Analyzing hardware accelerators’ reliability: from design exploration to fast evaluation with hyperscalers

The rising density and complexity of domain-specific accelerators make conventional design space exploration strategies prohibitively expensive. These costs further escalate when non-functional properties, such as resilience, must be assessed for safety-critical applications, highlighting the need for innovative early-stage evaluation frameworks. This work investigates methods for early estimation of design and resilience features in large hardware accelerators for edge computing and machine learning (e.g., vector processors, tensor cores, and stereo vision engines). We first outline design space exploration challenges at the micro-architectural level for configurable vector processors. We then assess the use of open-source hyperscaler frameworks to analyze fine-grained functional and non-functional properties of large accelerators via simulation and emulation. Results show that such frameworks are feasible solutions and provide a viable path to reducing the computational costs of early characterization and evaluation.