Crossbar IMCMaximizing compute SNRRepository

Accuracy evidence

The SNR-selected point remains within one accuracy point.

The device-specific circuit coordinates are compared with an exhaustive parameter search after mapping ResNet-20 across the evaluated MRAM, ReRAM, and FeFET crossbars.

See the resultOpen result data ↗
<1 pp

Accuracy gap from the exhaustive circuit search

Across the evaluated MRAM, ReRAM, and FeFET mappings, ResNet-20 accuracy at the SNR-selected parameters remains within one percentage point of the empirically found maximum.

Exact ResNet-20 / CIFAR-10 results reported in the paper
DeviceExhaustive maximumSNR-selectedGap
MRAM83.55%83.09%0.46 pp
ReRAM84.52%84.11%0.41 pp
FeFET84.30%83.53%0.77 pp

Fixed-point baseline: 84.94%. The mapped experiment used signed 5-bit inputs, ternary weights, and no retraining.

Network mapping

ResNet-20 maps onto three crossbar sizes.

The three residual blocks implement dot products of dimension N = 144, 288, and 576, and each block maps to a crossbar with its own sensing resistance. Rs1 is held at its optimum while (Rs2, Rs3) are swept. The SNR analysis picks its point without simulating the network.

ResNet-20 layer map3 crossbar arrays
  • 1
    Block 1 · 16 channelsN = 144 · 288 physical columns · Rs1 fixed at Rs1*
  • 2
    Block 2 · 32 channelsN = 288 · 576 physical columns · Rs2 swept
  • 3
    Block 3 · 64 channelsN = 576 · 1,152 physical columns · Rs3 swept

Accuracy surfaces

The SNR-selected point lands beside the exhaustive maximum.

Blue marks the exhaustive maximum. Red marks the SNR-selected point.

MRAM ResNet-20 accuracy surface over the block 2 and block 3 sensing resistances, with the exhaustive maximum and the SNR-selected point close together near the peak.
MRAM. Rs1 = Rs1* = 464 Ω. Adapted from Fig. 4(a) of the ISCAS 2022 paper.

MRAM keeps 83.09% at the SNR-selected point.

Exhaustive maximum
83.55% at (372 Ω, 177 Ω)
SNR-selected
83.09% at (464 Ω, 268 Ω)
Gap
0.46 pp

Evidence chain

Device-specific sensing coordinates survive three checks.

Maximizing array-level compute SNR also maximizes network accuracy, so each bank’s sensing resistance can come from the model instead of a simulation-based search.

1

Analytical selection

The model identifies the sensing resistance that maximizes compute SNR for each evaluated device and N.

2

Circuit agreement

In the paper’s specific ReRAM validation case, the behavioral SNR trend follows the precomputed SPICE result.

3

Network transfer

The selected device-specific points closely match the best mapped ResNet-20 accuracies.

The architect applies these sensing points to a workload partitioned across crossbar banks.

Open the architect