Want to call someone a quick-thinker? The easiest cliché for doing so is calling her a computer – in fact, “computers” was the literal job title of the “Hidden Figures” mathematicians who drove the success of early NASA.
But while modern electronic computers are far faster than even those mathematical geniuses were with paper, pencils, and slide rules, there’s another type of computing that leaves it eating space dust – optical computing, as fast as the speed of light because instead of using electrons, it uses light itself.
The need for faster computing isn’t just to satisfy impatient consumers screaming to stream movies while video-chatting, VR gaming, and 3D printing. It’s to handle the Big-Bang-level of data expansion of the modern digital world. The humble graphics processing units (GPUs) in standard computers simply can’t scale or work quickly enough to meet such massive overflow. Even worse, as The Smithsonian Magazine and Sustainability Magazine report, AI data centers filled with GPUs are consuming massive amounts of electricity (mostly from nonrenewable sources) and water (often in arid regions).
Enter Yufeng Zhang (Photonics Group at Aalto University’s Department of Electronics and Nanoengineering) and Xiaobing Liu (Chinese Academy of Sciences, Changchun). In their Nature Photonics paper “Direct tensor processing with coherent light,” lead author Zhang and colleagues reveal their computational method employing a single propagation of light, hence what is known as "single-shot tensor computing at light speed in a parallel optical matrix–matrix multiplication" (POMMM).
That kind of breakthrough brings artificial general intelligence one major step closer.
“Our method performs the same kinds of operations that today’s GPUs handle,” says Zhang, such as “convolutions and attention layers, but does them all at the speed of light.” Instead of standard binary coding information utilizing electronic circuits using signals of ones and zeros, the researchers use the amplitude and phase of light waves to store, process, and communicate data. Doing so not only saves energy, but provides a gigantic expansion of bandwidth and processing speed – in fact, it performs many of its processes simultaneously.
The combined interactions of light fields naturally produce mathematical operations including tensor multiplications. Tensor processing organizes data into multi-dimensional arrays (think of a 2D array as a single shelf in a filing cabinet, and a 3D array as a filing cabinet with several shelves stacked vertically) called data tensors. Tensor processing is everywhere in leading-edge technology, particularly in data analytics and artificial intelligence, including natural language processing and image recognition. It’s at the heart of deep learning algorithms.
However, current optical methods fare poorly with tensor-based tasks, rendering them inefficient for neural networks and other highly complex applications. That’s where the Aalto method offers a major leap forward, by employing multiple wavelengths of light which can handle even extremely demanding tensor operations.
To explain the difference between electronic and his team’s optical computation, Zhang uses an analogy of package-sorting to explain the difference between electronic and optical computation.
“Imagine you’re a customs officer who must inspect every parcel through multiple machines with different functions and then sort them into the right bins,” says Zhang. “Normally, you’d process each parcel one by one. Our optical computing method merges all parcels and all machines together [by creating] multiple ‘optical hooks’ that connect each input to its correct output. With just one operation, one pass of light, all inspections and sorting happen instantly and in parallel.”
Tensor parallel (also called tensor model parallel or TP) processing is an execution strategy for deep learning that deploys multiple devices to address components of the same model, allowing them to compute larger models more quickly. Put another way, many minds make lighter work. Or, in this case, speed-of-light work.
According to Zhipei Sun, leader of Aalto University’s Photonics Group, the new method can work on nearly any optical platform, and his team plans “to integrate this computational framework directly onto photonic chips, enabling light-based processors to perform complex AI tasks with extremely low power consumption.”
If the method works, they expect to deploy it for integration with existing hardware and major platforms within five years. Doing so “will create a new generation of optical computing systems, significantly accelerating complex AI tasks across a myriad of fields.”
Source: Aalto University
