The human brain is considered the most powerful processor in the world due to its unique structure of neurons and synapses. Mimicking this simulation setup could make a more powerful computer that saves time and energy by performing operations in series and processing data in memory rather than transferring data between different components. Neural networks exploit this principle, but they have their own hardware limitations.
Now, an MIT team may have cracked one of those limitations. Researchers have developed a new type of programmable resistors that are the building blocks of analog processors. The electrical conductivity of these devices can be switched to conduct or block ions as needed, and arrays of these resistors can process and transmit information like natural neurons and synapses.
In this case, the team made some improvements to the formula. First of all, the resistors conduct protons, the smallest ions that can move extremely fast with a little help. But the main change is the solid electrolyte, which is made of phosphosilicate glass (PSG) — essentially silicon dioxide with a little phosphorus added. The inorganic material was found to have high proton conductivity at room temperature, thanks to its nanoscale pores, which allow protons to pass through while blocking electrons.
When a strong electric field of up to 10 volts is applied, the protons travel through the device stack at lightning speed. This enables analog processors to transfer data a million times faster than previous versions — including the human brain’s synapses.
Importantly, even with all the energy going through it, the resistor doesn’t collapse over millions of cycles because the smaller size and mass of the protons means they don’t damage the material. And, because PSG is an insulator against electrons, very little current flows through the device, keeping it cool and reducing energy usage.
The researchers plan to tweak the design so that these resistors can be fabricated in large numbers to produce arrays of them to see how they work together. This could eventually lead to faster computer types.
The study was published inscience” magazine.
