Massachusetts Institute of Technology (MIT) researchers have developed a new photonic chip that can perform key computations of deep neural networks on a chip, promising energy-efficient machine-learning computations. The chip demonstrates 96% accuracy in training and 92% during inference, comparable to conventional electronic hardware.
Deep neural networks are being deployed for cutting-edge AI applications, but their high energy consumption is a major concern. Photonic hardware processes information using light, potentially addressing this issue. However, the technology still relies on electronic components, which slows down processing speeds.
A research team led by Dr. Dirk Englund made a breakthrough in developing photonic processors that can handle both linear and nonlinear operations. Nonlinear functions are crucial for DNNs to learn complex patterns. To overcome the challenge of nonlinear optics, the team designed devices called nonlinear optical function units (NOFUs) with electronic and photonic components on a single chip.
The new chip features three NOFU layers, enabling nonlinear functions while reducing energy consumption. The device siphons off light to photodiodes to convert it into electric current, eliminating the need for an amplifier. This approach allows the team to achieve ultra-low latency and train DNNs without excessive energy expenditure.
The research findings were published in Nature Photonics, and the team is now working on developing algorithms to improve training efficiency and energy consumption. The MIT photonic chip has the potential to power ultrafast AI applications with reduced energy requirements, making it an exciting breakthrough in the field of AI computing.
Source: https://interestingengineering.com/science/energy-efficient-photonic-chip-mit