Progress in artificial intelligence and machine learning over the past decade has been driven by the ability to train larger deep neural networks (DNNs), leading to a compute demand that far exceeds the growth in hardware performance afforded by Moore's law. Training DNNs is an extremely memory-intensive process, requiring not just the model weights but also activations and gradients for an entire minibatch to be stored. The need to provide high-density and low-leakage on-chip memory motivates the exploration of emerging non-volatile memory for training accelerators. Spin-Transfer-Torque MRAM (STT-MRAM) offers several desirable properties for training accelerators, including 3-4x higher density than SRAM, significantly reduced leakage power, high endurance and reasonable access time. On the one hand, MRAM write operations require high write energy and latency due to the need to ensure reliable switching. In this study, we perform a comprehensive device-to-system evaluation and co-optimization of STT-MRAM for efficient ML training accelerator design. We devised a cross-layer simulation framework to evaluate the effectiveness of STT-MRAM as a scratchpad replacing SRAM in a systolic-array-based DNN accelerator. To address the inefficiency of writes in STT-MRAM, we propose to reduce write voltage and duration. To evaluate the ensuing accuracy-efficiency trade-off, we conduct a thorough analysis of the error tolerance of input activations, weights, and errors during the training. We propose heterogeneous memory configurations that enable training convergence with good accuracy. We show that MRAM provide up to 15-22x improvement in system level energy across a suite of DNN benchmarks under iso-capacity and iso-area scenarios. Further optimizing STT-MRAM write operations can provide over 2x improvement in write energy for minimal degradation in application-level training accuracy.

Deep Neural Networks (DNNs) have gained significant interest in the recent past for plethora of applications such as image and video analytics, language translation, and medical diagnosis. High memory bandwidth is required to keep up with the needs of data-intensive DNN applications when implemented on a von-Neumann hardware architecture as majority of the data resides in the main memory. Therefore, processing in memory can provide a promising solution for the memory wall bottleneck for ML workloads. In this work, we propose a DRAM-based processing-in-memory (PIM) multiplication primitive coupled with intra-bank accumulation to accelerate matrix vector operations in ML workloads. Moreover, we propose a processing-in-memory DRAM bank architecture, data mapping and dataflow based on the proposed primitive. System evaluations performed on networks like AlexNet, VGG16 and ResNet18 show that the proposed architecture, mapping, and data flow can provide up to 23x and 6.5x benefits over a GPU and an ideal conventional (non-PIM) baseline architecture with infinite compute bandwidth, respectively.

Modern deep networks have millions to billions of parameters, which leads to high memory and energy requirements during training as well as during inference on resource-constrained edge devices. Consequently, pruning techniques have been proposed that remove less significant weights in deep networks, thereby reducing their memory and computational requirements. Pruning is usually performed after training the original network, and is followed by further retraining to compensate for the accuracy loss incurred during pruning. The prune-and-retrain procedure is repeated iteratively until an optimum tradeoff between accuracy and efficiency is reached. However, such iterative retraining adds to the overall training complexity of the network. In this work, we propose a dynamic pruning-while-training procedure, wherein we prune filters of the convolutional layers of a deep network during training itself, thereby precluding the need for separate retraining. We evaluate our dynamic pruning-while-training approach with three different pre-existing pruning strategies, viz. mean activation-based pruning, random pruning, and L1 normalization-based pruning. Our results for VGG-16 trained on CIFAR10 shows that L1 normalization provides the best performance among all the techniques explored in this work with less than 1% drop in accuracy after pruning 80% of the filters compared to the original network. We further evaluated the L1 normalization based pruning mechanism on CIFAR100. Results indicate that pruning while training yields a compressed network with almost no accuracy loss after pruning 50% of the filters compared to the original network and ~5% loss for high pruning rates (>80%). The proposed pruning methodology yields 41% reduction in the number of computations and memory accesses during training for CIFAR10, CIFAR100 and ImageNet compared to training with retraining for 10 epochs .