Shulou(Shulou.com)11/24 Report--

CTOnews.com, November 5, after a long period of joint research, the research team of Tsinghua University broke through the physical bottleneck of traditional chips, creatively proposed a new computing framework for optoelectronic fusion, and developed the international first fully analog optoelectronic intelligent computing chip (referred to as ACCEL). According to the actual measurement, the computing power of the chip in intelligent visual target recognition task is more than 3000 times that of the current high-performance commercial chip, which opens up a new path for the research and development of ultra-high performance chip.

CTOnews.com learned from the official account of Tsinghua University that the team used "All-analog Chip Combining Electronics and Light" to describe the characteristics of the optoelectronic fusion chip, and named the chip "ACCEL" with this initials, which happens to mean "acceleration".

In this small chip, the team creatively proposed the computing architecture of optoelectronic deep fusion-- combined with optical computing based on electromagnetic wave propagation in space, and fully analog electronic computing based on Kirchhoff's law, three international problems of large-scale computing unit integration, efficient nonlinearity and high-speed optoelectronic interface have been broken on a single chip. While ensuring high task performance, ultra-high computing energy efficiency and computing speed are achieved.

In order to ensure the reliability of the experimental data, the team carried out large-scale measurement and verification. The end-to-end cross-layer collaboration technology is applied to intelligent visual interaction for the first time, and the equivalent computing power is used to accurately evaluate the energy efficiency performance and delay data.

The computing principle and chip architecture of optoelectronic computing chip ACCEL are of super high performance. Under the measured performance, the system-level computing power of the ACCEL chip is 3000 times that of the existing high-performance chips. If the time of transportation is used to compare the calculation of the information flow in the chip, then the appearance of this chip is equivalent to shortening the eight-hour Beijing-Guangzhou high-speed railway to eight seconds.

Ultra low power consumption. The system-level energy efficiency is 74.8Peta-OPS / W, which is 4 million times higher than the existing high-performance GPU, TPU and other computing architectures. Figuratively speaking, the electricity originally used for the existing chip to work for one hour can supply it for more than 500 years.

Ultra low cost. The minimum linewidth of the optical part of the optoelectronic fusion chip is only 100 nm, and the circuit part only uses 180nm CMOS technology, which has achieved several orders of magnitude performance improvement compared with the high performance chip of 7nm process. At the same time, the materials used are simple and easy to obtain, and the cost is only a tenth of that of the latter.

Compared with advanced electronic computing chips, the new architecture achieves a thousand-fold increase in computing power, a thousand-fold reduction in computing latency and a million times higher energy efficiency. At the same time, compared with the existing training methods, the optical network training scale is increased ten times and the training speed is increased a hundred times.

The relevant achievements have been published in top international journals such as Nature, Natural Photonics, Natural Communications and Scientific Progress. They have been selected as highly cited papers in the world and more than ten invention patents have been authorized at home and abroad.

The optoelectronic intelligent computing architecture and chips studied by the team provide an important support for exploring the paradigm of high computing power and low power intelligent computing in the post-Moore era.

The advent of optoelectronic fusion chip is not easy. The research team first carried out theoretical modeling and simulation verification, then constantly optimized the fully analog optoelectronic computing architecture and manufacturing process, and used photocurrent to carry out fully analog electronic calculation based on Kirchhoff's law. a new computing system of "pre-sensing + sensing + near sensing" is completed.

Basically, it can be understood that this research gets rid of the maximum resistance that traditional photonic chips can not be widely used, and the density limitation of transistors, and promotes the computing speed to an unprecedented level.

This is where ACCEL chips break the shackles of Moore's Law.

The ACCEL chip is manufactured using the existing mature technology. At present, the principle sample has been completed, and the industrial sample is being developed.

In the measurement of high-speed visual tasks, the system-level computing time of the chip is only 3/10000 of that of the existing high-performance chip architecture, and the high computing speed is only one of the advantages of this chip. In the demonstration of intelligent vision tasks and traffic scene computing, its high performance and low power consumption, it can be used in large scenes, multi-object complex artificial intelligence vision tasks, unmanned systems (such as self-driving), industrial production lines, large models and other high computing power and low power consumption scenes.

In addition, the scenarios where ACCEL chips are battery powered and energy resources are very limited are also very promising for wearable devices at the edge end.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.