Graph Convolutional Networks (GCNs) are increasingly adopted in large-scale graph-based recommender systems. Training GCN requires the minibatch generator traversing graphs and sampling the sparsely located neighboring nodes to obtain their features. Since real-world graphs often exceed the capacity of GPU memory, current GCN training systems keep the feature table in host memory and rely on the CPU to collect sparse features before sending them to the GPUs. This approach, however, puts tremendous pressure on host memory bandwidth and the CPU. This is because the CPU needs to (1) read sparse features from memory, (2) write features into memory as a dense format, and (3) transfer the features from memory to the GPUs. In this work, we propose a novel GPU-oriented data communication approach for GCN training, where GPU threads directly access sparse features in host memory through zero-copy accesses without much CPU help. By removing the CPU gathering stage, our method significantly reduces the consumption of the host resources and data access latency. We further present two important techniques to achieve high host memory access efficiency by the GPU: (1) automatic data access address alignment to maximize PCIe packet efficiency, and (2) asynchronous zero-copy access and kernel execution to fully overlap data transfer with training. We incorporate our method into PyTorch and evaluate its effectiveness using several graphs with sizes up to 111 million nodes and 1.6 billion edges. In a multi-GPU training setup, our method is 65-92% faster than the conventional data transfer method, and can even match the performance of all-in-GPU-memory training for some graphs that fit in GPU memory.
With the increasing adoption of graph neural networks (GNNs) in the machine learning community, GPUs have become an essential tool to accelerate GNN training. However, training GNNs on very large graphs that do not fit in GPU memory is still a challenging task. Unlike conventional neural networks, mini-batching input samples in GNNs requires complicated tasks such as traversing neighboring nodes and gathering their feature values. While this process accounts for a significant portion of the training time, we find existing GNN implementations using popular deep neural network (DNN) libraries such as PyTorch are limited to a CPU-centric approach for the entire data preparation step. This "all-in-CPU" approach has negative impact on the overall GNN training performance as it over-utilizes CPU resources and hinders GPU acceleration of GNN training. To overcome such limitations, we introduce PyTorch-Direct, which enables a GPU-centric data accessing paradigm for GNN training. In PyTorch-Direct, GPUs are capable of efficiently accessing complicated data structures in host memory directly without CPU intervention. Our microbenchmark and end-to-end GNN training results show that PyTorch-Direct reduces data transfer time by 47.1% on average and speeds up GNN training by up to 1.6x. Furthermore, by reducing CPU utilization, PyTorch-Direct also saves system power by 12.4% to 17.5% during training. To minimize programmer effort, we introduce a new "unified tensor" type along with necessary changes to the PyTorch memory allocator, dispatch logic, and placement rules. As a result, users need to change at most two lines of their PyTorch GNN training code for each tensor object to take advantage of PyTorch-Direct.
Binary neural networks (BNNs) have 1-bit weights and activations. Such networks are well suited for FPGAs, as their dominant computations are bitwise arithmetic and the memory requirement is also significantly reduced. However, compared to start-of-the-art compact convolutional neural network (CNN) models, BNNs tend to produce a much lower accuracy on realistic datasets such as ImageNet. In addition, the input layer of BNNs has gradually become a major compute bottleneck, because it is conventionally excluded from binarization to avoid a large accuracy loss. This work proposes FracBNN, which exploits fractional activations to substantially improve the accuracy of BNNs. Specifically, our approach employs a dual-precision activation scheme to compute features with up to two bits, using an additional sparse binary convolution. We further binarize the input layer using a novel thermometer encoding. Overall, FracBNN preserves the key benefits of conventional BNNs, where all convolutional layers are computed in pure binary MAC operations (BMACs). We design an efficient FPGA-based accelerator for our novel BNN model that supports the fractional activations. To evaluate the performance of FracBNN under a resource-constrained scenario, we implement the entire optimized network architecture on an embedded FPGA (Xilinx Ultra96v2). Our experiments on ImageNet show that FracBNN achieves an accuracy comparable to MobileNetV2, surpassing the best-known BNN design on FPGAs with an increase of 28.9% in top-1 accuracy and a 2.5x reduction in model size. FracBNN also outperforms a recently introduced BNN model with an increase of 2.4% in top-1 accuracy while using the same model size. On the embedded FPGA device, FracBNN demonstrates the ability of real-time image classification.
High quality AI solutions require joint optimization of AI algorithms, such as deep neural networks (DNNs), and their hardware accelerators. To improve the overall solution quality as well as to boost the design productivity, efficient algorithm and accelerator co-design methodologies are indispensable. In this paper, we first discuss the motivations and challenges for the Algorithm/Accelerator co-design problem and then provide several effective solutions. Especially, we highlight three leading works of effective co-design methodologies: 1) the first simultaneous DNN/FPGA co-design method; 2) a bi-directional lightweight DNN and accelerator co-design method; 3) a differentiable and efficient DNN and accelerator co-search method. We demonstrate the effectiveness of the proposed co-design approaches using extensive experiments on both FPGAs and GPUs, with comparisons to existing works. This paper emphasizes the importance and efficacy of algorithm-accelerator co-design and calls for more research breakthroughs in this interesting and demanding area.
Quantization has been proven to be an effective method for reducing the computing and/or storage cost of DNNs. However, the trade-off between the quantization bitwidth and final accuracy is complex and non-convex, which makes it difficult to be optimized directly. Minimizing direct quantization loss (DQL) of the coefficient data is an effective local optimization method, but previous works often neglect the accurate control of the DQL, resulting in a higher loss of the final DNN model accuracy. In this paper, we propose a novel metric called Vector Loss. Based on this new metric, we develop a new quantization solution called VecQ, which can guarantee minimal direct quantization loss and better model accuracy. In addition, in order to speed up the proposed quantization process during model training, we accelerate the quantization process with a parameterized probability estimation method and template-based derivation calculation. We evaluate our proposed algorithm on MNIST, CIFAR, ImageNet, IMDB movie review and THUCNews text data sets with numerical DNN models. The results demonstrate that our proposed quantization solution is more accurate and effective than the state-of-the-art approaches yet with more flexible bitwidth support. Moreover, the evaluation of our quantized models on Saliency Object Detection (SOD) tasks maintains comparable feature extraction quality with up to 16$\times$ weight size reduction.
High quality AI solutions require joint optimization of AI algorithms and their hardware implementations. In this work, we are the first to propose a fully simultaneous, efficient differentiable DNN architecture and implementation co-search (EDD) methodology. We formulate the co-search problem by fusing DNN search variables and hardware implementation variables into one solution space, and maximize both algorithm accuracy and hardware implementation quality. The formulation is differentiable with respect to the fused variables, so that gradient descent algorithm can be applied to greatly reduce the search time. The formulation is also applicable for various devices with different objectives. In the experiments, we demonstrate the effectiveness of our EDD methodology by searching for three representative DNNs, targeting low-latency GPU implementation and FPGA implementations with both recursive and pipelined architectures. Each model produced by EDD achieves similar accuracy as the best existing DNN models searched by neural architecture search (NAS) methods on ImageNet, but with superior performance obtained within 12 GPU-hour searches. Our DNN targeting GPU is 1.40x faster than the state-of-the-art solution reported in Proxyless, and our DNN targeting FPGA delivers 1.45x higher throughput than the state-of-the-art solution reported in DNNBuilder.
To speedup Deep Neural Networks (DNN) accelerator design and enable effective implementation, we propose HybridDNN, a framework for building high-performance hybrid DNN accelerators and delivering FPGA-based hardware implementations. Novel techniques include a highly flexible and scalable architecture with a hybrid Spatial/Winograd convolution (CONV) Processing Engine (PE), a comprehensive design space exploration tool, and a complete design flow to fully support accelerator design and implementation. Experimental results show that the accelerators generated by HybridDNN can deliver 3375.7 and 83.3 GOPS on a high-end FPGA (VU9P) and an embedded FPGA (PYNQ-Z1), respectively, which achieve a 1.8x higher performance improvement compared to the state-of-art accelerator designs. This demonstrates that HybridDNN is flexible and scalable and can target both cloud and embedded hardware platforms with vastly different resource constraints.
Transformer-based models pre-trained on large-scale corpora achieve state-of-the-art accuracy for natural language processing tasks, but are too resource-hungry and compute-intensive to suit low-capability devices or applications with strict latency requirements. One potential remedy is model compression, which has attracted extensive attention. This paper summarizes the branches of research on compressing Transformers, focusing on the especially popular BERT model. BERT's complex architecture means that a compression technique that is highly effective on one part of the model, e.g., attention layers, may be less successful on another part, e.g., fully connected layers. In this systematic study, we identify the state of the art in compression for each part of BERT, clarify current best practices for compressing large-scale Transformer models, and provide insights into the inner workings of various methods. Our categorization and analysis also shed light on promising future research directions for achieving a lightweight, accurate, and generic natural language processing model.
Recent breakthroughs in Deep Neural Networks (DNNs) have fueled a growing demand for DNN chips. However, designing DNN chips is non-trivial because: (1) mainstream DNNs have millions of parameters and operations; (2) the large design space due to the numerous design choices of dataflows, processing elements, memory hierarchy, etc.; and (3) an algorithm/hardware co-design is needed to allow the same DNN functionality to have a different decomposition, which would require different hardware IPs to meet the application specifications. Therefore, DNN chips take a long time to design and require cross-disciplinary experts. To enable fast and effective DNN chip design, we propose AutoDNNchip - a DNN chip generator that can automatically generate both FPGA- and ASIC-based DNN chip implementation given DNNs from machine learning frameworks (e.g., PyTorch) for a designated application and dataset. Specifically, AutoDNNchip consists of two integrated enablers: (1) a Chip Predictor, built on top of a graph-based accelerator representation, which can accurately and efficiently predict a DNN accelerator's energy, throughput, and area based on the DNN model parameters, hardware configuration, technology-based IPs, and platform constraints; and (2) a Chip Builder, which can automatically explore the design space of DNN chips (including IP selection, block configuration, resource balancing, etc.), optimize chip design via the Chip Predictor, and then generate optimized synthesizable RTL to achieve the target design metrics. Experimental results show that our Chip Predictor's predicted performance differs from real-measured ones by < 10% when validated using 15 DNN models and 4 platforms (edge-FPGA/TPU/GPU and ASIC). Furthermore, accelerators generated by our AutoDNNchip can achieve better (up to 3.86X improvement) performance than that of expert-crafted state-of-the-art accelerators.