arXiv subjects

Theofilos Spyrou

Publications and source records attributed to Theofilos Spyrou.

3 recordsLinked to original sources

arXiv2026

IMPACT: Constraint-Aware Search for Imperfectly Factorized Accelerator Mappings

The efficiency of Deep Neural Network (DNN) accelerators depends on how each tensor operation is mapped onto the hardware. A mapping specifies how the computation is tiled across the memory hierarchy, ordered, and parallelized across the processing elements. These choices form a combinatorially large search space, where efficient mappings are hard to find. Most violate hardware constraints, and most remaining mappings underutilize the hardware. Existing mappers either search only part of this space to keep the search tractable, excluding the most efficient mappings, or sample the full space randomly, so most mappings they evaluate are infeasible. We present IMPACT, a mapper that searches the full space. IMPACT pairs a genetic algorithm with a constraint-aware repair operator that converts each infeasible mapping into a feasible one instead of discarding it. On a flexible accelerator, IMPACT outperforms state-of-the-art mappers in EDP, latency, and energy on every DeepBench workload. On a compute-in-memory (CIM) accelerator, whose constraints leave few mappings feasible, it reduces EDP by 66%, latency by 47%, and energy by 18% over the strongest prior mapper across five networks, and rivals hand-crafted expert mappings. Across flexible and highly constrained accelerators, IMPACT finds efficient mappings automatically. IMPACT will be released as open source.

cs.AR↗

arXiv2026

PdNeuRAM: forming-free, multi-bit Pd/HfO2 ReRAM for energy-efficient neuromorphic computing

Memristor technology shows great promise for energy-efficient computing, yet it grapples with challenges like resistance drift and inherent variability. For filamentary Resistive RAM (ReRAM), one of the most investigated types of memristive devices, the expensive electroforming step required to create conductive pathways results in increased power and area overheads and reduced endurance. In this study, we present novel HfO2-based forming-free ReRAM devices, PdNeuRAM, that operate at low voltages, support multi-bit functionality, and display reduced variability. Through a deep understanding and comprehensive material characterization, we discover the key process that allows this unique behavior: a Pd-O-Hf configuration that capitalizes on Pd innate affinity for integrating into HfO2. This structure actively facilitates charge redistribution at room temperature, effectively eliminating the need for electroforming. Moreover, the fabricated ReRAM device provides tunable resistance states for dense memory and reduces programming and reading energy by 43% and 73%, respectively, using spiking neural networks (SNN). This study reveals novel mechanistic insights and delineates a strategic roadmap for the realization of power-efficient and cost-effective ReRAM devices.

cond-mat.mtrl-sci↗

arXiv2024

SpikeFI: A Fault Injection Framework for Spiking Neural Networks

Neuromorphic computing and spiking neural networks (SNNs) are gaining traction across various artificial intelligence (AI) tasks thanks to their potential for efficient energy usage and faster computation speed. This comparative advantage comes from mimicking the structure, function, and efficiency of the biological brain, which arguably is the most brilliant and green computing machine. As SNNs are eventually deployed on a hardware processor, the reliability of the application in light of hardware-level faults becomes a concern, especially for safety- and mission-critical applications. In this work, we propose SpikeFI, a fault injection framework for SNNs that can be used for automating the reliability analysis and test generation. SpikeFI is built upon the SLAYER PyTorch framework with fault injection experiments accelerated on a single or multiple GPUs. It has a comprehensive integrated neuron and synapse fault model library, in accordance to the literature in the domain, which is extendable by the user if needed. It supports: single and multiple faults; permanent and transient faults; specified, random layer-wise, and random network-wise fault locations; and pre-, during, and post-training fault injection. It also offers several optimization speedups and built-in functions for results visualization. SpikeFI is open-source and available for download via GitHub at https://github.com/SpikeFI.

cs.NE↗