From Square to Cube: AI Hardware Processing Moves Towards 3D, Enhancing Processing Power

In a paper published in the journal "Nature Photonics," researchers from the University of Oxford, in collaboration with partners from the University of Münster, Heidelberg University, and the University of Exeter, reported the development of integrated photonic hardware that can process 3D data, greatly enhancing the parallelism of AI data processing tasks. Traditional computer chip processing efficiency doubles every 18 months, but the processing power required for modern AI tasks currently doubles every 3.5 months. This means that there is an urgent need for new computing paradigms to meet the growing demand. One approach is to use light instead of electrons, which allows for the representation of different datasets using different wavelengths and enables parallel processing of multiple computations. Indeed, in groundbreaking work published in the journal "Nature" in 2021, many of the same authors demonstrated an integrated optical processing chip that performs matrix-vector multiplication (critical for AI and machine learning applications) at speeds far exceeding the fastest electronic methods. This work led to the creation of the optical AI company Salience Labs, a spin-off from the University of Oxford. Now, the team has added an additional parallel dimension to the processing capabilities of the optical matrix-vector multiplier chip. This "higher-dimensional" processing is achieved by encoding data using multiple different radio frequencies, achieving levels of parallelism previously unattainable. As a test case, the team applied their novel hardware to the task of assessing the risk of sudden cardiac death in patients with heart disease. They were able to successfully analyze 100 electrocardiogram signals simultaneously and accurately determine the risk of sudden death with a 93.5% accuracy rate. The researchers further estimate that even with a modest expansion of 6 inputs × 6 outputs, this approach can surpass existing state-of-the-art electronic processors, potentially providing a 100-fold increase in energy efficiency and computational density. The team expects that in the future, further improvements in parallel computing can be achieved by leveraging additional degrees of freedom of light, such as polarization and mode-division multiplexing. Dr. Bowei Dong, the first author from the Department of Materials at the University of Oxford, said, "We initially thought that using light instead of electrons could only increase parallelism by using different wavelengths, but later we realized that using radio frequencies to represent data also opens up another dimension, enabling super-fast parallel processing for emerging AI hardware." Professor Harish Bhaskaran, a professor in the Department of Materials at the University of Oxford and co-founder of Salience Labs, who led this work, said, "Now is an exciting time for fundamental research in AI hardware, and this work is an example of how we can surpass what we thought were the limits." (Note: The content has been optimized and rewritten to meet the originality standards. HTML tags have been retained, while style and class attributes have been removed. Empty

tags have been deleted.)