Is the future direction of AI computing “water chips”?

The future of neural network computing may be worse than we expected - not solid chips that use electricity, but soaked in water.

Recently, a team composed of the Harvard School of Engineering and Applied Sciences (SEAS) and the startup DNA Script successfully developed a processor based on the movement of ions in an aqueous solution.

Physicists believe that such devices could be the next step in brain-inspired computing because they are closer to the way the brain transmits information.

"Ionic circuits in aqueous solutions use ions as charge carriers for signal processing," the researchers said in the paper. "We propose a water-based ion circuit... This functional ion circuit capable of analog calculations is a step towards more complex water-based ionics."

The research was published in the latest issue of the materials science journal Advanced Materials.

##Paper: https://onlinelibrary.wiley.com/doi/epdf/10.1002/adma.202205096

We know that chips in everything from smartphones to cloud servers handle computing tasks by manipulating electrons through solid semiconductors, which is different from the way biology works.

A major part of signal transmission in the brain is the movement of electrically charged molecules called ions in a liquid medium. Although the brain's incredible processing power is difficult to replicate artificially, scientists think computers could use a similar system: carrying ions in a water solution.

This approach will be slower than traditional silicon-based computing because of the changed medium, but it could have some interesting advantages. For example, ions can be produced from a variety of molecules, each with different properties that can be exploited in different ways.

But first, scientists need to show that it actually works.

A team led by Harvard University physicist Woo-Bin Jung has been working in this direction. The first step in building a computer is to design a functional ion transistor, a device that switches or enhances a signal. Their latest advance involves combining hundreds of transistors into an ionic circuit.

The transistor consists of a "bullseye" arrangement of electrodes, with a small disc-shaped electrode in the center and two concentric ring-shaped electrodes surrounding it. This comes into contact with an aqueous solution of quinone molecules. When used, a voltage applied to the central disk generates a hydrogen ion current in the quinone solution. At the same time, two ring electrodes adjust the pH of the solution, thereby increasing or decreasing the ionic current.

The chip (left) has an array of hundreds of transistors (right) in the center (middle).

Quinones are a class of organic compounds containing conjugated cyclohexadienedione or cyclohexadienedimethylene structures. Based on this substance The transistor performs the physical multiplication of the weight parameter and the disk voltage set by the ring pair gating, producing the answer to the ionic current.

You may know the concept of "biological computer", which refers to the use of biological Materials to replace the semiconductor chips and storage media currently used in computers are considered to be another major direction for the future of computers in addition to quantum computing. But much previous research has focused on individual ion diodes and transistors, rather than circuits containing many such devices.

Current neural networks that require extremely high computing power rely heavily on matrix multiplication operations, which involve multiple multiplications. So the team designed a 16-by-16 array of transistors, each capable of multiplication, to produce an ionic circuit that could perform matrix multiplication. They are implemented on the surface of and operated by complementary metal-oxide semiconductor (CMOS) electronic chips.

The researchers demonstrated the utility of this array-scale ion circuit by performing physical or simulated multiply-accumulate (MAC) operations. Analog MAC operations based on physical phenomena - In contrast to digital MAC operations based on many digital logic gates and Boolean algebra, new methods bring direction towards reducing the power consumption of artificial neural networks.

Schematic diagram of an ion transistor.

Since each crosspoint conductance acts as a network synaptic weight, the input voltage fed into the array rows is multiplied by the weight by Ohm's law and according to the basis Erhoff's law accumulates the resulting current in each column. Therefore, each column current is physically produced as a dot product between the input data vector and the column's synaptic weight vector.

In each ion transistor, the current Iout of the applied voltage Vin is gated by Ig, we can find a region of Vin where Iout = W × Vin, the proportionality constant or weight W can be adjusted by Ig, i.e. the region in which the ion transistor physically multiplies between the weight and the input voltage.

Multiply and accumulate operations.

"Matrix multiplication is the most commonly used calculation in artificial intelligence neural networks, and our ionic circuit performs matrix multiplication in water in a completely electrochemical-mechanical simulation. ”, said Woo-Bin Jung.

Of course, this technology currently has significant limitations, including that operations must be performed sequentially rather than simultaneously, which greatly slows down the method.

However, the research team believes that the next step is not to increase the speed, but to introduce a wider range of molecules into the system. So far, the team has used only three or four ionic species, such as hydrogen and quinone ions, to achieve gating and ion transport in aqueous ion transistors. This research attempts to complete more complex ion calculations and let the circuits process more complex information.

The research team pointed out: The ultimate goal of this research is not to use ion technology to compete with or replace electronic products, but to use hybrid technology to complement each other's strengths.

The above is the detailed content of Is the future direction of AI computing “water chips”?. For more information, please follow other related articles on the PHP Chinese website!