An energy-efficient deep convolutional neural network accelerator featuring conditional computing and low external memory access

With its algorithmic success in many machine learning tasks and applications, deep convolutional neural networks (DCNNs) have been implemented with custom hardware in a number of prior works. However, such works have not exploited conditional/approximate computing to the utmost toward eliminating redundant computations of CNNs. This article presents a DCNN accelerator featuring a novel conditional computing scheme that synergistically combines precision cascading (PC) with zero skipping (ZS). To reduce many redundant convolutions that are followed by max-pooling operations, we propose precision cascading, where the input features are divided into a number of low-precision groups and approximate convolutions with only the most significant bits (MSBs) are performed first. Based on this approximate computation, the full-precision convolution is performed only on the maximum pooling output that is found. This way, the total number of bit-wise convolutions can be reduced by ∼ 2× with < 0.8% degradation in ImageNet accuracy. PC provides the added benefit of increased sparsity per low-precision group, which we exploit with ZS to eliminate the clock cycles and external memory accesses. The proposed conditional computing scheme has been implemented with custom architecture in a 40-nm prototype chip, which achieves a peak energy efficiency of 24.97 TOPS/W at 0.6-V supply and a low external memory access of 0.0018 access/MAC with VGG-16 CNN for ImageNet classification and a peak energy efficiency of 28.51 TOPS/W at 0.9-V supply with FlowNet for Flying Chair data set.