Reducing Errors and Powers in LPDDR for DNN Inference: A Compression and IECC-Based Approach

Abstract

In modern edge systems, the demand for data processing, especially for complex DNN tasks, is rapidly increasing. To address this, various compression schemes have been proposed to enable on-device AI while meeting the strict power and storage constraints of edge devices. However, despite these advancements, the compatibility of the compression methods with edge device memory, such as LPDDR, has not been thoroughly investigated. LPDDR operates at low voltage and faces reliability challenges like cell leakage, which is particularly concerning for applications where accuracy is critical, such as Advanced Driver Assistance Systems (ADAS) or medical devices. To address these reliability concerns, an ECC engine, known as IECC, is employed within each LPDDR bank. While IECC improves reliability, it also incurs performance penalties due to Read-Modify-Write (RMW) operations and parity storage overheads. This paper introduces RELIA, a DNN weight compression scheme with three-stage protection, aimed at enabling power-efficient and reliable DNN operations in mobile environments. RELIA reduces the operation granularity of the IECC engine to eliminate RMW overhead. Additionally, it proposes a SEC-FOEC(72,64) scheme (Single Error Correction-Frequently Occurring Error Correction) that can correct 99.97% of LPDDR errors. To mitigate the added storage overhead, a compression scheme based on DNN weight characteristics is introduced. Experimental results show RELIA outperforms traditional IECC schemes, reducing power by 16.12%, cycles by 12.6%, energy by 30.62%, and storage by 22.78%, while offering superior reliability in DNN inference.