Number 11 of 12 · November 21, 2026

Scalable fault vulnerability analysis for ReRAM-based process-in-memory CNN accelerators

The authors propose a fast, formula-based way to estimate how much stuck-cell faults in ReRAM chips hurt CNN accuracy, replacing days or weeks of fault injection tests with minutes or hours.

What they found

The authors build an analytical framework that scores how vulnerable a CNN is to permanent faults on ReRAM-based processing-in-memory chips. These faults are cells stuck at a high or low resistance state, which corrupt the stored weights. The scores work at three levels: single weights, kernels and whole layers. The authors say the scores match accuracy loss measured with fault injection simulations, which confirms the method. They report a speedup of about 10⁴×-10⁵× over exhaustive fault injection, bringing analysis of large CNNs down to a few minutes or hours.

Why it matters

Fault injection means breaking parts of a network one by one and measuring the damage. It is accurate but very slow for large models. A quick estimate of which weights, kernels and layers matter most could help designers protect only the fragile parts. The abstract points to edge and safety-critical uses, where unreliable hardware is a real concern. The abstract gives no detail on which CNNs or datasets were tested, or on how close the match is.

How it changes the race

This is a niche hardware reliability result and incremental for frontier SI. If it holds, it could make fault-tolerant design of efficient SI chips cheaper and faster, mainly for edge devices.

Safety and control

The paper bears on the reliability of SI hardware in safety-critical settings, where stuck cells can silently degrade accuracy. It is about hardware faults, not about alignment or control of SI behavior.

The abstract, by the authors

Resistive random-access memory (ReRAM)-based processing-in-memory (PIM) accelerators offer a highly efficient platform for parallel matrix-vector multiplications (VMM) in neural networks, leveraging analog computation, high density, low leakage, and non-volatility. However, their susceptibility to permanent faults, such as stuck-at-high (SaH) and stuck-at-low (SaL) resistance states, permanently corrupts the convolutional neural networks (CNNs) weights mapped onto ReRAM cell conductance values, thereby degrading inference accuracy and leading to system unreliability in safety-critical applications. Fault injection is a commonly used technique to assess the vulnerability of neural networks by systematically injecting faults into individual parameters and measuring the resulting accuracy degradation. Although this method is accurate, it is prohibitively time-consuming and computationally expensive, especially for large-scale CNNs. In this paper, we propose a fast, scalable, and analytical framework to quantify the fault vulnerability of convolutional neural networks deployed on ReRAM-based PIM accelerators. This method quantifies vulnerability at different granularities, including weights, kernels, and layers, without the need for exhaustive simulation, enabling the design of fault-tolerant accelerators for edge and safety-critical applications. The comparison between the results of the proposed vulnerability score (VS) metrics with performance metric of CNNs captured with fault injection simulations confirms the effectiveness of our proposed method. Unlike exhaustive fault injection, which requires days or weeks, whereas the proposed approach achieves a speedup of about 10⁴×-10⁵×, enabling fast and accurate vulnerability analysis for large-scale CNNs within only a few minutes or hours.

Quoted from arXiv:2611.26307. The research and its words belong to its authors.

The authors

Aniseh Dorostkar, Hamed Farbeh, Hamid R. Zarandi, Meysam Sharifzadeh

Credit for this research belongs to its authors. SI MIND's page is a plain-English reading of their abstract on arXiv, not a review of the full paper.