GitHub

Today's artificial intelligence runs on powerful systems that calculate quickly, but consume energy just as quickly. One reason is that data is constantly moving between memory and the processor. Researchers from Cornell have shown a completely different, very unusual way: a system that writes information electrically, but reads it through microscopic vibrations. This is not a ready AI processor yet, but an interesting example of how the computers of the future can start calculating using the physics of materials alone.

AI has a big problem with data crunching

In a classical computer, memory and computation are usually separated. The data must be retrieved from memory, transferred to the computational unit, the operation performed and the result saved. For regular tasks this works great. In the case of artificial intelligence, it becomes problematic because the models perform enormous numbers of simple operations, mainly multiplications and additions.

It is the multiplication and accumulation of results that are the basis of the work of neural networks. Each input signal is multiplied by a weight, i.e. a number describing the importance of a given connection. Then such results are added up. In large models, this operation is repeated billions of times. The more data that has to be moved between memory and the processor, the higher the power consumption and latency.

That's why scientists are looking for systems in which memory and computation are closer together. This approach is called in-memory computing. Instead of treating memory as a passive store of numbers, you can make the memory element itself help with calculations.

Electrical recording, but mechanical reading

The new arrangement with Cornell works on a rather unusual principle. Information is written electrically but read by movement. Researchers built a ferroelectric microelectromechanical system, abbreviated as FeMEMS. In its center there is a microscopic suspended beam with a layer of hafnium oxide and zirconium 20 nm thick.

Ferroelectric material has domains, i.e. microscopic areas in which the electrical polarization can be set in a specific way. Electrical impulses change the orientation of these domains, and thus program the state of the system. Then a small reading signal causes the beam to vibrate. How the beam moves reveals the recorded value.

The most important thing here is to separate writing from reading. In many ferroelectric systems, the same electrical path is responsible for writing information, storing it and later reading it. This can be extremely problematic because the reading itself may introduce distortions or affect the saved state. In the Cornell solution, the reading is done differently, by observing microscopic vibrations of the beam. Thanks to this, there is no need to intensively probe the memory with an electrical signal to check what has been stored in it.

One beam can store much more than 0 and 1

Perhaps the most interesting thing is that the system is not limited to two states, such as the classic 0 and 1. The Cornell team showed about 200 different electromechanical states that can be distinguished from each other. This is important in analog calculations, where instead of a simple yes or no, the ability to save many intermediate values is important.

In neural networks, such values correspond to weights. The more accurately you can record weight, the less risk that small errors will start to accumulate in large calculations. When a network performs thousands or millions of operations, every small error can make a difference. A large number of stable states therefore provides greater precision.

This is what sets this chip apart from many previous analog memory demonstrations. The mere fact that a device can store intermediate values is not enough. What also matters is how many such levels can be reliably distinguished, whether they can be set precisely and whether the reading does not affect the stored data.

A computer that multiplies with motion

The vibrating beam does more than just reveal the state of memory. Because the input signal and the programmed material state interact in one device, the motion of the beam becomes the physical equivalent of multiplication. If the stored state represents one number and the input signal represents another, the mechanical response of the system corresponds to their product.

While this sounds quite complicated, it's actually something that AI models do all the time. Neural networks constantly multiply input data by weights stored in memory. If such operations could be performed without constantly transferring data between memory and the processor, the computer could use less energy and spend less time simply transferring information.

The new chip is described as a potential piece of neuromorphic hardware. This does not mean that the computer should act like a brain. It is an architecture inspired by the fact that in the biological nervous system, information storage and processing are not as brutally separated as in classical computers.

Old ideas are coming back because silicon is getting harder and harder to come by

The history of computers has not always been a history of only CMOS transistors. For decades, various physical methods of recording and processing information have been experimented with. Silicon's dominance has been so successful that many alternatives have disappeared from the mainstream. CMOS turned out to be cheap, scalable, fast and perfectly suited to industrial production.

Now, however, classic scaling is getting more and more difficult. Transistors continue to improve, but they can no longer count on a simple, automatic increase in performance with each subsequent reduction in size. At the same time, AI increases the demand for calculations at a pace that forces the search for new solutions.