Back to search

University of Patras

Συνελικτικά vευρωνικά δίκτυα σε υπολογιστικές διατάξεις τεχνολογίας IMC (In-Memory Computing)

Abstract

dc:description

The rapid evolution of Artificial Intelligence (AI) applications and the continuously increasing complexity of Convolutional Neural Networks (CNNs) have led to a significant rise in computational and energy demands, particularly in edge systems with strict resource constraints. The efficient acceleration of the inference process in such environments requires architectures that drastically reduce the cost of accessing and transferring data to and from memory, addressing the so-called memory wall, which is exacerbated by the large volume of data that must be moved at each computational stage. At the same time, efficient support for convolution operations, which constitute the main computational workload of modern CNNs and can be equivalently reformulated as Matrix–Vector Multiplication (MVM) operations, is required. The need to reduce data movement and efficiently implement MVM operations has led to the development of new architectures, with In-Memory Computing (IMC) being a prominent example. IMC exploits the massive parallelism offered by memory arrays by moving computation closer to where data are stored. In this way, a significant reduction in data transfer cost is achieved, while energy efficiency and processing throughput are enhanced, particularly for MVM operations that dominate CNN workloads. This dissertation focuses on the use of Non-Volatile Memories (NVMs) for implementing IMC arrays, as they offer higher storage density. In these implementations, MVM operations are performed at the analog level, where weights are encoded as conductance values and computations are carried out through analog currents and voltages. However, the non-idealities introduced by these analog implementations, such as noise, variability, and nonlinearities, as well as the heterogeneous nature of hybrid architectures that integrate IMC accelerators with conventional digital units in System-on-Chip (SoC) platforms designed for CNN inference acceleration, make the evaluation of overall system behavior particularly challenging. Existing literature mainly focuses on low level simulation tools and device-level studies, as well as experimental silicon implementations, while limited emphasis is placed on tools that support system-level evaluation at a full architectural scale. Such tools, including emulators, have been successfully employed in other technological domains, significantly facilitating analysis and design decision-making through parameterization, observability, and repeatability. This dissertation introduces the In-Memory Computing Emulator (IMCE), a system-level emulation platform designed for studying CNN execution on hybrid analog–digital architectures (AIMC-Digital). IMCE is implemented as a complete system that integrates heterogeneous Processing Units (PUs), including units that emulate the functional behavior of analog IMC accelerators (AIMC PUs) and fully digital processing units(DPUs), together with their corresponding interconnection, coordination, and execution mechanisms. Through a comprehensive software stack, the platform executes real neural models and collects extensive runtime statistics, enabling controlled and repeatable experiments across the entire inference pipeline. The architecture of IMCE represents a typical edge AI chip, where heterogeneous computing units coexist under strict resource constraints. At the same time, it is inherently scalable and distributed, supporting parameterization of the number, type, and topology of PUs, as well as the extension beyond a single chip through dynamic addition of local or remote computing units. This capability enables the execution and systematic evaluation of complex and heterogeneous neural models requiring multiple cooperating PUs within a unified framework. The IMCE platform provides extensive, configurable execution statistics, which are used at multiple levels of evaluation: from accuracy estimation and performance assessment to optimization of CNN node-to-PU mapping and understanding of system dynamic behavior at runtime. The functional correctness and reliability of the platform are experimentally validated on convolutional and related inference workloads, such as ResNet and YOLO networks. Under ideal memory conditions, that is, ignoring the analog non-idealities of NVM-based implementations, the accuracy of IMCE is comparable to that of corresponding quantized digital implementations (INT8), demonstrating the feasibility of reliable CNN model execution. Furthermore, by leveraging the data collected during platform operation, the problem of mapping CNN graph nodes to heterogeneous AIMC and DPU resources is formulated and studied. The analysis focuses on nodes executed on AIMC units, as they represent the vast majority and include the most computationally demanding operations. The study is based on execution-time metrics, memory requirements for storing per node weights (storage footprint), and the topology of the computational graph, with emphasis on the critical path, that is, the path with the largest total execution time, and on parallel branches. The evaluation is conducted on ResNet-type networks as well as on controlled synthetic CoMes27 networks, enabling the isolation and analysis of topological and architectural factors. The results show that node-mapping algorithms that take into account overall computational load and graph topology (load-aware and topology aware)achieve higher throughput and lower end-to-end latency. In contrast, simplified algorithms, such as random assignment, round-robin scheduling, or methods based solely on equalizing memory usage or node count per PU, exhibit noticeable performance degradation. For realistic evaluation of hybrid AIMC–Digital systems, the analysis focuses on noise introduced by analog units based on Phase-Change Memory (PCM). To this end, parameterizable PCM noise models are incorporated, and the General PCM Model (GPM) is introduced as a unified and flexible framework for noise description. GPM covers two representative PCM models, enabling consistent and comparable noise-aware analysis of analog node mapping across different processing units. In this way, quantitative assessment of accuracy–performance trade-offs under realistic PCM noise conditions becomes possible. In conclusion, this dissertation demonstrates that effective utilization of hybrid AIMC–Digital architectures for CNN inference requires co-optimization of architecture, interconnection, node-to-PU mapping, and robustness to analog noise phenomena. IMCE provides a controlled and extensible emulation environment that bridges theoretical analysis and realistic silicon implementation, supporting informed architectural design prior to final deployment.

Degree

thesis:*
Grantor dc:publisher
University of Patras
Year dc:date
2026

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Bougioukou, Eleni
  • Μπουγιούκου, Ελένη

Subjects

dc:subject × 23

Rights

Language dc:language
gre

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/61368

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Bougioukou, Eleni; Μπουγιούκου, Ελένη. Συνελικτικά vευρωνικά δίκτυα σε υπολογιστικές διατάξεις τεχνολογίας IMC (In-Memory Computing). University of Patras, 2026. http://hdl.handle.net/10442/hedi/61368