IN-DEPTH / IN-DEPTH
Chinese research team proposes "Future" chip: computing power up to 3,000 times higher that of current high-performance commercial chips
Published: Nov 01, 2023 10:22 PM Updated: Nov 01, 2023 10:15 PM
Photo: Tsinghua Univeristy

Photo: Tsinghua Univeristy



 In the science fiction movie "The Wandering Earth," artificial intelligence system "Moss" is able to explore all solutions to save the Earth in just a few seconds.

This miraculous scene is gradually transitioning from science fiction to reality. The ultra-high-performance optoelectronic chip proposed by the research team at Tsinghua University adopts a new architecture of optoelectronic fusion, which is disruptive to existing chip technologies, the team told the Global Times on Wednesday.

The technology not only opens up a new path for this future technology to become part of daily life, but also provides inspiration for the integration of other future high-performance technologies such as quantum computing and in-memory computing with current electronic information systems.

The results, titled "Purely Analog Optoelectronic Chips for High-Speed Visual Tasks," had been published in the recent issue of the journal "Nature."

In 1965, Gordon Moore, one of the founders of Intel, proposed "Moore's Law," which has influenced the chip industry for over half a century. It predicts that the number of transistors on integrated circuits will double approximately every two years.

The semiconductor field has prospered for decades based on Moore's Law, and "chips" have become an important engine for humanity's entry into the era of digital intelligence. However, as transistor sizes approach their physical limits, Moore's Law has slowed down or even faced failure in the past decade. How to build a new generation of computing architecture and establish a "new" order of chips in the era of artificial intelligence has been a frontier hotspot of international concern.

To address this challenge, a joint research team from Tsinghua University, including Academician Dai Qionghai from the Department of Automation, Assistant Professor Wu Jiamin, Associate Professor Fang Lu from the Department of Electronic Engineering, and Associate Researcher Qiao Fei, proposed a new computing architecture that "breaks free" from Moore's Law: optoelectronic analog chips. In practical tests for visual tasks, the computing power of these chips reached over 3,000 times that of current high-performance commercial chips.

From a physical perspective, optoelectronic chips are based on a disruptive technology that is different from existing chip technologies, the research team explained in an interview with the Global Times on Wednesday. They noted that, at current stage, their work is focused on intelligent visual tasks, and they are also conducting further exploration to see if the new technology can achieve tasks of the same or even higher complexity as current chips, such as large language models.

In this small chip, the Tsinghua University research team creatively proposed an optoelectronic deep fusion computing framework. Starting from the most fundamental physical principles, it combines optical computing based on electromagnetic wave propagation in space with pure analog electronic computing based on Kirchhoff's law. It "breaks free" from the physical bottlenecks of data conversion speed, accuracy, and power consumption that constrain traditional chip architectures, and overcomes three international challenges: large-scale computing unit integration, efficient nonlinearity, and high-speed optoelectronic interfaces.

In the demonstrated intelligent visual scene tests in the paper, the system-level computing power of the optoelectronic fusion chip was thousands of times higher than that of existing high-performance chip architectures. However, such astonishing computing power is just one of the many advantages of this chip. In the intelligent visual tasks and traffic scene calculations demonstrated by the research team, the system-level energy efficiency (the number of operations that can be performed per unit of energy) of the optoelectronic fusion chip reached 74.8 Peta-OPS/W, which is over 4 million times that of current high-performance chips. In other words, the amount of electricity that can power existing chips for one hour can power this chip for over 500 years.

One key factor currently limiting chip integration limits is the heat dissipation problem caused by high density. The optoelectronic fusion chip, which operates at ultra-low power consumption, will greatly improve the chip's heat dissipation problem and bring all-round breakthroughs to the future chip design. Furthermore, the minimum linewidth of the chip's optical part is only in the hundreds of nanometers, while the circuit part uses 180 nanometers Complementary metal-oxide-semiconductor technology, achieving a performance improvement of multiple orders of magnitude compared to 7 nanometers process high-performance chips. At the same time, the materials used are simple and easily obtainable, and the cost is only a fraction of the latter.

Will the emergence of this cutting-edge chip technology help China achieve a "overtaking on the curve" in chip research and development? In response to this question, the research team told the Global Times that the research and development competition in the traditional chip technology field is becoming increasingly fierce, and facing international challenges such as the slowing down of Moore's Law, the world is seeking new computing architectures. It can indeed be understood as a kind of "curve" opportunity. Whether we can achieve "curve overtaking" depends on the joint efforts and ecological construction of all sectors. We are also working towards this goal.

A special review of this research, invited by the journal Nature, pointed out that the appearance of this chip may allow the new generation of computing architecture which will be integrated into daily life much earlier than expected. Academician Dai Qionghai, one of the corresponding authors of the paper, stated, "Developing a new computing architecture for the era of artificial intelligence is a summit, but truly implementing the new architecture in real life to solve major national and livelihood needs is a more important challenge and our responsibility."