MIT engineers say stacking circuit components on top of each other could be a solution to creating more energy-efficient artificial intelligence (AI) chips. Logic and memory components each perform calculations and store data. These components can transfer data more easily when they are in direct contact than when they are separated.
The research team created so-called “memory transistors,” which consist of both logic elements (transistors) and memory elements that can perform calculations. The nanoscale devices have relatively few electrical defects, meaning they can operate faster with less power, the scientists said in two studies presented at the International Electronic Devices Conference in San Francisco on Dec. 9 and Dec. 10.
you may like
A single operation of ChatGPT can generate as much heat as a bottle of water for cooling. However, most of the energy associated with AI is used in passing data between components rather than performing calculations. Scientists believe that even small savings on chips can have a big impact.
“In the future, we will need to minimize the amount of energy we use for AI and other data-centric computations, because that is simply not sustainable,” study lead author Yanjie Shao, a postdoctoral fellow at MIT, said in a statement. “Continuing that progress will require new technologies like this integrated platform.”
Stacking can save energy, but it’s not easy
Modern chips contain logic circuits made of transistors. These are on/off switches that control the flow of current. Together, these transistors represent the binary numbers 1 and 0, and this is how the chip processes information. It also has a memory circuit that includes transistors along with other materials that can store data.
Logic and memory circuits are traditionally separated, and data must be moved between them via wires and interconnects, wasting energy in the process. Stacking active ingredients may seem like an obvious solution, but the challenge is to do so without causing damage. For example, some transistors cannot withstand heat, so deposition, the controlled formation of ultrathin layers that form these components, must be done at low temperatures.
To overcome this problem, scientists built logic transistors with an active channel layer (the area through which electricity flows) made of indium oxide. Importantly, the material can be deposited in 2-nanometer layers at about 302 degrees Fahrenheit (150 degrees Celsius). This is a low enough temperature that it does not affect other transistors.
Scientists stacked memory components (10 nanometer layers of ferroelectric hafnium zirconium oxide) vertically on top of indium oxide transistors, allowing the device to store and process data. The resulting memory transistor can be switched on or off in just 10 nanoseconds and operates at less than 1.8 volts. Typical ferroelectric memory transistors tend to have orders of magnitude slower switching speeds and require voltages of 3-4V.
Memory transistors are built on the “back end” of the chip, where the wires and metal bonds that connect the front-end active components are located, further increasing efficiency. Shao said that doing this allows the chip to be much more densely integrated.
In two studies, memory transistors were placed only on chip-like structures rather than functional circuits. The researchers hope to improve the transistor’s performance so that it can be integrated first into single circuits and then into larger electronic systems.
“We can now build a versatile electronics platform on the back end of the chip, enabling energy efficiency and a wide range of functionality in very small devices,” said Shao. “We have superior device architectures and materials, but we must continue to innovate to uncover the ultimate performance limits.”
Source link
