No abstract

No abstract

No abstract

Recurrent Neural Networks, in particular One-dimensional and Multidimensional Long Short-Term Memory (1D-LSTM and MD-LSTM) have achieved state-of-the-art classification accuracy in many applications such as machine translation, image caption generation, handwritten text recognition, medical imaging and many more. However, high classification accuracy comes at high compute, storage, and memory bandwidth requirements, which make their deployment challenging, especially for energy-constrained platforms such as portable devices. In comparison to CNNs, not so many investigations exist on efficient hardware implementations for 1D-LSTM especially under energy constraints, and there is no research publication on hardware architecture for MD-LSTM. In this article, we present two novel architectures for LSTM inference: a hardware architecture for MD-LSTM, and a DRAM-based Processing-in-Memory (DRAM-PIM) hardware architecture for 1D-LSTM. We present for the first time a hardware architecture for MD-LSTM, and show a trade-off analysis for accuracy and hardware cost for various precisions. We implement the new architecture as an FPGA-based accelerator that outperforms NVIDIA K80 GPU implementation in terms of runtime by up to 84× and energy efficiency by up to 1238× for a challenging dataset for historical document image binarization from DIBCO 2017 contest, and a well known MNIST dataset for handwritten digits recognition. Our accelerator demonstrates highest accuracy and comparable throughput in comparison to state-of-the-art FPGA-based implementations of multilayer perceptron for MNIST dataset. Furthermore, we present a new DRAM-PIM architecture for 1D-LSTM targeting energy efficient compute platforms such as portable devices. The DRAM-PIM architecture integrates the computation units in a close proximity to the DRAM cells in order to maximize the data parallelism and energy efficiency. The proposed DRAM-PIM design is 16.19 × more energy efficient as compared to FPGA implementation. The total chip area overhead of this design is 18 % compared to a commodity 8 Gb DRAM chip. Our experiments show that the DRAM-PIM implementation delivers a throughput of 1309.16 GOp/s for an optical character recognition application.

The Artificial Neural Networks (ANNs), like CNN/DNN and LSTM, are not biologically plausible. Despite their initial success, they cannot attain the cognitive capabilities enabled by the dynamic hierarchical associative memory systems of biological brains. The biologically plausible spiking brain models, e.g., cortex, basal ganglia, and amygdala, have a greater potential to achieve biological brain like cognitive capabilities. Bayesian Confidence Propagation Neural Network (BCPNN) is a biologically plausible spiking model of the cortex. A human-scale model of BCPNN in real-time requires 162 TFlop/s, 50 TBs of synaptic weight storage to be accessed with a bandwidth of 200 TBs. The spiking bandwidth is relatively modest at 250 GBs/s. A hand-optimized implementation of rodent scale BCPNN has been done on Tesla K80 GPUs require 3 kWs, we extrapolate from that a human scale network will require 3 MWs. These power numbers rule out such implementations for field deployment as cognition engines in embedded systems. The key innovation that this paper reports is that it is feasible and affordable to implement real-time BCPNN as a custom tiled application-specific integrated circuit (ASIC) in 28 nm technology with custom 3D DRAM - eBrainII - that consumes 3 kW for human scale and 12 watts for rodent scale. Such implementations eminently fulfill the demands for field deployment.