China has presented LightGen, a photonic processor designed to accelerate generative artificial intelligence models, which is positioned as the most powerful today.

According to data published in the magazine Science, In certain tasks it exceeds more than 100 times the performance of an Nvidia A100 GPU in relation between calculation and energy consumption.

It uses light pulses to process information, eliminating electrical friction and allowing a scale of processing greater than today.

Jiao Tong and Tsinghua’s photonic processor

It should be noted that LightGen is the result of a joint project between researchers from Shanghai Jiao Tong University and Tsinghua University.

This is not a commercial chip, but rather a laboratory prototype described as an optical computing platform for AI. Therefore, its design is oriented towards typical loads of generative models: creation and transformation of images, video and 3D scenes.

In essence, the chip replaces electronic transistors with photonic “neurons”. These units manipulate beams of light to perform the operations that electronic circuits execute in a classic neural network.

The objective is to take advantage of the properties of photons – speed and lower heat dissipation – to gain efficiency when the computing volume is very high.

More than two million photonic neurons on a chip

One of the advances that the authors point out is density. LightGen uses 3D packaging to integrate more than two million artificial neurons into a surface area about a quarter of a square inch.

In previous optical processors, the scale was typically a few thousand neurons, enough for simple classification tasks but not for complex generation.

With this density, The chip can now address tasks such as generating high-resolution video or handling 3D modelswhich typically run on GPU farms.

Each photonic neuron acts on the light that circulates through the chip, adjusting parameters such as intensity or phase to implement network operations.

It is important to mention that the work also introduces the concept of “optical latent space”. In generative models, latent space is the compressed representation of information from which images or other content are generated.

In LightGen, this representation is manipulated directly with light and, to achieve this, the design uses ultrathin metasurfaces and fiber matrices.

These elements allow multidimensional data to be compressed and processed without having to fragment the images into blocks.

This better preserves the statistical structure of the input data and reduces the number of steps required to generate the result. Researchers have tested LightGen in various generative AI scenarios.

The system has been able to produce semantic images with good quality and perform 3D manipulations comparable to those of advanced electronic neural networks.

Current limitations of LightGen

The work itself highlights several limitations, which is that the system depends on external lasers to generate and control the optical signal, which complicates assembly and makes the solution more expensive.

Additionally, manufacturing the chip requires specific processes that are not integrated into the current semiconductor industry.

Integrating LightGen into real data centers would require addressing issues of scaling, cost, integration with existing hardware, and long-term reliability.

That is why the authors present the chip as a promising line of research, not as an immediate replacement for current GPUs.

Potential impact on generative AI

If this type of design is consolidated, one of the clearest effects would be the reduction of energy consumption by using generative AI.

And training and executing large models today requires a lot of power and energy; A more efficient specific accelerator could reduce costs in data centers and make the deployment of advanced models more affordable.

LightGen also shows that China is not only competing in electronic chips, but is exploring alternative avenues such as photonic computing.

It is likely that, if these technologies advance, we will see hybrid systems: electronic processors for certain parts of the calculation and optical modules for very specific phases where calculation with light provides a measurable advantage.

According to Science, researchers from Shanghai Jiao Tong University and Tsinghua have shown that a photonic chip can be built capable of executing generative AI tasks with much higher efficiency than electronic hardware.

The transition from these prototypes to commercial products will depend on how the technical and economic barriers are resolved in the coming years.