Special Session: Sustainable Deployment of Deep Neural Networks on Non-Volatile Compute-in-Memory Accelerators

2025-08-16

AI Summary

Non-volatile memory-based compute-in-memory (NVM-CIM) accelerators suffer from degraded DNN inference accuracy due to device-level random variations, necessitating high-overhead write-verification during deployment. Method: This paper proposes Variation-Aware Forward Training (OVF), a training methodology that explicitly models and compensates for hardware variations during the forward pass using a negative-feedback optimization mechanism. OVF enhances model robustness via targeted variational forward propagation, thereby reducing epistemic uncertainty. Contribution/Results: Experimental evaluation across multiple DNNs demonstrates that OVF improves inference accuracy by up to 46.71%, while simultaneously reducing energy consumption and deployment latency. By mitigating reliance on write-verification, OVF significantly enhances the practicality and sustainability of NVM-CIM architectures.

Technology Category

• Machine Learning: Hardware-aware ML
• Computer Vision: Learning & Optimization for CV
• Cognitive Modeling & Cognitive Systems: Conceptual Inference and Reasoning

Application Category

• User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendation
• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications
• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs

Abstract

Non-volatile memory (NVM) based compute-in-memory (CIM) accelerators have emerged as a sustainable solution to significantly boost energy efficiency and minimize latency for Deep Neural Networks (DNNs) inference due to their in-situ data processing capabilities. However, the performance of NVCIM accelerators degrades because of the stochastic nature and intrinsic variations of NVM devices. Conventional write-verify operations, which enhance inference accuracy through iterative writing and verification during deployment, are costly in terms of energy and time. Inspired by negative feedback theory, we present a novel negative optimization training mechanism to achieve robust DNN deployment for NVCIM. We develop an Oriented Variational Forward (OVF) training method to implement this mechanism. Experiments show that OVF outperforms existing state-of-the-art techniques with up to a 46.71% improvement in inference accuracy while reducing epistemic uncertainty. This mechanism reduces the reliance on write-verify operations and thus contributes to the sustainable and practical deployment of NVCIM accelerators, addressing performance degradation while maintaining the benefits of sustainable computing with NVCIM accelerators.

Problem

• Addressing performance degradation in NVM-based CIM accelerators for DNNs
• Reducing energy and time costs of write-verify operations in NVCIM
• Enhancing inference accuracy and robustness in sustainable DNN deployment

Innovation

• Negative feedback training for robust DNN deployment
• Oriented Variational Forward method improves accuracy
• Reduces write-verify operations for sustainable NVCIM