.png){kind=icon}
.png){kind=icon}
Return
Reconfigurable spatial-parallel stochastic computing for accelerating sparse convolutional neural networks
.png){kind=icon}
PRE
AI
DOI：10.1007/s11432-021-3519-1.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Edge devices play an increasingly important role in the convolutional neural network (CNN) inference. However, the large computation and storage requirements are challenging for resource- and power-constrained hardware. These limitations might be overcome by exploring the following: (a) error tolerance via approximate computing, such as stochastic computing (SC); (b) data sparsity, including the weight and activation sparsity. Although SC can perform complex calculations with compact and simple arithmetic circuits, traditional SC-based accelerators suffer from the low reconfigurability and long bitstream, further making it difficult to benefit from the data sparsity. In this paper, we propose spatial-parallel stochastic computing (SPSC), which improves the spatial parallelism of the SC-based multiplier to the full extent while consuming fewer logic gates than the fixed-point implementation. Moreover, we present SPA, a highly reconfigurable SPSC-based sparse CNN accelerator with the proposed hybrid zero-skipping scheme (HZSS), to efficiently take advantage of different zero-skipping strategies for different types of layers. Comprehensive experiments show that SPA with up to 2477.6 Gops/W outperforms existing several binary-weight accelerators, SC-based accelerators, and the sparse CNN accelerator considering energy efficiency.
Keywords:
convolutional neural networks
stochastic computing
sparse neural networks
energy-efficient accelerator
high reconfigurability
spatial parallelism
Journal
IF:
7.6
Papers:
4.9K
Citations:
8.9K