Chen Hongwei's research group from the Department of Electronics has made progress in the field of integrated optical computing
Tsinghua News, February 20th. Machine learning technology has been widely used in the field of high-performance information processing. At the same time, the requirements for computing capacity, computing speed, and energy consumption are getting higher and higher when solving various complex tasks. However, the calculation speed of existing hardware is severely limited by the traditional von Neumann architecture. As the time required for the calculation process increases, the calculation efficiency will become low and the energy consumption will be greater. In recent years, photonic methods have shown extraordinary potential in performing deep learning processes involving complex calculations. Many research institutions at home and abroad have successively proposed new architectures for integrated photonic neural networks, such as based on Mach-Zehnder interferometer (MZI), micro Ring resonator (MRR) and photonic neural network designed by WDM system, etc. However, at this stage, integrated photonic neural networks generally have the problem of limited large-scale expansion of computing units, which seriously limits the further improvement of computing capacity.
Figure 1. Schematic diagram of integrated diffractive photonic neural network
The research group of Professor Chen Hongwei from the Department of Electronics, Tsinghua University proposed an integrated diffractive photon neural network (DONN) based on subwavelength structure, which overcomes the volume limitation of spatial diffractive photon neural network, not only greatly improves the integration of computing units, but also reduces the Errors due to bulky bulk optics and system calibration. For other integrated photonic neural networks, the DONN chip gets rid of the restriction of the number of waveguides, and it is easier to realize the large-scale expansion of computing units on the chip, thus solving the problem of high computing capacity of integrated photonic neural networks. The DONN optical computing chip implemented in this study has a computing throughput of 1.38×10 4 TOPS (TOPS: Trillions of operations per second, trillion operations per second), and a chip computing power density of up to 10 16 FLOPS/mm 2 (FLOPS: Floating-point operations per second, the number of floating-point operations per second), the energy consumption is about 10 -17 J/FLOP (FLOP: Floating-point operations, floating-point operations). The DONN chip has complete domestic independent intellectual property rights, and the preparation process is also completely realized in China. It is compatible with the standard complementary metal oxide semiconductor (CMOS) process and meets the conditions of large-scale and low-cost production.
Chen Hongwei's research group has systematically completed the whole process verification of the integrated diffractive photonic neural network chip, including theoretical exploration, simulation verification, structural design, layout drawing, chip processing, packaging and system error compensation. This achievement will significantly improve the practicability of integrated photonic neural network chips, and it is expected to realize the integration of multiple DONN computing units on one chip, so that the on-chip optical computing system has stronger processing capabilities, which will greatly promote integrated optical computing, photonics The rapid development of intelligence and other fields.
Figure 2. DONN package physical picture, chip structure and test results
The achievement was published in Nature Communications under the title of
" Photonic machine learning with on-chip diffractive optics
