An optimum sensing resistance
Compute SNR peaks where ADC clipping and quantization noise are equal.
ISCAS 2022 · Resistive crossbars
Balance the error coordinates to maximize compute SNR.
Statistical signal and noise models identify the sensing resistance that balances ADC clipping and quantization for MRAM-, ReRAM-, and FeFET-based crossbars.
Why a limit matters
Resistive IMCs give up accuracy to gain energy efficiency.
Embedded non-volatile memories such as MRAM, ReRAM, and FeFET make in-memory computing dense and non-volatile. However, a resistive crossbar computes in the analog domain, and empirical or simulation-based methods cannot pinpoint the limit on its accuracy or the non-ideality that sets it. That limit matters most when single-bank macros scale to multi-bank systems.
Key results
Four results define the crossbar design space.
Results assume signed 5 b inputs, a ±2 µA ADC range, and 4% mismatch and variation.
A ceiling on useful contrast
Beyond this Roff/Ron, mismatch and variation dominate.
A minimum ADC precision
MRAM needs 6 b, while ReRAM and FeFET need 7 b at N = 512.
A network-level match
SNR-selected ResNet-20 accuracy is within one point of an exhaustive search.
Evaluated devices
Three eNVMs span a 500-fold range of resistive contrast.
SNRmax depends on the contrast Roff/Ron, not on the absolute resistance.
| Device | Ron | Roff/Ron | BADC* at N = 512 | Trend with dimension N |
|---|---|---|---|---|
| MRAM | 3 kΩ | 2 | 6 b | Rolls off beyond N ≈ 500 |
| ReRAM | 25 kΩ | 12 | 7 b | About 5 dB above MRAM, rolls off beyond N ≈ 2,000 |
| FeFET | 1 MΩ | 1,000 | 7 b | No roll-off, since Rs* stays near 10 kΩ |
The dimension trend assumes a 6 b ADC and Rs,min = 1 kΩ.
Research route
Move from circuit structure to verified accuracy.
Model the signal path
The array, sensing resistance, TIA, ADC, and four noise sources.
Open Architecture →Balance the errors
How clipping and quantization set Rs*, and what moves the maximum.
Open SNR Optimum →Check network accuracy
The SNR-selected point against an exhaustive ResNet-20 search.
Open Accuracy Evidence →Map a workload
Partition a larger workload across banks at the paper-selected Rs*.
Open Architect →The analysis continues in parallel-bar architectures and a measured 22 nm MRAM macro.
See related publications