The ambition of two young Chinese chip scientists: new materials to achieve the integration of storage and computing to break the bottleneck of AI computing power

  Recently, a research team from the Federal Institute of Technology in Lausanne, Switzerland published a paper titled “Construction of Storage and Computing Units from Atomic Thick Semiconductor Materials” on Nature. The research results effectively integrate the two functional modules of logical operation and data storage through a single architecture, which may pave the way for the emergence of more efficient computers. It is worth noting that this technology is particularly suitable for artificial intelligence computing. Doctoral students Zhao Yanfei and Wang Zhenyu from China also participated in the writing of this thesis.
  Corresponding author Andras Kis of the paper and one of the authors of the paper, Yanfei Zhao, said that this research was initiated and directed by Professor Andras Kis of the Nanoelectronics and Structure Laboratory (LANEs) of the Federal Institute of Technology Lausanne (EPFL) in Switzerland. At the same time, Andras Kis Also the corresponding author of the paper, doctoral student Guillherme Migliato Marega, with the cooperation of Zhao Yanfei and others, completed the preparation of the above-mentioned new computing and storage two-in-one chip.
  Regarding computing and storage, the current popular practice in the industry is to shorten the communication “path” between the storage unit and the computing unit as much as possible. Take the A64FX, which is currently the number one Japanese supercomputer “Fuyue”, as an example. Its chip adopts a general architecture of fusion CPU+GPU, and has a built-in 7nm HBM2 memory. The memory bandwidth of each chip is up to 1024GB/s. .
  But the problem is still not solved fundamentally. These storage units and computing units are still separated. Is it possible to “combine the two into one”?
  The computer we currently use usually processes data in the CPU, and then transfers the data to the hard disk or solid-state hard disk for storage.
  This model has been running for decades, but there are obviously more efficient ways, such as the human brain. It is known as the most powerful computer in the world. The neurons in the brain can process and store information at the same time.

Logic in memory

  Based on this, Professor Andras Kis tried to imitate the human brain to develop a chip that combines a storage unit and a computing unit. After the idea was determined, the team used molybdenum disulfide (MoS2) as the channel material and used it to develop logic devices and circuits in memory based on floating gate field effect transistors (FGFETs). After demonstrating programmable NOR gates, FGFETs, as a building block suitable for reconfigurable logic circuits, can be applied to more complex programmable logic.
  According to Zhao Yanfei, in this research, doctoral student Guillherme Migliato Marega built a test device and independently completed electrical measurements. Zhao Yanfei used organic metal chemical vapor deposition (MOCVD) to prepare a single layer of single crystal MoS2 material.
  Under the guidance of Professor Aleksandra Radenovic, PhD student Wang Zhenyu, who is also the corresponding author, conducted Raman spectroscopy and wafer-level monolayer MoS2 thin film growth; postdoctoral fellow Mukesh Tripathi performed high-resolution transmission electron microscope (HRTEM) measurement and simulation ;Guillherme Migliato Marega, postdoctoral fellow Ahmet Avsar and Professor Andras Kis are jointly responsible for data analysis and summary, and with the assistance of all authors, the manuscript of the paper is written. The paper can be roughly divided into four main points: brain-like computer structure, gate circuit, and Programming design, semiconductor material selection, and sample production in the laboratory.

Schematic diagram of Fuyue A64FX structure

  It is reported that the chip is made of molybdenum disulfide, a compound composed of molybdenum and sulfur, which has great potential in the manufacture of very small transistors, light-emitting diodes (LEDs) and solar cells.
  Andras Kis said that in a Mos2 wafer with a thickness of 0.65 nanometers, electrons can move as efficiently as in a silicon wafer with a thickness of 2 nanometers, and its energy consumption can be reduced by 100,000 times compared with the latter.
  Mos2 material has very good electrical characteristics, and can accurately and continuously control the conductivity of the transistors it builds, which is unmatched by graphene. In addition, molybdenum disulfide is very sensitive to the charge stored in floating gate field effect transistors, so researchers can develop circuits that can be used as both memory storage cells and programmable transistors. In addition, molybdenum disulfide can integrate multiple processing functions into a single circuit and change it as needed.
  The chip is built on the basis of floating gate field effect transistors (FGFETs). These transistors can store charge for a long time, that is, “non-volatile storage” can be realized. In the future, it can be widely used in flash memory systems of cameras, smartphones and computers .
  The floating gate memory structure involved in this research is shown in the figure below. It mainly contains a local chromium/palladium bottom gate and a thin-film platinum floating gate. This composition gives it a continuous and smooth surface, which in turn brings metal The surface roughness is reduced, and the dielectric disorder at the junction of the top tunnel oxide layer and the 2D channel can also be reduced by this, and the final device performance and stability can be improved.
  As shown in the figure below, all the device components in this study are assembled in an expandable manner, so no peeling materials (peeling: a phenomenon produced by physical and chemical actions) are used.
  Regarding the benefits of putting logic operations and data storage in the same architecture, Andras Kis said: “Our circuit design has multiple advantages. It can reduce the energy loss associated with transferring data between the memory unit and the processor, and reduce computing operations. It takes time and reduces the space required. This opens the door to smaller, more powerful and more energy-efficient equipment.”
  The breakthrough day of the integration of storage and calculation is also the breakthrough day of AI computing power bottleneck. In order to solve the above-mentioned problems, companies such as Microsoft and Intel have invested in this technology, but there is still no product that can be applied in a wide range.
  The new method proposed by the Swiss Federal Institute of Technology Lausanne this time is expected to solve the above-mentioned problems. Professor Andras Kis, the leader of the study, said: “The ability of this circuit to perform two functions is similar to the way the human brain works, (just like) neurons (like) both storing memory and performing mental calculations.”

  Zhao Yanfei reported that the team has made a lot of achievements in the field of two-dimensional material semiconductors. They designed the first transistor chip composed of a single-layer MoS2 with a thickness of only 0.7nm in 2010, and subsequently published other important results based on single-layer transition metal sulfide (TMDC), including the release in 2013 The first flash memory device based on MoS2.
  The first batch of electronic chips prepared by the team were all based on hand-stripped Mos2 materials, and the success rate was relatively low. Under the knowledge system and experimental conditions at the time, researchers did not know how to enhance chip reliability. Since then, the team discovered that with current knowledge, a circuit combining flash memory and transistors can be made, and all tasks such as memory and calculations can be completed with flash memory alone.
Post-90s “chip scholar” from Huanggang, Hubei

  Zhao Yanfei, 27 years old, was born in Huanggang, Hubei. When he was young, he studied and settled in Wuhan with his parents. In 2011, she was admitted to Zhejiang University, studied at the School of Materials Science and Engineering, and received a bachelor’s degree in engineering.

Memory device architecture

  In the second half of the undergraduate course, teachers from Zhejiang University helped her obtain many opportunities to participate in laboratory projects. These experiences also helped her apply for the opportunity to study in EPFL. After coming to EPFL, she continued her studies in materials subject and obtained a master of science degree in 2018. While taking courses, she joined three laboratories one after another, during which she also practiced in the R&D department of ABB in Switzerland for half a year.
  She said that these different experiences gave her the opportunity to explore her own scientific research interests. During her master’s thesis, she became very curious about the two-dimensional materials and devices field done by her current tutor Professor Andras Kis, so she stayed The group finally decided to study for her doctorate with the topic of “improving the electrical properties of two-dimensional materials”.
  The first author of the thesis Guillherme Migliato Marega and Zhao Yanfei had similar experiences. He once said that he was born in 1994 in San Carlos, a rural university city in Sao Paulo, Brazil. After graduating from high school, he entered the University of Sao Paulo (USP) to study electronic engineering. Professional, and participated in the dual degree program jointly opened by USP and Lyon Central Institute of Technology, and finally obtained the engineer diploma. During this period, he participated extensively in various small-scale research projects. A year before starting his Ph.D. career, he had a conversation about two-dimensional materials with a professor in Brazil. This conversation also became the enlightenment for his follow-up research. Since then, he has been passionate about the electronic applications that such materials may bring, which prompted him to contact Professor Andras Kis, who pioneered TMDC in this field, and finally decided to “Develop the most “Advanced Memory Processor” as the subject to pursue his Ph.D.
  The two post-90s scholars from different countries stated in this paper: “This direct integration of memory and logic can increase processing speed and realize machine learning, Internet of Things and non-volatile memory for energy-saving circuits based on 2D materials. Sexual computing has opened the way.”
  Talking about application scenarios, Zhao Yanfei said that in theory, this method can replace the processor in a mobile phone or computer. Even if the above goals cannot be achieved in the end, it can also be used for signal processing, such as image or sound processing on a chip.
  Finally, she added that the research is still in its early stages, and only small-scale circuits have been implemented this time. The question to be solved in the future is whether these circuits can be manufactured on a large scale and whether they will bring economic benefits.