Work

New techniques for the Reliability Evaluation of AI-oriented Hardware Accelerators

Year: 2026
Type: dissertation
Abstract: Artificial Intelligence (AI) has become a cornerstone of modern computing, enabling applications in di!erent domains such as autonomous vehicles, robotics, healthcare, finance, and High-Performance Computing (HPC).These domains demand high throughput, energy e"ciency, and low latency, which has driven the widespread adoption of specialized AI-oriented hardware accelerators.In practice, these platforms include devices such as Graphics Processing Units (GPUs), Deep Learning Accelerators (DLAs), and Neural Processing Units (NPUs), often integrating dedicated tensor-acceleration engines to speed up matrix operations.They execute Machine Learning (ML) and Deep Learning (DL) workloads through massive parallelism, using architectures composed of many arithmetic units performing Multiply-Accumulate (MAC) operations.As these accelerators move beyond data centers and into safety-critical domains-such as autonomous driving, aerospace, and advanced robotics-their reliability becomes a fundamental design concern.When used in HPC and data centers, their reliability is also a major concern, since the number of devices involved requires an extremely low per-device failure rate to guarantee an acceptable overall failure rate at the system level.AI-oriented accelerators are increasingly fabricated in deeply scaled semiconductor technologies that, while improving performance and power e"ciency, also increase susceptibility to hardware faults.Radiation-induced upsets, process variations, and device aging can cause transient or permanent faults in logic and memory structures.Depending on where they occur and how they propagate, such faults may lead to Silent Data Corruptions (SDCs), where results are corrupted without detection, or Detected Unrecoverable Errors (DUEs), which may trigger exceptions, hangs, or application/system crashes.In safety-critical deployments and HPC environments, these outcomes can translate into mispredictions, accuracy degradation, or catastrophic failures.The doctoral research presented in this thesis addresses these challenges by introducing new techniques for the reliability evaluation of AI-oriented hardware accelerators.The goal is twofold: (i) to develop methodologies for accurate and scalable reliability assessment across multiple abstraction levels, and (ii) to design lightweight mitigation strategies, co-optimized across hardware and software, that enhance fault resilience under realistic constraints.iii • First and foremost, I would like to express my deepest gratitude to my advisors,
Source: PORTO@Iris
Language: English
Open Access status: green
OpenAlex ID: W7216741214
Cites: 99
Cited by:
Related to: 10