An Efficient Hybrid Deep Learning Accelerator for Compact and Heterogeneous CNNs

Abstract

Resource-efficient Convolutional Neural Networks (CNNs) are gaining more attention. These CNNs have relatively low computational and memory requirements. A common denominator among such CNNs is having more heterogeneity than traditional CNNs. This heterogeneity is present at two levels: intra-layer type and inter-layer type. Generic accelerators do not capture these levels of heterogeneity, which harms their efficiency. Consequently, researchers have proposed model-specific accelerators with dedicated engines. When designing an accelerator with dedicated engines, one option is to dedicate one engine per CNN layer. We refer to accelerators designed with this approach as single-engine single-layer (SESL). This approach enables optimizing each engine for its specific layer. However, such accelerators are resource-demanding and unscalable. Another option is to design a minimal number of dedicated engines such that each engine handles all layers of one type. We refer to these accelerators as single-engine multiple-layer (SEML). SEML accelerators capture the inter-layer-type but not the intra-layer-type heterogeneity. We propose the Fixed Budget Hybrid CNN Accelerator (FiBHA), a hybrid accelerator composed of an SESL part and an SEML part, each processing a subset of CNN layers. FiBHA captures more heterogeneity than SEML while being more resource-aware and scalable than SESL. Moreover, we propose a novel module, Fused Inverted Residual Bottleneck (FIRB), a fine-grained and memory-light SESL architecture building block. The proposed architecture is implemented and evaluated using high-level synthesis (HLS) on different Field Programmable Gate Arrays representing various resource budgets. Our evaluation shows that FiBHA improves the throughput by up to 4 x and 2.5 x compared to state-of-the-art SESL and SEML accelerators, respectively. Moreover, FiBHA reduces memory and energy consumption compared to an SEML accelerator. The evaluation also shows that FIRB reduces the required memory by up to 54%, and energy requirements by up to 35% compared to traditional pipelining.