Sparsity-accelerated offset-diagonal optical processors
Abstract
Sparse neural networks facilitate accelerated computation and reduced energy consumption by minimizing the total number of operations required. However, conventional digital and analog computing platforms typically yield only modest performance gains when executing sparse models due to hardware primitives that remain unoptimized for these data structures. This work demonstrates a high-performance, diagonal-sparsity-accelerated optical computing system utilizing an offset-diagonal architecture that achieves near-ideal reductions in compute time as sparsity increases. Experimental validation on MNIST benchmarks confirms that the system maintains performance parity with digital baselines across a wide range of sparsity levels, from 0 % (dense) to 99 %. By concurrently leveraging intrinsic optical parallelism and high-performance structured sparse models, this platform establishes a strategic pathway for next-generation optical accelerators capable of delivering substantial improvements in both computational speed and energy efficiency.