Recently, Academician Dai Qionghai and Assistant Professor Wu Jiamin of the Department of Automation and Associate Professor Fang Lu and Associate Researcher Qiao Fei of the Department of Electronic Engineering of Tsinghua University jointly tackled key problems and achieved key achievements in the chip field. The ACCEL photoelectric computing chip developed by them has achieved the same accuracy of existing high-performance chips in a number of complex intelligent vision tasks, while increasing the real measured force by more than 3,000 times and the energy efficiency by more than 4 million times for the first time in the world, opening up a new path for the research and development of ultra-high performance chips. The results are in the journal Nature.

After years of exploration, the Tsinghua University team creatively proposed the architecture of pure analog electric fusion optical chip to make photoelectric computing chip. Using the advantages of optical computing performance, it breaks through the physical bottleneck of data conversion speed, accuracy and power consumption, and solves the three difficulties of large-scale integration, efficient nonlinear and high-speed photoelectric interface on a single chip.

ACCEL photoelectric computing chip not only successfully completed the traffic scene discrimination, high-resolution image recognition, low-light computing and other intelligent vision tasks in the actual scene, and the time and power consumption is less than one-thousandth of the current high-performance chip. At the same time, based on the developed adaptive training algorithm, the chip can achieve performance improvement of several orders of magnitude compared with advanced process chips by using only hundred-nanometer process accuracy.