Exploiting sparsity is a key technique in accelerating quantized convolutional neural network (CNN) inference on mobile devices. Prior sparse CNN accelerators largely exploit unstructured sparsity and achieve significant speedups. Due to the unbounded, largely unpredictable sparsity patterns, however, exploiting unstructured sparsity requires complicated hardware design with significant energy and area overhead, which is particularly detrimental to mobile/IoT inference scenarios where energy and area efficiency are crucial.We propose to exploit structured sparsity, more specifically, Density Bound Block (DBB) sparsity for both weights and activations. DBB block tensors bound the maximum number of non-zeros per block. DBB thus exposes statically predictable sparsity patterns that enable lean sparsity-exploiting hardware and efficient memory access. We propose new hardware primitives to implement DBB sparsity for (static) weights and (dynamic) activations, respectively, with very low overheads.Building on top of the primitives, we describe S2TA, a systolic array-based CNN accelerator that exploits joint weight and activation DBB sparsity and new dimensions of data reuse unavailable on the traditional systolic array. S2TA in 16nm achieves more than 2× speedup and energy reduction compared to a strong baseline of a systolic array with zero-value clock gating, over five popular CNN benchmarks. Compared to two recent non-systolic sparse accelerators, Eyeriss v2 (65nm) and SparTen (45nm), S2TA in 65nm uses about 2.2× per and 3.1× less energy inference, respectively.

The Straight-Through Estimator (STE) is widely used for back-propagating gradients through the quantization function, but the STE technique lacks a complete theoretical understanding. We propose an alternative methodology called alpha-blending (AB), which quantizes neural networks to low precision using stochastic gradient descent (SGD). Our AB method avoids STE approximation by replacing the quantized weight in the loss function by an affine combination of the quantized weight w_q and the corresponding full-precision weight w with non-trainable scalar coefficient alpha and (1- alpha). During training, alpha is gradually increased from 0 to 1; the gradient updates to the weights are through the full precision term, (1-alpha) * w, of the affine combination; the model is converted from full-precision to low precision progressively. To evaluate the AB method, a 1-bit BinaryNet on CIFAR10 dataset and 8-bits, 4-bits MobileNet v1, ResNet_50 v1/2 on ImageNet are trained using the alpha-blending approach, and the evaluation indicates that AB improves top-1 accuracy by 0.9\%, 0.82\% and 2.93\% respectively compared to the results of STE based quantization.