Unstructured pruning introduces significant sparsity in deep neural networks (DNNs), enhancing accelerator hardware efficiency. However, three critical challenges constrain performance gains: 1) complex fetching logic for nonzero (NZ) data pairs; 2) load imbalance across processing elements (PEs); and 3) PE stalls from write-back contention. This brief proposes an energy-efficient accelerator addressing these inefficiencies through three innovations. First, we propose a Cartesian-product output-row-stationary (CPORS) dataflow that inherently matches NZ data pairs by sequentially fetching compressed data. Second, a multilevel partial sum reduction (MLPR) strategy minimizes write-back traffic and converts random PE stalls into manageable load imbalance. Third, a kernel sorting and load scheduling (KSLS) mechanism resolves PE idle/stall and achieves PE array-level load balancing, attaining 76.6% average PE utilization across all sparsity levels. Implemented in 22-nmCMOS, the accelerator delivers $1.85\times $ speedup and $1.4\times $ energy efficiency over baseline and achieves 25.8 TOPS/W peak energy efficiency at 90% sparsity.

Deep neural networks (DNNs) have achieved remarkable success in critical domains such as computer vision. However, their substantial model scale and computational demands hinder deployment on resource-constrained edge devices. Bit-serial accelerators (BSAs) leverage significant bit-level sparsity (BLS) in weights and activations to accelerate inference. Unstructured BLS causes hardware inefficiency, while existing static pruning methods cannot adapt to real-time activations. To address these challenges, we propose BitFleX, a BSA enabling runtime semi-structured pruning for both weights and activations. Specifically, we introduce Bit-Term Decomposition (BTD) encoding to enhance inherent BLS and reduce pruning complexity. Additionally, a pruning-error predictor dynamically selects operands for sparsification with minimal error. Experiments show BitFleX achieves 87.5% BLS in ViT-B with <1% Top-1 accuracy loss on ImageNet, yielding 5.86× speedup over baseline and 23.61 TOPS/W peak energy efficiency.