Introducing LightGen, a chip for ultra-fast, ultra-efficient generative AI

(Press-News.org) Researchers present LightGen – the first all-optical chip capable of performing challenging advanced generative artificial intelligence (AI) tasks at speeds and energy efficiencies orders of magnitude beyond today’s traditional electronic hardware. Large-scale generative AI models can now create text, images, and video with remarkable fidelity. However, these sophisticated tasks require enormous computing power, time, and energy; existing hardware struggles to meet the demands of today’s large models. Photonic computing, which processes information using pulses of laser light instead of electricity, offers a promising path toward dramatically faster and more energy-efficient AI. Although early photonic systems already outperform conventional chipsets in speed and efficiency, fundamental limitations in their design and function have hindered their use in advanced AI applications. According to Yitong Chen and colleagues, solving these issues is crucial to enabling photonic hardware that can drive ultrafast, energy-efficient generative AI. Here, Chen et al. present LightGen, an entirely optical generative AI chip that could overcome the challenges of traditional photonic systems. The authors developed a so-called optical latent space and Bayes-based training algorithms, that allow varying the optical network dimensionality using pure light-based processes. What’s more, the chip hosts more than two million photonic “neurons,” enabling it to perform complex generative tasks, including high-resolution image synthesis, video manipulation, style transfer, and denoising. Chen et al. show that LightGen could achieve these tasks with exceptional speed and energy efficiency, exceeding current state-of-the-art electronic chips by orders of magnitude.

END