[Paper Review] SWIM: Selective Write-Verify for Computing-in-Memory Neural Accelerators

Zheyu Yan, Xiaobo Sharon Hu+1 more|arXiv (Cornell University)|Feb 17, 2022

Advanced Memory and Neural Computing4 citations

TL;DR

This paper proposes SWIM, a selective write-verify framework for computing-in-memory (CiM) neural accelerators that reduces programming time by identifying and verifying only the most sensitive weights using second derivatives. By requiring only a single forward and backpropagation pass, SWIM achieves up to 10x speedup over full write-verify while maintaining near-ideal model accuracy despite non-idealities in emerging non-volatile memory devices.

ABSTRACT

Computing-in-Memory architectures based on non-volatile emerging memories have demonstrated great potential for deep neural network (DNN) acceleration thanks to their high energy efficiency. However, these emerging devices can suffer from significant variations during the mapping process i.e., programming weights to the devices), and if left undealt with, can cause significant accuracy degradation. The non-ideality of weight mapping can be compensated by iterative programming with a write-verify scheme, i.e., reading the conductance and rewriting if necessary. In all existing works, such a practice is applied to every single weight of a DNN as it is being mapped, which requires extensive programming time. In this work, we show that it is only necessary to select a small portion of the weights for write-verify to maintain the DNN accuracy, thus achieving significant speedup. We further introduce a second derivative based technique SWIM, which only requires a single pass of forward and backpropagation, to efficiently select the weights that need write-verify. Experimental results on various DNN architectures for different datasets show that SWIM can achieve up to 10x programming speedup compared with conventional full-blown write-verify while attaining a comparable accuracy.

Motivation & Objective

• To address the high programming time of write-verify schemes in non-volatile computing-in-memory (nvCiM) DNN accelerators.
• To reduce the number of weights requiring write-verify without sacrificing model accuracy.
• To develop an efficient, single-pass method to identify the most sensitive weights for write-verify.
• To demonstrate that selective write-verify is viable and significantly faster than full write-verify.
• To provide a practical, hardware-friendly alternative to in-situ training or retraining for handling device variations in nvCiM systems.

Proposed method

• SWIM uses second-order derivatives of the DNN weights as sensitivity indicators to prioritize which weights need write-verify.
• It computes second derivatives efficiently using a single forward and backpropagation pass, avoiding expensive finite difference approximations.
• The method selects only a small subset of weights—those with high second derivative magnitude—for iterative write-verify.
• The selection process is based on the hypothesis that weights with higher second derivatives are more sensitive to programming errors.
• The framework is applied during the weight mapping phase before deployment on nvCiM hardware.
• It enables significant reduction in write-verify cycles while preserving model accuracy under device variations.

Experimental results

Research questions

• RQ1 Can we reduce the number of weights requiring write-verify in nvCiM DNN accelerators without degrading model accuracy?
• RQ2 Is there a computationally efficient way to identify the most sensitive weights for write-verify?
• RQ3 Can second derivatives serve as a reliable metric for selecting critical weights in the presence of device variations?
• RQ4 How does selective write-verify compare to full write-verify and in-situ training in terms of speed and accuracy?
• RQ5 What is the minimum number of write-verify cycles needed to maintain acceptable accuracy across diverse DNN architectures and datasets?

Key findings

• SWIM achieves up to 10x programming speedup over conventional full write-verify on MNIST, CIFAR-10, and Tiny ImageNet.
• With only 10% of the write cycles (NWC = 0.1), SWIM maintains an accuracy drop of less than 0.1% on MNIST and 0.5% on CIFAR-10, compared to full write-verify.
• On Tiny ImageNet, SWIM achieves less than 3% accuracy drop with only 10% of the write cycles, outperforming magnitude-based and random selection.
• SWIM reduces accuracy standard deviation across 3,000 Monte Carlo runs, indicating superior robustness to device variation compared to baselines.
• Compared to in-situ training, SWIM achieves comparable accuracy with up to 9x less write-verify cycles, avoiding the need for additional hardware and training iterations.
• For ResNet-18 on CIFAR-10, SWIM maintains an accuracy drop of less than 0.5% at 10% write-verify cycles, while other methods see drops exceeding 2%.