The unsung hero behind multimedia experiences is a tiny electronic chip—often no bigger than a credit card—that brings productivity and entertainment to life. Found in smart devices and computers, the graphics processing unit (GPU) performs numerous simultaneous calculations to render images and videos seamlessly.
“Being able to process more data in each clock cycle means you can run algorithms faster and with higher energy efficiency,” explained Bowei Dong, a Principal Investigator at A*STAR’s Institute of Microelectronics (IME). This capability stems from hardware-based 'accelerators': specially designed architectures that handle multiple data streams in a single step (parallelism). It's an essential feature for the complex, data-heavy demands of machine learning (ML).
Everyday computers used in homes and offices that rely on traditional electronic architectures falter under high parallelism, struggling to double their capacity every few months. In response, Dong and his colleagues have proposed an innovative solution: adding an extra dimension to data processing to boost parallel processing capabilities.
Think of how adding more lanes to a congested highway, each designated for a different vehicle type, allows for smoother, faster travel without bottlenecks. Similarly, advanced photonic processors can use multiple channels—like lanes—to manage several data streams on a single chip. These processors exploit the colour of light and numerous parallel optical fibres, integrating them into a data-processing photonic tensor core.
“The access to many degrees of freedom is the reason why photonics can achieve higher parallelism than electronics,” said Dong.
The architecture of a photonic tensor core for in-memory computing. Multiple radiofrequency (RF) components are multiplexed into a single optical wavelength as it enters the tensor core, which processes them in parallel. In the tensor core, each cell (inset image) contains a tuneable power splitter; a phase-change memory (PCM); two multimode interferometers (MMI); a waveguide crossing for interconnects; and a directional coupler.
Building on this architecture, Dong and collaborators from the University of Oxford and University of Exeter in the UK; and University of Muenster, Germany; introduced a third data channel using radio-frequency (RF) multiplexing, enhancing the photonic tensor core's capacity to process even more information concurrently.
The team showcased the superiority of their system using real-time heart activity recordings, processing clinical electrocardiogram (ECG) data from 100 patients. Their advanced processor handled the data with a parallelism of 100, two orders of magnitude higher than current methods. By applying an ML model to this data, the system also identified patients at risk of sudden death with 93.5 percent accuracy.
This breakthrough offers exciting real-world benefits: beyond improved real-time patient monitoring, it has broader applications in the Internet of Things (IoT)—a network of interconnected devices— and in edge cloud computing, which processes data closer to where it's generated for faster results.
The team plans to build on this promising scheme by exploring methods to encode information onto additional light channels. "Meanwhile, we will also explore new electronics-photonics integration to improve the performance of each individual computing channel," said Dong, pointing to a future where hardware evolves to meet the colossal computational demands of ML.
The A*STAR-affiliated researchers contributing to this research are from the Institute of Microelectronics (IME).
