Memory and AI devices

# Resistive Memory # Selector # Crossbar Array # AI application

Machine learning including deep learning is achieved by huge amount of calculation. It needs much lower power consumption and calculation hardware than by software.

We are research for Resistive Memory, Selector and Crossbar Array.

Resistive Memory

Resistive memory(ReRAM) is one of the non-volatile memories.
ReRAMs store data changing its resistance by interface barrier difference or forming/rupturing conductive filament
They change into states with different resistance with low power consumption and fast switching speed.

Selector

The selector is the element in the crossbar array for individual access of memory. Since the crossbar array has many parallel paths, parallel resistance occurs. If current and voltage are applied through these parallel resistors during the read / write process of the memory, a leak current will occur and the correct read / write operation will not be performed, which will increase power. The selector is a device that has a nonlinear current-voltage characteristic, and blocks leakage current flowing at a low voltage to ensure that the current and voltage flow through the individual memory elements.

Crossbar Array

The crossbar array is a stacked structure of two-terminal non-volatile memory. Resistive random access memory(ReRAM) and phade change memory(PCM) are being studied for memory used in crossbar arrays. These next-generation memories have excellent performance such as non-volatile, low power, high speed I/O, and high integration.

Intel has released Optain memory to launch crossbar array based products.

To implement a crossbar memory, it is necessary to develop a memory element that stores data and a selector to access the memory.

AI application

2T-0C DRAM

Machine learning including deep learning is achieved by huge amount of calculation using GPU and DRAM. Crossbar memory composing ReRAMs and selectors can be replaced these devices. A machine learning system implemented by hardware needs much lower power consumption and calculation time than by software.

Page updated

Google Sites

Report abuse