2022IEEE ISCAS
Fundamental Limits on the Computational Accuracy of Resistive Crossbar-based In-memory Architectures
Saion K. Roy, Ameya Patil, and Naresh R. Shanbhag · Pages 384–388 · DOI 10.1109/ISCAS48785.2022.9937336
IEEE ISCAS 2022
The paper behind the crossbar accuracy map.
The publication derives the compute-SNR limits, verifies the behavioral route with circuit simulation, and connects the selected circuit parameters to mapped neural-network accuracy.
Summary
What the paper shows.
The paper derives the limits on compute SNR for MRAM-, ReRAM-, and FeFET-based crossbars. SNRmax occurs at the Rs* that balances ADC clipping and quantization, contrast helps only up to 12 to 15, and the SNR-optimal parameters also maximize ResNet-20 accuracy.
- Analytical framework. Signal and noise expressions for voltage-driven crossbars.
- Circuit validation. Behavioral models checked against 22 nm SPICE simulation.
- Design limits. SNRmax versus Rs, N, ADC precision, and contrast.
- Network transfer. SNR-selected points within one point of an exhaustive search.
Citation
Cite this work.
The repository also includes a CITATION.cff file.
IEEE
S. K. Roy, A. Patil, and N. R. Shanbhag, “Fundamental Limits on the Computational Accuracy of Resistive Crossbar-based In-memory Architectures,” in 2022 IEEE International Symposium on Circuits and Systems (ISCAS), 2022, pp. 384–388, doi: 10.1109/ISCAS48785.2022.9937336.
BibTeX
@inproceedings{roy2022crossbarlimits,
author = {Saion K. Roy and Ameya Patil and Naresh R. Shanbhag},
title = {Fundamental Limits on the Computational Accuracy of Resistive Crossbar-based In-memory Architectures},
booktitle = {2022 IEEE International Symposium on Circuits and Systems (ISCAS)},
pages = {384--388},
year = {2022},
doi = {10.1109/ISCAS48785.2022.9937336}
}Companion repository
The behavioral route is open and reproducible.
The repository turns the original scripts into a reproducible behavioral model and serves this website.
- Open model. A voltage-driven 1T1R crossbar with the four compute-SNR error terms.
- SPICE points. Published comparison points only. The commercial 22 nm PDK is not redistributed.
- Network results. Reported values in
data/network/paper_points.csv. - Parasitic extension. Wire IR drop, input-rail sag, and DAC and device nonlinearity, fitted to IBM, NeuRRAM, and HP crossbar chips.