A young American company offers another way to operate a generative artificial intelligence. Its model Un-0 produces images using a system of coupled oscillators, instead of relying solely on the classic numerical calculations of graphics processors. The promise: a thousandfold reduction in energy consumption for equal performance.
Oscillators are systems that evolve over time according to a frequency, a phase, and interactions with their environment. In this architecture, their physical states replace part of the operations usually performed by transistors. The computation then relies on the collective dynamics of the system.
This approach belongs to the field of physical computing. It seeks to directly exploit the properties of a hardware medium, such as signal variations or the synchronization of small oscillators. The goal is to perform certain operations with fewer steps, and therefore less energy.
Un-0 was tested on 64 by 64 pixel images from the ImageNet database. The company reports an FID score of 6.74, a metric used to compare the statistical quality of generated images. This result is comparable to that of several early classic image generation models.
However, the model is not currently executed on a dedicated physical chip. It runs on a computer simulation of an oscillator architecture. The announced energy gains therefore correspond to a hardware prospect, not to a consumption already measured with a complete prototype.
The operation also differs from the most common diffusion models. Instead of progressively reconstructing an image from noise with a vast neural network, the system exploits the collective evolution of its oscillators to produce a coherent representation.
The company estimates that this architecture could greatly reduce the energy needed for inference. Some communications mention a factor potentially reaching 1,000, but this value remains for now a projection depending on the hardware, software, and chosen comparison.
Moving to larger images constitutes a major challenge. It will also be necessary to build the circuits, limit errors, program the interactions between oscillators, and compare performances. A promising simulation therefore does not yet guarantee an industrial alternative to GPUs.
This path nevertheless remains interesting as image generation increases the demand for computing and electricity. It is representative of a broader trend: designing artificial intelligence systems that directly use physics to reduce the energy cost of operations.