Analog non-volatile memory (NVM)-based accelerators for Deep Neural Networks (DNNs) can achieve high-throughput and energy-efficient multiply-accumulate (MAC) operations by taking advantage of massively parallelized analog compute, implemented with Ohm's law and Kirchhoff's current law on arrays of resistive memory devices. Competitive end-to-end DNN accuracies can be obtained, provided that weights are accurately programmed onto NVM devices and MAC operations are sufficiently linear. In this paper, we report architectural and circuit advances for such Analog NVM-based accelerators. We describe a highly heterogeneous and programmable accelerator architecture for DNN inference that combines analog NVM memory-array “Tiles” for weight-stationary, energy-efficient MAC operations, together with heterogeneous special-function Compute-Cores for auxiliary digital computation. Massively parallel vectors of neuron-activation data are exchanged over short distances using a dense and efficient circuit-switched 2D mesh, enabling a wide range of DNN workloads, including CNNs, LSTMs, and Transformers. We also show a 14-nm inference chip consisting of multiple$\mathbf{512}\times \mathbf{512}$arrays of Phase Change Memory (PCM) devices which implements multiple DNN benchmarks using such a circuit-switched 2D mesh.

We report on aggressive quantization strategies that greatly accelerate inference of Recurrent Neural Network Transducers (RNN-T).We use a 4 bit integer representation for both weights and activations and apply Quantization Aware Training (QAT) to retrain the full model (acoustic encoder and language model) and achieve near-iso-accuracy.We show that customized quantization schemes that are tailored to the local properties of the network are essential to achieve good performance while limiting the computational overhead of QAT.Density ratio Language Model fusion has shown remarkable accuracy gains on RNN-T workloads but it severely increases the computational cost of inference.We show that our quantization strategies enable using large beam widths for hypothesis search while achieving streaming-compatible runtimes and a full model compression ratio of 7.6× compared to the full precision model.Via hardware simulations, we estimate a 3.4× acceleration from FP16 to INT4 for the end-to-end quantized RNN-T inclusive of LM fusion, resulting in a Real Time Factor (RTF) of 0.06.On the NIST Hub5 2000, Hub5 2001, and RT-03 test sets, we retain most of the gains associated with LM fusion, improving the average WER by >1.5%.

Analog non-volatile memory (NVM)-based accelerators for deep neural networks perform high-throughput and energy-efficient multiply-accumulate (MAC) operations (e.g., high TeraOPS/W) by taking advantage of massively parallelized analog MAC operations, implemented with Ohm’s law and Kirchhoff’s current law on array-matrices of resistive devices. While the wide-integer and floating-point operations offered by conventional digital CMOS computing are much more suitable than analog computing for conventional applications that require high accuracy and true reproducibility, deep neural networks can still provide competitive end-to-end results even with modest (e.g., 4-bit) precision in synaptic operations. In this paper, we describe a 14-nm inference chip, comprising multiple 512$\times$ 512 arrays of Phase Change Memory (PCM) devices, which can deliver software-equivalent inference accuracy for MNIST handwritten-digit recognition and recurrent LSTM benchmarks, by using compensation techniques to finesse analog-memory challenges such as conductance drift and noise. We also project accuracy for Natural Language Processing (NLP) tasks performed with a state-of-art large Transformer-based model, BERT, when mapped onto an extended version of this same fundamental chip architecture.

We investigate the impact of aggressive low-precision representations of weights and activations in two families of large LSTM-based architectures for Automatic Speech Recognition (ASR): hybrid Deep Bidirectional LSTM -Hidden Markov Models (DBLSTM-HMMs) and Recurrent Neural Network -Transducers (RNN-Ts).Using a 4-bit integer representation, a naïve quantization approach applied to the LSTM portion of these models results in significant Word Error Rate (WER) degradation.On the other hand, we show that minimal accuracy loss is achievable with an appropriate choice of quantizers and initializations.In particular, we customize quantization schemes depending on the local properties of the network, improving recognition performance while limiting computational time.We demonstrate our solution on the Switchboard (SWB) and CallHome (CH) test sets of the NIST Hub5-2000 evaluation.DBLSTM-HMMs trained with 300 or 2000 hours of SWB data achieves <0.5% and <1% average WER degradation, respectively.On the more challenging RNN-T models, our quantization strategy limits degradation in 4-bit inference to 1.3%.

By performing parallelized multiply-accumulate operations in the analog domain at the location of weight data, crossbar-array "tiles" of analog non-volatile memory (NVM) devices can potentially accelerate the forward-inference of deep neural networks. To be successful, such systems will need to achieve two related but challenging goals. First is the achievement of high neural network classification accuracies, indistinguishable from those achieved with conventional approaches, despite the difficulties of programming NVM devices accurately in the presence of significant device-to-device variability. Towards this first goal, we describe row-wise Phase-Change Memory (PCM) programming schemes for rapid yet accurate weight- programming. The second goal is highly energy-efficient forward- inference of multi-layer neural networks, requiring efficiency in both the massively-parallel analog-AI operations performed at each tile, as well as efficiency in how the resulting neuron-excitation data vectors get conveyed from tile to tile. Towards this second goal, micro-architectural design ideas including source-follower-based readout, array segmentation, and transmit-by- duration are described.

Neuromorphic computation based on analog nonvolatile memories (NVMs) holds great promise to improve Deep Neural Networks inference performance. In virtue of an architecture that executes the computation at the location of the stored weight data, remarkable gains in energy efficiency and speed are projected over competing von Neumann architectures leveraged by existing digital accelerators. Here we describe two optimization strategies for NVMs: one for programming the memory elements and one of cell design, both aimed at mitigating the effect of NVM non-idealities on the performance of analog, phase-change memory-based accelerators. We then demonstrate the advantages realized by such strategies on the inference accuracy of Long Short Term Memory networks and evaluate the energy requirements of such networks.