A defect-tolerant accelerator for emerging high-performance applications

Due to the evolution of technology constraints, especially energy constraints which may lead to heterogeneous multi-cores, and the increasing number of defects, the design of defect-tolerant accelerators for heterogeneous multi-cores may become a major micro-architecture research issue. Most custom circuits are highly defect sensitive, a single transistor can wreck such circuits. On the contrary, artificial neural networks (ANNs) are inherently error tolerant algorithms. And the emergence of high-performance applications implementing recognition and mining tasks, for which competitive ANN-based algorithms exist, drastically expands the potential application scope of a hardware ANN accelerator. However, while the error tolerance of ANN algorithms is well documented, there are few in-depth attempts at demonstrating that an actual hardware ANN would be tolerant to faulty transistors. Most fault models are abstract and cannot demonstrate that the error tolerance of ANN algorithms can be translated into the defect tolerance of hardware ANN accelerators. In this article, we introduce a hardware ANN geared towards defect tolerance and energy efficiency, by spatially expanding the ANN. In order to precisely assess the defect tolerance capability of this hardware ANN, we introduce defects at the level of transistors, and then assess the impact of such defects on the hardware ANN functional behavior. We empirically show that the conceptual error tolerance of neural networks does translate into the defect tolerance of hardware neural networks, paving the way for their introduction in heterogeneous multi-cores as intrinsically defect-tolerant and energy-efficient accelerators.