What happened
Companies developing processors that compute with light rather than electronics are attracting new interest as AI inference pushes up data center power demand, IEEE Spectrum reported. The combined market valuation of firms in the sector has reached US $2.22 billion, compared with $0.4 billion in 2022, and new optical computing ventures have emerged in the United States, China, Germany and Israel.
Lumai, based in Oxford, England, uses a free-space design in which light passes through air and a programmable spatial light modulator instead of on-chip waveguides. Its head of product, Phillip Burr, said this approach is what will allow the company to outperform GPUs, and that it completes an entire vector-matrix multiply in one optical cycle. Optalysys, also in England, is placing optical elements in data center networking paths so that data can be processed while it travels.
Optical computing companies are targeting energy efficiency gains of 10 to 80 times over current GPUs. Cornell professor Peter McMahon, whose team in 2022 showed an optical neural network using under one photon of energy per multiplication, cautioned that such results leave out the cost of encoding data into light. He said optics only begins to win at vector sizes of roughly 1,000 to 10,000 elements, and that parts of inference work cannot be moved to optics at all.
Context
Computing with light has been studied for more than 70 years but stayed mostly in research labs. In 1978, Joseph W. Goodman and colleagues proposed performing matrix-vector multiplication, the core operation of AI inference, using light-emitting diodes, lenses and an attenuator. Inference accounts for most AI energy use, and the industry already spends billions of dollars designing chips to run it more efficiently.
El boletín
La semana tecnológica, en un solo correo.
Las crónicas que merecen su tiempo, cada semana. Cancele la suscripción cuando quiera.
