We are pushing towards a bright future that will harness the power of ever-more powerful and energy-saving computing. The relentless energy demands of artificial intelligence (AI) are driving an exciting new change that will see AI’s energy-intense operations not simply managed, but considerably reduced. By using light-based chips, known as photonic chips, instead of conventional computer chips that use electrons, this dream is becoming a reality. It is not science fiction. Optical (light-based) computing, which has been a topic of interest for decades, is at the forefront of innovation today as photons are poised to compete with electrons in powering the deep-learning era.
And for six decades, it has been the electron, not the photon, that has underpinned the field of computing. That is soon to change. Today, researchers are increasingly exploring ways to process information using photons. Their advantages are clear: photons carry more information, operate faster and can be more energy-efficient than their electronic counterparts. The key to reaping these benefits lies in the ability to perform matrix multiplication, a fundamental operation critical to the way neural networks work.
Matrix multiplication powers AI, essential for both training neural networks and manipulating new data. This new leap to use light for such computations doesn’t just innovate, it could revolutionise. Photonic systems, or optical neural networks (ONNs), have suggested that, in fact, photons might be better for matrix operations than electrons.
Recently – most recently in 2017 – the biggest milestone has been the addition of optical computing to the legacy of modern computing. A team at the Massachusetts Institute of Technology announced they had found a way to use matrix multiplication to solve fundamental problems in computing using an optical chip. What they achieved was using beams of light to encode data, and then guiding these beams through multiple optical components to perform the computation.
However, a 2017 seminal study, often described as a turning point for the field of optical computing, provoked a new burst of creativity. Researchers developed and demonstrated new classes of optical computing systems with prospects of scaling up throughput. For instance, HITOP network exhibits promising potential of scaling up data processing. It relies on additional dimensions of light – time, space and wavelength – to process more data at lower energy consumption.
For all this progress, optical computing still has some way to go before it can compete with the performance of electronic systems. Even the output of the most advanced optical chips cannot come close to that of electronic giants such as Nvidia. Nonetheless, the advantages that the technology has the potential to offer could be too compelling to ignore for applications that could profit from its unusual qualities.
But it could turn out that the immediate future for optical AI systems is in specific applications, where their distinctive advantages can be used most effectively. Another promising area is in improving wireless communications, where optical networks could reduce the problems of interference and energy use. Most of the researchers interviewed saw the path forward as difficult but hopeful: it is possible, and desirable, to build an optimum optical neural network that can outperform electronic systems.
In the decades to come, as we continue to explore the possibilities of the deep-learning revolution, the development of light-based chips offers a glimpse of what’s possible: an optical computing future might truly change how we build computers – and ultimately how we power the technologies of our AI-driven future.
At the centre of it all is the matrix – not the sci-fi one, but the mathematical one. Matrix multiplication is a key operation in AI, and forms the basis of how neural networks are trained to learn and make decisions. Moving this operation from electrons to photons could bring tremendous advantages: matrices of brightness could be pushed through optical chips much faster than electrons can move through silicon. In other words, by migrating matrix multiplication from the electronics laboratory to the laser lab, researchers are taking a big step towards a future where optical chips could become the next computing paradigm for AI and beyond.
In the end, with this new ocean we are about to enter, the synthesis of light and computation is not just another vindication of human ingenuity – it foreshadows a future in which technological abundance and sustainability go hand in hand. The hundred-year history of optical computing is the story of how we got here – and the story of where we might end up.
© 2024 UC Technology Inc . All Rights Reserved.