Dynamic Reliability Management in Neuromorphic Computing
نویسندگان
چکیده
Neuromorphic computing systems execute machine learning tasks designed with spiking neural networks. These are embracing non-volatile memory to implement high-density and low-energy synaptic storage. Elevated voltages currents needed operate memories cause aging of CMOS-based transistors in each neuron synapse circuit the hardware, drifting transistor’s parameters from their nominal values. If these circuits used continuously for too long, parameter drifts cannot be reversed, resulting permanent degradation performance over time, eventually leading hardware faults. Aggressive device scaling increases power density temperature, which further accelerates aging, challenging reliable operation neuromorphic systems. Existing reliability-oriented techniques periodically de-stress all at fixed intervals, assuming worst-case operating conditions, without actually tracking run-time. To circuits, normal must interrupted, introduces latency spike generation propagation, impacting inter-spike interval hence, (e.g., accuracy). We observe that contrast long-term permanently damages short-term scaled CMOS is mostly due bias temperature instability. The latter heavily workload-dependent and, more importantly, partially reversible. propose a new architectural technique mitigate aging-related reliability problems by designing an intelligent run-time manager (NCRTM), dynamically de-stresses response during execution workloads, objective meeting target. NCRTM only when it absolutely necessary do so, otherwise reducing impact scheduling operations off critical path. evaluate state-of-the-art workloads on hardware. Our results demonstrate significantly improves marginal performance.
منابع مشابه
ReRAM-Based Neuromorphic Computing
Artificial neural networks are currently being used as a promising means to solve problems in machine learning and computer science [1–3]. These networks are loosely inspired by massively parallel biological neural systems, but they are often implemented on mostly sequential hardware computing platforms, ignoring key features of real neural processing systems, such as their ability to carry out...
متن کاملAutomotive Computing, Neuromorphic Computing, and Beyond
Welcome to 2018 and IEEE Micro’s 38th year. I’m delighted about last year’s issues. In addition to the traditional special issues on Hot Chips and Top Picks, we had special issues on cognitive architectures, architectures for the post-Moore era, Cool Chips and Hot Interconnects, and ultra-low-power processors. I deeply appreciate the efforts of everyone who reviewed the many articles submitted ...
متن کاملLarge-scale neuromorphic computing systems.
Neuromorphic computing covers a diverse range of approaches to information processing all of which demonstrate some degree of neurobiological inspiration that differentiates them from mainstream conventional computing systems. The philosophy behind neuromorphic computing has its origins in the seminal work carried out by Carver Mead at Caltech in the late 1980s. This early work influenced other...
متن کاملBackpropagation for Energy-Efficient Neuromorphic Computing
Solving real world problems with embedded neural networks requires both training algorithms that achieve high performance and compatible hardware that runs in real time while remaining energy efficient. For the former, deep learning using backpropagation has recently achieved a string of successes across many domains and datasets. For the latter, neuromorphic chips that run spiking neural netwo...
متن کاملذخیره در منابع من
با ذخیره ی این منبع در منابع من، دسترسی به آن را برای استفاده های بعدی آسان تر کنید
ژورنال
عنوان ژورنال: ACM Journal on Emerging Technologies in Computing Systems
سال: 2021
ISSN: ['1550-4832', '1550-4840']
DOI: https://doi.org/10.1145/3462330