How far away could artificial brains be?
It may be a very long time before we understand exactly how our brains function, but there are some promising signs that we’re close.
Electrochemical RAM (ECRAM)
That’s why a device that draws its power from batteries now seems surprisingly well-suitable for running artificial neural networks.
ECRAM is giving traditional transistor-based AI an unexpected run for its money. It’s quickly becoming one of the leading contenders in the race to create the perfect artificial synapse.
IEEE International Electron Device Meeting
Researchers recently reported several new developments at this week’s IEEE International Electron Device Meetings (IEDM 2022), including ECRAM chips that use less power, store more data, and require less space.
Analog AI
Artificial neural networks are programs that model a large number of electronic neurons, along with their many connections.
Instead of using software to represent artificial intelligence, some scientists believe that faster, more efficient AI could be achieved by building actual hardware for the brain’s neurons.
This concept, called “analog artificial intelligence,” requires a type of computer chip that has a number of challenging requirements: It must store a wide variety of digital numbers, change from one number to another reliably and quickly, retains its value for a long period of time, and be easy to manufacture at scale.
The State of Charge of a Battery
Back in 2015, a team of scientists from Sandia National Labs realized that the solution to storing digital data in an analog way was right in front of us: the state of charging a battery.
A rechargeable lithium-ion (Li-Ion) cell uses an electrochemical reaction to store electrical charge. When chemical reactions occur inside the cell, they move positive and negative charges from one material to another.
“We found that we can store similar types of data using the same process.”
How Does ECRAM Work?
Basically, as many electrons as there are in the battery determines an analog voltage. The ECRAM uses these ideas by using a third gate terminal to control how much charge is in its battery.
Recent advances are rapidly bringing ECRAMS closer to having all the characteristics required for an ideal analog RAM.
Lower Energy
Ionized particles (ions) don’t shrink down to anything smaller than a single proton. Del Alamo’s team at MIT has chosen ions as their information carriers because they’re so fast.
A few years ago, they showed devices that moved ions around in mere nanoseconds, about 10,000 times faster than synapses in the brain, but fast wasn’t good enough.
“We can see the device responds very quickly to small voltage spikes,” he says, “but we need to improve its response time when the voltage is larger.”
Bottleneck
Researchers from MIT published new findings today at the International Electron Devices Meeting (IEDM) showing how they’ve built an ultrafast, low-power transistor using graphene.
They found that the protons travel easily through the electrolyte layer but need an extra voltage boost at the boundary between the electrolyte and channel.
Researchers believe they can engineer the materials used to create transistors so that they require less power to switch between states. This could lead to greater energy efficiency and scalability.
Longer Memory
Researchers from Sandia National Labs and the University of Michigan found a way to extend battery life by ten times. Their findings were recently reported in the journal Advanced Electronic Material.
To keep their memories intact for so long, the team, led by Yiyang Li at MIT, used an even heavier oxygen atom than the one in the conventional MIT device. However, what they saw was surprising.
They speculate that the lack of volatility comes from their choice of materials, tungsten oxide (WO3), and the arrangement of oxygen atoms within them.
However, this long memory property has come at the cost of switching speeds, which are in the minutes for Li’s device. Nevertheless, the researchers say they were able to identify other materials that showed both long memories and fast-switching properties simultaneously.
Tinier Footprint
Adding a third terminal to these devices makes them bulkier than competing two terminals, which limits their scalability.
To help compress the devices and pack them into an array, researchers from the Pohang University of Sciences and Technology, in South Korea, played them on their sides. They found that when they did so, the density increased by up to 30 percent.
He was able to shrink the size of the transistors down to a mere 30 nanometers, which is about 1/5th the size of previous generations while maintaining their performance.
They also presented their results at IEEE IEDM 2022.
