startup explores Aragofounded in early 2024 in Paris. Its technology is based on a Photonico-electronic architectureready to be industrialized. After nine months of existence, the startup raises 24 million euros in seed to accelerate its development.
Arago was founded by Nicolas Muller (CEO),, Ambroise Müller (CSO) And Eliott Sarrey (CTO). Located in Arcueil, his team of twelve people is divided between France, Israel and North America.
The financing tour was co-directed by Earlybird Venture Capital,, Protagonist And Tomorrow Visionarrieswith the participation of C4 ventures,, Acequia Capital,, Generatifiqand strategic business angels such as Bertrand Serlet (ex-Apple), Olivier Pomel (Datadog), Thomas Wolf (Hugging Face), Christophe Frey (Arm), David Fattal,, Jean-Laurent PhilippeAnd Pierre Boudier.
