RADHARD 2023

Radiation Hardness Assurance

Abstract

(c) Puchner Helmut
(c) Puchner Helmut

Radiation Hard AI Memory for Space Applications

H. Puchner, V. Agrawal, M. Iskarus1

1 Infineon Technologies, San Jose, CA, USA

 

Abstract

We will present the fundamental capabilities of our SONOS charge trapping memory with respect to radiation tolerance and discretization capability which allows SONOS to be the only true radiation hard analog memory. We have demonstrated up to 64 levels of discretization without significant read disturb and excellent data retention. Several neural network model implementations will be presented to showcase the capability of the analog memory AI implementation for edge inference. In memory compute performance of up to 50 TOPS/W is achievable with this technology. SONOS is currently the only matured solution in a radiation and power limited space for AIM (AI Memory) implementation.

References

[1] M. Marinella et al., “Achieving Accurate In-Memory Neural Network Inference with Highly Overlapping Nonvolatile Memory State Distributions”, 2022 6th IEEE Electron Devices Technology & Manufacturing Conference (EDTM)
[2] P. Xiao et al., “Ionizing Radiation Effects in SONOS-Based Neuromorphic Inference Accelerators”, IEEE Transactions on Nuclear Science 2021 Vol. 68(5)
[3] H. Puchner et al., ”Impact of Total Ionizing Dose on the Data Retention of a 65 nm SONOS-Based NOR Flash”, IEEE Transactions on Nuclear Science 2014 Vol. 61(6)

 

    

Acknowledgments

This work has been partially conducted in cooperation with Sandia National Laboratories, Albuquerque, NM, USA and we respectfully acknowledge their contribution.