Hybrid Memristor AI chip is scalable

Applications of AI

Scientists have combined atomically thin devices with conventional microchips to mimic the brain, which could help implement neural network artificial intelligence systems in a much more energy efficient way than standard electronics. A new study finds that they have created a hybrid electronic device that

As electronics get smaller and smaller, scientists are investigating atomically thin 2D materials for next-generation electronics. For example, graphene consists of a single layer of carbon atoms, while molybdenum disulfide is made of a sheet of molybdenum atoms sandwiched between two layers of sulfur atoms.

“2D materials not only have state-of-the-art electrical performance, but also excellent thermal, mechanical, optical, and chemical properties, which could lead to new applications that do not exist today. There is,” said Mario, senior author of the study. King Lanza, Associate Professor of Materials Science and Engineering at King Abdullah University of Science and Technology in Twal, Saudi Arabia, said:

“Most people specialize in semiconductors. We specialize in insulators.”
—Mario Lanza, King Abdullah University of Science and Technology

Several research teams have developed prototype devices based on 2D materials. However, none have demonstrated the ability to compute or store data. Moreover, their manufacture has largely relied on synthetic and processing methods that are incompatible with standard industrial techniques. Moreover, it is difficult to manipulate monolayer His 2D materials. This is because defects can occur when the material is transferred from the surface on which it was grown to a more useful substrate for the application. These defects reduce device consistency and yield.

Scientists now have created what they say is the first densely integrated microchip made of 2D materials, using a process compatible with the semiconductor industry. “We not only had excellent properties, but also high yields and low variability,” he says.

In a new study, researchers experimented with hexagonal boron nitride. This atomically thin ceramic is often used as an insulating material in 2D electronics. “Most people’s expertise is in semiconductors,” he says Lanza. “We are insulation specialists.”

Scientists wanted to overcome many of the challenges faced by previous devices based on 2D materials. For example, instead of trying to fabricate transistors from 2D materials, Lanza and his colleagues aimed to create memristors. A memristor or memory he register is basically a switch that can remember the switched electrical state after the power is turned off.

This hybrid 2D/CMOS microchip is promising for memristive applications.Mario Lanza

“Most groups focus on transistors, probably because they are the workhorse components of electronics,” says Lanza. “Instead, we focused on memristors, which have a much smaller current market size but great potential for data storage, computation, encryption, and communication.”

Scientists around the world want to use memristors and similar components to build electronics that can both compute and store data like neurons. These memristive devices have the potential to significantly reduce the energy and time lost when conventional microchips transfer data between processors and memory. Such brain-inspired neuromorphic hardware may also prove ideal for implementing neural networks. These AI systems are increasingly being used in applications such as supporting self-driving cars and analyzing medical scans.

A memristor is a “fault-tolerant simple device,” says Lanza. In contrast, transistors “need perfect crystalline materials,” he explains. Lanza says memristors also don’t suffer from other transistor problems, such as contact resistance (electrical resistance at points of contact with other components).

Moreover, while most previous studies have relied on 2D materials with a thickness of one or two layers, Lanza and his colleagues have demonstrated a total thickness of about 6 nanometers from about 18 layers. I used a sheet of 2D material that looks like this: “This thick material doesn’t split easily,” says Lanza.

Furthermore, instead of building 2D devices on blank substrates such as traditional silica-silicon wafers, researchers fabricated 2D devices on standard CMOS microchips. Microchips helped control the current and switching of memristors, leading to the success of 2D devices.

Researchers who manufacture transistors for computation usually use the so-called front-end-of-line step. In contrast, Lanza and his colleagues built memristors on top of the back-end-of-line interconnect that links devices on the wafer. Memristors are typically built into microchips this way, “except that he used a 2D material rather than some other material,” Lanza said.

The research team transferred multilayer sheets of hexagonal boron nitride to the back-end-of-line interconnects of a 4-square-centimeter silicon microchip containing 180-nanometer-node CMOS transistors on a 200-millimeter silicon wafer. They then created circuits from this combination by etching hexagonal boron nitride and patterning and depositing electrodes onto it. Each of these circuits consisted of his 5-by-5 ​​crossbar array of cells, each consisting of his one transistor and his one memristor.

Most devices made using 2D materials are over 1 square micrometer in size, while the memristors in the new study are only 0.053 μm.2, the researchers point out. These memristors “can be miniaturized very easily if more advanced microchips are available,” he says Lanza.

A CMOS transistor helped control the current across the 2D memristor. This allowed us to achieve a memristor endurance of about 5 million switching cycles. This is roughly equivalent to existing resistive RAM and phase change memory. Without CMOS transistors, memristors could only last about 100 cycles.

Researchers have shown that they can build “or” and “implicit” logic gates to perform in-memory computing operations on the device. They note that by changing the interconnections between devices, more advanced operations can be performed.

Furthermore, scientists note that the electrical conductivity of hybrid microchips can be dynamically tuned to different levels by applying electrical pulses. This is a property called spike timing dependent plasticity. This feature suggests that the device could be useful in implementing spiking his neural networks, which mimic the human brain more closely than traditional neural networks.

The main component of the “spike” in a spiking neural network (that is, producing an output signal) only occurs after receiving a certain amount of input signal within a certain amount of time. Because spike neural networks rarely produce spikes, they shuffle much less data than typical artificial neural networks, and in principle require much less power and communication bandwidth. Conventional electronics are not well-suited to running spiking neural networks, so the market will have to develop new neuromorphic hardware to run them, scientists say.

As a proof of principle, the researchers created a spike neural network using a device with 784 input neurons, an excitation layer of 400 neurons, and an inhibition layer of 400 neurons. When tested on the standard task of classifying images from the Correction of Handwritten Digits National Institute of Standards and Technology (MNIST) database, this simple device still achieved an accuracy of about 90%.

Scientists point out that their devices require about 1.4 to 5 volts to switch. This is low compared to other prototypes in the 2D materials area that may require 20 V or more. 180 nm CMOS node. However, they suggest that this voltage may not hinder the development of this technology. For example, state-of-the-art 3D-NAND flash memory is programmed at around 20V. Also, all bipolar CMOS microchips for automotive applications require up to 40 V.

Earlier, IBM researchers experimented with the benefits of placing 2D materials on microchips. In 2011 they made a circuit containing his 1 graphene transistor and his 2 inductors, and in 2014 a larger circuit containing 3 graphene transistors, 4 inductors, 3 capacitors, and 2 resistors. developed the circuit. However, IBM seems to have abandoned this approach. “Probably because it’s hard to transfer his 2D materials in a single layer,” he says. In contrast, Lanza and his colleagues used his 18-ply thick material, which is more durable. He predicts that “more discoveries will come as many other scientists now prototype on working microchips instead of his non-functioning SiO2 substrates.” increase.

Lanza also says that 2D materials are typically the domain of materials scientists rather than microchip engineers. “To do the experiments that we did, we used a specific software to design a microchip and then multi-project he created a tape-out of his wafer or, as in our case, the entire wafer. should be created,” he says. “If you use CMOS technology at the 180 nm node, as in our case, the first costs $25,000 and the second costs $100,000. I can’t even afford this.In our case, a colleague from Tsinghua University provided the wafers and I integrated the materials.”

Lanza said their research has already attracted the interest of major semiconductor companies.Scientists are now aiming for over 4cm2 Silicon microchips are “for making whole 300 mm wafers,” Lanza says.

Scientists detailed their findings in a journal on March 27 Nature.

Updated April 6, 2023

from an article on your site

Related articles on the web

Source link

Leave a Reply

Your email address will not be published. Required fields are marked *