@article {SahDuw20A, title = {CyNAPSE: A Low-power Reconfigurable Neural Inference Accelerator for Spiking Neural Networks}, journal = {Journal of Signal Processing Systems}, volume = {92}, year = {2020}, abstract = {While neural network models keep scaling in depth and computational requirements, biologically accurate models are becoming more interesting for low-cost inference. Coupled with the need to bring more computation to the edge in resource-constrained embedded and IoT devices, specialized ultra-low power accelerators for spiking neural networks are being developed. Having a large variance in the models employed in these networks, these accelerators need to be flexible, user-configurable, performant and energy efficient. In this paper, we describe CyNAPSE, a fully digital accelerator designed to emulate neural dynamics of diverse spiking networks. Since the use case of our implementation is primarily concerned with energy efficiency, we take a closer look at the factors that could improve its energy consumption. We observe that while majority of its dynamic power consumption can be credited to memory traffic, its on-chip components suffer greatly from static leakage. Given that the event-driven spike processing algorithm is naturally memory-intensive and has a large number of idle processing elements, it makes sense to tackle each of these problems towards a more efficient hardware implementation. With a diverse set of network benchmarks, we incorporate a detailed study of memory patterns that ultimately informs our choice of an application-specific network-adaptive memory management strategy to reduce dynamic power consumption of the chip. Subsequently, we also propose and evaluate a leakage mitigation strategy for runtime control of idle power. Using both the RTL implementation and a software simulation of CyNAPSE, we measure the relative benefits of these undertakings. Results show that our adaptive memory management policy results in up to 22\% more reduction in dynamic power consumption.}, author = {Saunak Saha and Henry Duwe and Joseph Zambreno} } @conference {SahDuw19A, title = {An Adaptive Memory Management Strategy Towards Energy Efficient Machine Inference in Event-Driven Neuromorphic Accelerators}, booktitle = {Proceedings of the International Conference on Application-specific Systems, Architectures and Processors (ASAP)}, year = {2019}, month = {July}, abstract = {Spiking neural networks are viable alternatives to classical neural networks for edge processing in low-power embedded and IoT devices. To reap their benefits, neuromorphic network accelerators that tend to support deep networks still have to expend great effort in fetching synaptic states from a large remote memory. Since local computation in these networks is event-driven, memory becomes the major part of the system{\textquoteright}s energy consumption. In this paper, we explore various opportunities of data reuse that can help mitigate the redundant traffic for retrieval of neuron meta-data and post-synaptic weights. We describe CyNAPSE, a baseline neural processing unit and its accompanying software simulation as a general template for exploration on various levels. We then investigate the memory access patterns of three spiking neural network benchmarks that have significantly different topology and activity. With a detailed study of locality in memory traffic, we establish the factors that hinder conventional cache management philosophies from working efficiently for these applications. To that end, we propose and evaluate a domain-specific management policy that takes advantage of the forward visibility of events in a queue-based event-driven simulation framework. Subsequently, we propose network-adaptive enhancements to make it robust to network variations. As a result, we achieve 13-44\% reduction in system power consumption and 8-23\% improvement over conventional replacement policies.}, author = {Saunak Saha and Henry Duwe and Joseph Zambreno} }