If AI is trying to meet our expectations, it requires a longer context window, better consistency in image generation, and other progressive advances. The pace of rapid progress in this field over the past few years has led to a broad belief that we are on the forefront of creating tightly-held machines. But in reality, we seem to be against some difficult technical limitations that are slowing further advancements.
Without another breakthrough in the order of trans-architecture development, the improvements to exponential algorithms could soon become a thing of the past. The counting of model parameters and scaling of training dataset sizes has been obtained for some time, but this kind of further scaling is not practical due to computational overhead and energy consumption. However, thanks to the efforts of a group of researchers at the University of Florida, there is a possibility that mitigation from these issues is on the horizon. They developed a Light-based chip This not only speeds up commonly used calculations, but also reduces energy consumption by up to 100 times.
Comparison of conventional and optical processes of convolutional (📷: H. Yang et al.)
This chip is specially designed to handle convolution, one of the most power-hungry operations in AI. Convolution is the backbone of modern deep learning systems, allowing neural networks to recognize images, video and text patterns. Although essential, they are also very demanding hardware, often making up more than 90% of the power consumed by convolutional neural networks.
Instead of relying solely on electronics to perform these operations, the team integrated small optical components directly into the silicon chip. Using laser light and a microscopic Fresnel lens (flat, ultra-thin lens etched on the chip itself), the convolution operation could be performed using most energy. By passing optically encoded data to these lenses, the system performs the necessary Fourier transforms optically, returning the results back to the digital signal for further processing.
The prototype chip already demonstrates competitive performance, achieving approximately 98% accuracy when classifying handwritten numbers from standard MNIST datasets. The results are comparable to traditional electronic chips, but only a small portion of the power consumption. In additional testing, the system remained resilient, achieving accuracy of over 95%, even when timing delays were introduced into the input signal.
Microscopic images of the chip (📷: H. Yang et al.)
Another advantage of photonics is that it allows multiple data streams to be processed simultaneously. By using different wavelengths or colors of laser light, researchers have shown that the chips can perform parallel calculations within the same device. This technique, known as wavelength multiplexing, could provide a scalable pathway to dramatically increase AI throughput without a corresponding increase in energy usage.
If technology can be commercialized, it not only will AI models be faster, but it also promises a solution to the looming energy crisis brought about by growing demand for data centers. With the efficiency gains measured several orders of magnitude, the team's optical power chips could be something like the breakthrough needed to keep AI momentum from halting.
