MIT’s chip stacking breakthrough could cut energy use in power-hungry AI processes
MIT engineers say stacking circuit components on top of each other could be the answer to creating more energy-efficient artificial intelligence (AI) chips. Logic and memory components, which perform calculations and store data respectively, can transfer data more easily when they are in direct contact rather than when they are separated.
The team created what is called a “memory transistor” comprising both a logic element capable of performing calculations (the transistor) and a memory element. This nanoscale device has relatively few electrical defects, meaning it can operate faster while consuming less electricity, the scientists said in two studies presented December 9-10 at the International Electronic Devices Meeting in San Francisco.
A single interaction with ChatGPT can generate enough heat for you to need it. the equivalent of a bottle of water to cool off. But most of the energy associated with AI is used to data transfer between components rather than performing calculations. According to scientists, even a small saving on the chip could have a huge impact.
“We need to minimize the amount of energy we use in the future for AI and other data-centric computations, because it is simply not sustainable,” lead author of the study. Yanjie Shaopostdoctoral researcher at MIT, said in a statement. “We will need new technologies like this integration platform to continue this progress.”
Stacking saves 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. These transistors combine to represent binary 1s and 0s, which is how chips process information. They also have memory circuits, containing transistors as well as other materials capable of storing data.
Logic and memory circuits are traditionally separated and data must travel between them via wires and interconnections, which wastes energy. While stacking active components may seem like an obvious solution, the challenge is achieving it without causing damage. Deposition, the controlled formation of ultrathin layers that form these components, must be done at low temperatures, for example, because some transistors cannot handle heat.
To overcome this problem, the scientists built their logic transistor with an active channel layer (the region where electricity flows) made of indium oxide. Importantly, the material can be deposited in a two-nanometer layer at around 302 degrees Fahrenheit (150 degrees Celsius). This is a low enough temperature to not affect other transistors.
Beyond the indium oxide transistor, the scientists vertically stacked a memory component — a 10-nanometer layer of ferroelectric hafnium-zirconium oxide — that allows the device to store data as well as process it. The resulting memory transistor can turn on or off in just 10 nanoseconds and operates at less than 1.8 volts. Switching speeds of typical ferroelectric memory transistors tend to be orders of magnitude lower and require voltages between 3 and 4 V.
The memory transistor is made even more efficient by being built on the “back end” of the chip, where the wires and metal bonds that connect the active components of the front end are located. Shao said this makes the chip’s integration density much higher.
For both studies, the memory transistor was installed only on a chip-like structure rather than in a functional circuit. The team hopes to improve the transistor’s performance so that it can be integrated first into a single circuit and then into larger electronic systems.
“Now we can build a versatile electronics platform on the back of a chip that allows us to achieve high energy efficiency and many different features in very small devices,” Shao said. “We have good device architecture and materials to work with, but we must continue to innovate to discover the ultimate limits of performance.”
