a sparsity-aware training accelerator

What is it about?

Currently most TinyML devices only focus on inference, as trainingrequires much more hardware resources. In this paper, we introduceSPARK, an efficient hybrid acceleration architecture with run-timesparsity-aware scheduling for TinyML learning. Besides a stand-alone accelerator, an in-pipeline acceleration unit is integratedwithin the CPU pipeline to support simultaneous forward and back-ward propagation. To better utilize sparsity and improve hardwareutilization, a sparsity-aware acceleration scheduler is implementedto schedule the workload between two acceleration units. A unifiedmemory system is also constructed to support transposable datafetch, reducing memory access. We implement SPARK using TSMC22nm technology and evaluate different TinyML tasks. Comparedwith the baseline accelerator, SPARK achieves 4.1× performanceimprovement in average with only 2.27% area overhead. SPARKalso outperforms off-shelf edge devices in performance by 9.4×with 446.0× higher efficiency.

Featured Image

Photo by Vishnu Mohanan on Unsplash

Photo by Vishnu Mohanan on Unsplash