Bound the bank size
Wire IR drop compresses the signal as N grows, so the largest usable bank depends on the device and the metal.
Review the signal path →Interactive crossbar map
Size each bank for the SNR its compensation can recover.
Every point comes from the full network model with wire resistance, DAC and device nonlinearity, and 4% bitcell variation. Accuracy uses ResNet-20 noise tolerance measured on NeuRRAM, scaled to the 84.94% ISCAS baseline.
Model check
The model reproduces measured crossbar chips.
Each chip keeps its published circuit parameters. One unknown is fitted to one reported result, and a second result checks the prediction.
| Chip | Result | Role | Reported | Model |
|---|
Fits come from scripts/fit_crossbar_chips.py. The Tsinghua ReRAM chips are left out because their conductance and programming data are not openly available.
Crossbar scaling rule
Keep analog optimization local.
Each bank keeps its local Rs*, and the converted partial sums are combined digitally.
Remove amplitude error
Pulse-width inputs avoid DAC nonlinearity, which no calibration removes.
Review the SNR optimum →Accumulate digitally
Partition larger dot products across banks and combine their converted partial sums outside the array.
Review the network evidence →