Efficient image processing via memristive-based approximate in-memory computing

• Verfasst von: Seiler, Fabian [VerfasserIn]
• Verfasst von: Taherinejad, Nima [VerfasserIn]
• Titel: Efficient image processing via memristive-based approximate in-memory computing
• Verf.angabe: Fabian Seiler and Nima TaheriNejad, member, IEEE
• Jahr: 2024
• Umfang: 12 S.
• Illustrationen: Illustrationen, Diagramme
• Fussnoten: Gesehen am 21.05.2025
• Titel Quelle: Enthalten in: Institute of Electrical and Electronics EngineersIEEE transactions on computer-aided design of integrated circuits and systems
• Ort Quelle: New York, NY : Institute of Electrical and Electronics Engineers, 1982
• Jahr Quelle: 2024
• Band/Heft Quelle: 43(2024), 11, Seite 3312-3323
• ISSN Quelle: 1937-4151
• DOI: doi:10.1109/TCAD.2024.3438113
• Datenträger: Online-Ressource
• Sprache: eng
• Sach-SW: Adders
• Sach-SW: Approximate
• Sach-SW: Approximate computing
• Sach-SW: Computational efficiency
• Sach-SW: Energy consumption
• Sach-SW: image processing
• Sach-SW: Image processing
• Sach-SW: IMPLY
• Sach-SW: in-memory computing
• Sach-SW: In-memory computing
• Sach-SW: Logic
• Sach-SW: memristor
• Sach-SW: Memristors
• Sach-SW: Power demand
• Sach-SW: Topology
• K10plus-PPN: 1926221559

Abstract

Image processing algorithms continue to demand higher performance from computers. However, computer performance is not improving at the same rate as before. In response to the current challenges in enhancing computing performance, a wave of new technologies and computing paradigms is surfacing. Among these, memristors stand out as one of the most promising components due to their technological prospects and low power consumption. With efficient data storage capabilities and their ability to directly perform logical operations within the memory, they are well-suited for in-memory computation (IMC). Approximate computing emerges as another promising paradigm, offering improved performance metrics, notably speed. The tradeoff for this gain is the reduction of accuracy. In this article, we are using the stateful logic material implication (IMPLY) in the semi-serial topology and combine both the paradigms to further enhance the computational performance. We present three novel approximated adders that drastically improve speed and energy consumption with an normalized mean error distance (NMED) lower than 0.02 for most scenarios. We evaluated partially approximated Ripple carry adder (RCA) at the circuit-level and compared them to the State-of-the-Art (SoA). The proposed adders are applied in different image processing applications and the quality metrics are calculated. While maintaining acceptable quality, our approach achieves significant energy savings of 6%-38% and reduces the delay (number of computation cycles) by 5%-35%, demonstrating notable efficiency compared to exact calculations.