Engineers have developed a chip for accelerated training of neural networks

Engineers from the University of Illinois at Urbana-Champaign integrated electrochemical random access memory

(ECRAM) with silicon semiconductors. The technology will speed up deep learning of neural networks and reduce the cost of operating artificial intelligence systems.

Researchers used to create memorymaterials compatible with modern semiconductor technologies (CMOS): tungsten oxide for the gate and channel, zirconium oxide for the electrolyte, and protons as mobile ions. This made it possible to integrate the device into standard microelectronics.

ECRAM manufacturing process. Image: Jinsong Cui et al., Nature Electronics

ECRAM is a memory cell or device thatuses the same space for storing data and computing. This custom architecture eliminates the energy cost of transferring data between memory and the processor, enabling data-intensive operations to be performed very quickly and efficiently.

Electrochemical memory encodes informationby moving mobile ions between the gate and the channel. Electrical pulses applied to the gate terminal either introduce ions into the channel or draw them out, the resulting change in the channel's electrical conductivity retains information. It is read by measuring the electrical current flowing through the channel. The electrolyte between the gate and channel prevents unwanted ion flow, allowing the memory to operate in a non-volatile mode.


Electrochemical memory circuit. Images: Jinsong Cui et al., Nature Electronics

Researchers have shown that thethe device exhibited high switching speeds, withstood over 100 million read-write cycles, and was much more efficient than standard memory technology. At the same time, the channel reliably holds ions for hours, which is enough to train most deep neural networks. Because the materials are compatible with microfabrication technologies, devices can be scaled down to micro and nano sizes, enabling high density and processing power.

Read more:

The key theory of quantum physics has finally been proven. Main

Biologists discover how cancer cells elude the immune system

Found a way to lower blood sugar without insulin injections

Cover: ECRAM array. Image: The Grainger College of Engineering at the University of Illinois Urbana-Champaign