Abstract
Reliability issues stemming from device level non-idealities of non-volatile emerging technologies like ferroelectric field-effect transistors (FeFET), especially at scaled dimensions, cause substantial degradation in the accuracy of In-Memory crossbar-based AI systems. In this work, we present a variation-aware design technique to characterize the device level variations and to mitigate their impact on hardware accuracy employing a Bayesian Neural Network (BNN) approach. An effective conductance variation model is derived from the experimental measurements of cycle-to-cycle (C2C) and device-to-device (D2D) variations performed on FeFET devices fabricated using 28 nm high-$k$ metal gate technology. The variations were found to be a function of different conductance states within the given programming range, which sharply contrasts earlier efforts where a fixed variation dispersion was considered for all conductance values. Such variation characteristics formulated for three different device sizes at different read voltages were provided as prior variation information to the BNN to yield a more exact and reliable inference. Near-ideal accuracy for shallow networks (MLP5 and LeNet models) on the MNIST dataset and limited accuracy decline by ~3.8-16.1% for deeper AlexNet models on CIFAR10 dataset under a wide range of variations corresponding to different device sizes and read voltages, demonstrates the efficacy of our proposed device-algorithm co-design technique.
Introduction
In-Memory Computing (IMC) is a promising paradigm for accelerating artificial intelligence by performing computations directly within memory arrays, thereby mitigating the energy bottlenecks associated with data movement. Ferroelectric Field-Effect Transistors (FeFETs) are leading candidates for IMC due to their CMOS compatibility and multi-level cell capabilities. However, as FeFET technology scales to nanometer dimensions, it faces significant reliability challenges. Physical variations during manufacturing (device-to-device, D2D) and operational fluctuations (cycle-to-cycle, C2C) cause programmed conductance weights to deviate from their target values, severely degrading the accuracy of traditional neural networks.
Experimental Characterization of FeFET Devices
The researchers characterized FeFET devices fabricated using an industrial 28nm high-k metal gate (HKMG) process. To understand the impact of scaling, three different device sizes were evaluated: 0.24µm/0.24µm, 0.5µm/0.24µm, and 1µm/1µm.
As shown in Figure 1, the FeFETs exhibit controllable conductance levels through programming voltage ($V_{PRG}$) pulses. However, these levels are not perfectly stable, showing clear dependencies on device dimensions and programming range.
Modeling Device Variations
A key contribution of this work is the discovery that FeFET variation is not uniform across all conductance states. Instead, it is a function of the programmed conductance and the read voltage ($V_{Read}$).
Figure 3 illustrates the C2C and D2D variations observed experimentally. Smaller devices and lower read voltages were found to exhibit significantly higher relative variation. The authors developed an empirical polynomial model to capture this state-dependent noise, moving beyond the simplified fixed-noise models used in previous literature.
Bayesian Neural Network for Variation Resilience
To mitigate these hardware non-idealities, the authors employed a hardware-software co-design strategy using Bayesian Neural Networks (BNNs). Unlike standard networks with fixed weights, BNNs treat weights as probability distributions. By incorporating the realistic, state-dependent FeFET variation model as a "prior" during training (using the "Bayes by Backprop" method), the network learns to identify weight distributions that are inherently robust to the specific noise characteristics of the underlying hardware. This optimization process steers the model toward "flatter" regions of the loss landscape where weight fluctuations have minimal impact on performance.
Results and Discussion
The framework was evaluated on standard benchmarks including MNIST (using MLP5 and LeNet) and CIFAR-10 (using AlexNet).
- Near-Ideal Accuracy: For shallow networks on MNIST, the BNN approach achieved accuracy nearly identical to software-only models with zero noise.
- Robustness in Deep Networks: On the more complex AlexNet/CIFAR-10 task, the proposed method significantly outperformed traditional models. For the most challenging case (smallest device at 0.6V read voltage), the BNN improved accuracy by 46.6% compared to a standard non-Bayesian network.
- Importance of Realistic Priors: The study demonstrated that using device-specific variation models as priors is far more effective than using generic, fixed-noise Bayesian models.
Conclusion
This work demonstrates that algorithm-level awareness of hardware imperfections can compensate for device-level unreliability. While challenges such as the power-accuracy trade-off at different read voltages and potential spatial correlations in crossbar arrays remain for future study, this co-design approach provides a viable path for scalable and reliable FeFET-based In-Memory Computing.
Summary updated