Recording of a practice run I did

1

Predictive Modeling of Resistive Memory Device Retention

Ari Shtein, Noah Geisler, Dongjae Shin, Yiyang Li

2

How ReRAM Works

3

Setting

Resetting

HfO₂

HfO₂

Low resistance state

High resistance state

TiO_x

TiO_x

TiN

TiN

TiN

TiN

Non-volatility is important

4

ReRAM is faster and more space-efficient than flash, and can be less volatile & less energy-intensive than DRAM

​

Industry standards for non-volatile memory:

• 10 years
• 85° C

​

My rule-of-thumb

• ~10 hrs at ~250° C means ~10 years at ~100° C

Retention failure

5

HfO₂

TiO_x

TiN

TiN

My work

6

• Can we identify the features that make devices more or less likely to fail?
• Can we determine in advance whether a device is non-volatile?

​

Yes, much of the time! (on a very small dataset)

Predictions with one feature

(standard deviation of set voltage)

Feature extraction from switching curve

Retention plots for all devices in the dataset

Correctly predicted, non-volatile

Correctly predicted, volatile

Incorrectly predicted, volatile