SIOM
Scientists from the Shanghai Institute of Optics and Precision Mechanics under the Chinese Academy of Sciences have announced the creation of the first processor for ultra-high-parallel optical computing.
It is noted that the theoretical performance of the new optical processor is 2 thousand 560 terra operations per second (TOPS) at an optical clock frequency of 50 GHz. Typically, optical processors transmit information at a single wavelength of light. However, the design of the chip, developed in China, divides the laser beam into more than 100 channels that simultaneously pass through one integrated circuit the size of a fingernail.
According to the project manager Xie Peng, his team managed to achieve a significant increase in bandwidth while maintaining the chip size and clock frequency. This was made possible by using soliton microcombs — tiny ring resonators that divide a dense laser beam into a number of evenly spaced spectral «teeth», each of which carries an independent stream of bits.
Since light is not affected by resistive heating, which represents a problem for electronic circuits, parallel strips can run adjacent to each other with minimal energy loss and a small risk of heating of individual areas. The developers claim an optical bandwidth of more than 40 nm, low energy losses and fully reconfigurable routing. This allows the processor to perform a variety of tasks, from image recognition to real-time signal processing.
Chinese researchers say that the large number of parallel channels running through the chip could allow modern artificial intelligence models work more energy efficiently than with GPUs. Neural networks, which rely on many identical mathematical operations, are naturally mapped onto a multi-band chip structure. The low latency makes the technology attractive for edge devices ranging from high-frequency trading servers to swarms of drones, where milliseconds count and power budgets are limited.
In addition to AI systems, this chip can speed up modeling processes, medical imaging, and other large-scale workloads that are difficult to perform on electronic equipment alone. The developers emphasized that they were able to significantly increase the number of bandwidths without limiting the speed.
The results of the study were published in the journal eLight
Source: Interesting Engineering