U-SWIM: Universal Selective Write-Verify for Computing-in-Memory Neural Accelerators

TL;DR

U-SWIM addresses the significant programming-time bottleneck in weight mapping for nvCiM neural accelerators by selectively applying write-verify to a small, high-sensitivity subset of weights. It introduces a second-derivative–based sensitivity metric, computed via a single forward-backward pass, to identify weights that most impact accuracy under device variations. Across MNIST, CIFAR-10, and Tiny ImageNet, U-SWIM achieves up to 10× programming acceleration while preserving accuracy, outperforming magnitude-based and random selection baselines and the authors' earlier SWIM approach, especially in non-uniform device scenarios. This work enables practical, energy-efficient CiM-DNN deployment on edge devices by dramatically reducing programming time without sacrificing performance, even with heterogeneous device variations.

Abstract

Architectures that incorporate Computing-in-Memory (CiM) using emerging non-volatile memory (NVM) devices have become strong contenders for deep neural network (DNN) acceleration due to their impressive energy efficiency. Yet, a significant challenge arises when using these emerging devices: they can show substantial variations during the weight-mapping process. This can severely impact DNN accuracy if not mitigated. A widely accepted remedy for imperfect weight mapping is the iterative write-verify approach, which involves verifying conductance values and adjusting devices if needed. In all existing publications, this procedure is applied to every individual device, resulting in a significant programming time overhead. In our research, we illustrate that only a small fraction of weights need this write-verify treatment for the corresponding devices and the DNN accuracy can be preserved, yielding a notable programming acceleration. Building on this, we introduce USWIM, a novel method based on the second derivative. It leverages a single iteration of forward and backpropagation to pinpoint the weights demanding write-verify. Through extensive tests on diverse DNN designs and datasets, USWIM manifests up to a 10x programming acceleration against the traditional exhaustive write-verify method, all while maintaining a similar accuracy level. Furthermore, compared to our earlier SWIM technique, USWIM excels, showing a 7x speedup when dealing with devices exhibiting non-uniform variations.

Figures (6)

• Figure 1: Diagram of the crossbar array architecture: Inputs are introduced horizontally and multiplied by weights stored in the NVM devices at each intersection. The products are accumulated vertically to produce the final output.
• Figure 2: Effect of device variation-induced weight perturbation on LeNet DNN architecture targeting MNIST dataset: (a) Relationship between accuracy drop and weight magnitude, showing minimal correlation. (b) Correlation between accuracy drop and each weight's second derivative, indicating a strong correlation.
• Figure 3: Uniform Device Variations: Comparison of Accuracy vs. Normalized Write Cycles (NWC) for U-SWIM and baseline methods across three models, ConvNet and ResNet-18 for CIFAR-10, and ResNet-18 for Tiny ImageNet. Solid lines denote mean accuracy, while shaded regions indicate standard deviation, derived from 3,000 Monte Carlo simulations utilizing the device variation model.
• Figure 4: Non-Uniform Device Variations: Comparison of Accuracy vs. Normalized Write Cycles (NWC) for U-SWIM and baseline methods across three models, ConvNet and ResNet-18 for CIFAR-10, and ResNet-18 for Tiny ImageNet—using Device $R_4$. Solid lines denote mean accuracy, while shaded regions indicate standard deviation. These metrics are based on 3,000 Monte Carlo simulations using the specified device variation model.
• Figure 5: Accuracy vs. Normalized Write Cycles (NWC) in ConvNet for CIFAR-10: Comparison between U-SWIM and baseline methods using three different devices, $F_2$, $R_4$, and $F_6$. Solid lines indicate average accuracy, while shaded regions represent the standard deviation. These results are obtained from 3,000 Monte Carlo simulations using the respective device variation models.