FPGA-based high-performance neural network acceleration
MetadataShow full item record
In the last ten years, Artificial Intelligence through Deep Neural Networks (DNNs) has penetrated virtually every aspect of science, technology, and business. Advances are rapid with thousands of papers being published annually. Many types of DNNs have been and continue to be developed -- in this thesis, we address Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and Graph Neural Networks (GNNs) -- each with a different set of target applications and implementation challenges. The overall problem for all of these Neural Networks (NNs) is that their target applications generally pose stringent constraints on latency and throughput, but also have strict accuracy requirements. Much research has therefore gone into all aspects of improving NN quality and performance: algorithms, code optimization, acceleration with GPUs, and acceleration with hardware, both dedicated ASICs and off-the-shelf FPGAs. In this thesis, we concentrate on the last of these approaches. There have been many previous efforts in creating hardware to accelerate NNs. The problem designers face is that optimal NN models typically have significant irregularities, making them hardware unfriendly. One commonly used approach is to train NN models to follow regular computation and data patterns. This approach, however, can hurt the models' accuracy or lead to models with non-negligible redundancies. This dissertation takes a different approach. Instead of regularizing the model, we create architectures friendly to irregular models. Our thesis is that high-accuracy and high-performance NN inference and training can be achieved by creating a series of novel irregularity-aware architectures for Field-Programmable Gate Arrays (FPGAs). In four different studies on four different NN types, we find that this approach results in speedups of 2.1x to 3255x compared with carefully selected prior art; for inference, there is no change in accuracy. The bulk of this dissertation revolves around these studies, the various workload balancing techniques, and the resulting NN acceleration architectures. In particular, we propose four different architectures to handle, respectively, data structure level, operation level, bit level, and model level irregularities. At the data structure level, we propose AWB-GCN, which uses runtime workload rebalancing to handle Sparse Matrices Multiplications (SpMM) on extremely sparse and unbalanced input. With GNN inference as a case study, AWB-GCN achieves over 90% system efficiency, guarantees efficient off-chip memory access, and provides considerable speedups over CPUs (3255x), GPUs (80x), and a prior ASIC accelerator (5.1x). At the operation level, we propose O3BNN-R, which can detect redundant operations and prune them at run time. This works even for those that are highly data-dependent and unpredictable. With Binarized NNs (BNNs) as a case study, O3BNN-R can prune over 30% of the operations, without any accuracy loss, yielding speedups over state-of-the-art implementations on CPUs (1122x), GPUs (2.3x), and FPGAs (2.1x). At the bit level, we propose CQNN. CQNN embeds a Coarse-Grained Reconfigurable Architecture (CGRA) which can be programmed at runtime to support NN functions with various data-width requirements. Results show that CQNN can deliver us-level Quantized NN (QNN) inference. At the model level, we propose FPDeep, especially for training. In order to address model-level irregularity, FPDeep uses a novel model partitioning schemes to balance workload and storage among nodes. By using a hybrid of model and layer parallelism to train DNNs, FPDeep avoids the large gap that commonly occurs between training and testing accuracy due to the improper convergence to sharp minimizers (caused by large training batches). Results show that FPDeep provides scalable, fast, and accurate training and leads to 6.6x higher energy efficiency than GPUs.