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%.

Hardware acceleration of Artificial Intelligence (AI) workloads has gained widespread popularity with its potential to deliver unprecedented performance and efficiency. An important challenge remains in how AI accelerators are programmed to sustain high utilization without impacting end-user productivity. Prior software optimizations start with an input graph and focus on node-level optimizations, viz. dataflows and hierarchical tiling, and graph-level optimizations such as operation fusion. However, little effort has been devoted to inter-node on-chip scratchpad memory (SPM) management in Deep Learning (DL) accelerators, whose significance is bolstered by the recent trends in complex network topologies and the emergence of eager execution in DL frameworks. We characterize and show that there exists up to a 5.2× performance gap in DL inference to be bridged using SPM management and propose OnSRAM, a novel SPM management framework integrated with the compiler runtime of a DL accelerator. We develop two variants, viz. OnSRAM-Static, which works on static graphs to identify data structures that can be lucratively held on-chip based on their size, liveness and significance, and OnSRAM-Eager, which targets an eager execution model (no graph) and uses a history-based speculative scheme to hold/discard data structures. We integrate OnSRAM with TensorFlow and analyze it on multiple accelerator configurations. Across a suite of 12 images, objects, and language networks, on a 3 TFLOP system with a 2 MB SPM and 32 GBps external memory bandwidth, OnSRAM-Static and OnSRAM-Eager achieve 1.02–4.8× and 1.02–3.1× reduction in inference latency (batch size of 1), over a baseline with no SPM management. In terms of energy savings, we observe average reductions of 1.51× (up to 4.1×) and 1.23× (up to 2.9×) for the static and eager execution scenarios, respectively.

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%.

The growing prevalence and computational demands of Artificial Intelligence (AI) workloads has led to widespread use of hardware accelerators in their execution. Scaling the performance of AI accelerators across generations is pivotal to their success in commercial deployments. The intrinsic error-resilient nature of AI workloads present a unique opportunity for performance/energy improvement through precision scaling. Motivated by the recent algorithmic advances in precision scaling for inference and training, we designed RaPiD1, a 4-core AI accelerator chip supporting a spectrum of precisions, namely, 16 and 8-bit floating-point and 4 and 2-bit fixed-point. The 36mm2RaPiD chip fabricated in 7nm EUV technology delivers a peak 3.5 TFLOPS/W in HFP8 mode and 16.5 TOPS/W in INT4 mode at nominal voltage. Using a performance model calibrated to within 1% of the measurement results, we evaluated DNN inference using 4-bit fixed-point representation for a 4-core 1 RaPiD chip system and DNN training using 8-bit floating point representation for a 768 TFLOPs AI system comprising 4 32-core RaPiD chips. Our results show INT4 inference for batch size of 1 achieves 3 - 13.5 (average 7) TOPS/W and FP8 training for a mini-batch of 512 achieves a sustained 102 - 588 (average 203) TFLOPS across a wide range of applications.

A prevalent challenge for Deep Learning (DL) accelerators is how they are programmed to sustain utilization without impacting end-user productivity. Little prior effort has been devoted to the effective management of their on-chip Scratch-Pad Memory (SPM) across the DL operations of a Deep Neural Network (DNN). This is especially critical due to trends in complex network topologies and the emergence of eager execution. This work demonstrates that there exists up to a 5.2x performance gap in DL inference to be bridged using SPM management, on a set of image, object and language networks. We propose OnSRAM, a novel SPM management framework integrated with a DL accelerator runtime. OnSRAM has two variants, viz. OnSRAM-Static, which works on static graphs to identify data structures that should be held on-chip based on their properties, and OnSRAM-Eager, which targets an eager execution model (no graph) and uses a speculative scheme to hold/discard data structures. On a prototypical DL accelerator, OnSRAM-Static and OnSRAM-Eager achieve reductions in inference latency (batch size of 1) of 1.02-4.8 x and 1.02-3.1 x, respectively, over a baseline with no SPM management.

Large-scale distributed training of Deep Neural Networks (DNNs) on state-of-the-art platforms are expected to be severely communication constrained. To overcome this limitation, numerous gradient compression techniques have been proposed and have demonstrated high compression ratios. However, most existing compression methods do not scale well to large scale distributed systems (due to gradient build-up) and / or lack evaluations in large datasets. To mitigate these issues, we propose a new compression technique, Scalable Sparsified Gradient Compression (ScaleComp), that (i) leverages similarity in the gradient distribution amongst learners to provide a commutative compressor and keep communication cost constant to worker number and (ii) includes low-pass filter in local gradient accumulations to mitigate the impacts of large batch size training and significantly improve scalability. Using theoretical analysis, we show that ScaleComp provides favorable convergence guarantees and is compatible with gradient all-reduce techniques. Furthermore, we experimentally demonstrate that ScaleComp has small overheads, directly reduces gradient traffic and provides high compression rates (70-150X) and excellent scalability (up to 64-80 learners and 10X larger batch sizes over normal training) across a wide range of applications (image, language, and speech) without significant accuracy loss.

Recent breakthroughs in deep neural networks (DNNs) have fueled a tremendous demand for intelligent edge devices featuring on-site learning, while the practical realization of such systems remains a challenge due to the limited resources available at the edge and the required massive training costs for state-of-the-art (SOTA) DNNs. As reducing precision is one of the most effective knobs for boosting training time/energy efficiency, there has been a growing interest in low-precision DNN training. In this paper, we explore from an orthogonal direction: how to fractionally squeeze out more training cost savings from the most redundant bit level, progressively along the training trajectory and dynamically per input. Specifically, we propose FracTrain that integrates (i) progressive fractional quantization which gradually increases the precision of activations, weights, and gradients that will not reach the precision of SOTA static quantized DNN training until the final training stage, and (ii) dynamic fractional quantization which assigns precisions to both the activations and gradients of each layer in an input-adaptive manner, for only "fractionally" updating layer parameters. Extensive simulations and ablation studies (six models, four datasets, and three training settings including standard, adaptation, and fine-tuning) validate the effectiveness of FracTrain in reducing computational cost and hardware-quantified energy/latency of DNN training while achieving a comparable or better (-0.12%~+1.87%) accuracy. For example, when training ResNet-74 on CIFAR-10, FracTrain achieves 77.6% and 53.5% computational cost and training latency savings, respectively, compared with the best SOTA baseline, while achieving a comparable (-0.07%) accuracy. Our codes are available at: https://github.com/RICE-EIC/FracTrain.

In this paper, we propose a number of novel techniques and numerical representation formats that enable, for the very first time, the precision of training systems to be aggressively scaled from 8-bits to 4-bits. To enable this advance, we explore a novel adaptive Gradient Scaling technique (Gradscale) that addresses the challenges of insufficient range and resolution in quantized gradients as well as explores the impact of quantization errors observed during model training. We theoretically analyze the role of bias in gradient quantization and propose solutions that mitigate the impact of this bias on model convergence. Finally, we examine our techniques on a spectrum of deep learning models in computer vision, speech, and NLP. In combination with previously proposed solutions for 4-bit quantization of weight and activation tensors, 4-bit training shows a non-significant loss in accuracy across application domains while enabling significant hardware acceleration (> 7X over state-of-the-art FP16 systems).

Advances in deep neural networks (DNNs) and the availability of massive real-world data have enabled superhuman levels of accuracy on many AI tasks and ushered the explosive growth of AI workloads across the spectrum of computing devices. However, their superior accuracy comes at a high computational cost, which necessitates approaches beyond traditional computing paradigms to improve their operational efficiency. Leveraging the application-level insight of error resilience, we demonstrate how approximate computing (AxC) can significantly boost the efficiency of AI platforms and play a pivotal role in the broader adoption of AI-based applications and services. To this end, we present RaPiD, a multi-tera operations per second (TOPS) AI hardware accelerator core (fabricated at 14-nm technology) that we built from the ground-up using AxC techniques across the stack including algorithms, architecture, programmability, and hardware. We highlight the workload-guided systematic explorations of AxC techniques for AI, including custom number representations, quantization/pruning methodologies, mixed-precision architecture design, instruction sets, and compiler technologies with quality programmability, employed in the RaPiD accelerator.

The resilience of Deep Learning (DL) training and inference workloads to low-precision computations, coupled with the demand for power-and area-efficient hardware accelerators for these workloads, has led to the emergence of 16-bit floating point formats as the precision of choice for DL hardware accelerators. This paper describes our optimized 16-bit format that has 6 exponent bits and 9 fraction bits, derived from a study of the range of values encountered in DL applications. We demonstrate that our format preserves the accuracy of DL networks, and we compare its ease-of-use for DL against IEEE-754 half-precision (5 exponent bits and 10 fraction bits) and bfloat16 (8 exponent bits and 7 fraction bits). Further, our format eliminated sub-normals and simplifies rounding modes and handling of corner cases. This streamlines floating-point unit logic and enables realization of a compact power-efficient computation engine.

Fixed-point implementations (FxP) are prominently used to realize Deep Neural Networks (DNNs) efficiently on energy-constrained platforms. The choice of bit-width is often constrained by the ability of FxP to represent the entire range of numbers in the datastructure with sufficient resolution. At low bit-widths (< 8 bits), state-of-the-art DNNs invariably suffer a loss in classification accuracy due to quantization/saturation errors.

Hardware accelerators are a promising solution to the stringent computational requirements of Deep Neural Networks (DNNs). Ranging from low-power IP cores to server class systems, various accelerator architectures with high TOPS/W peak processing efficiencies and flexibility to execute different DNN topologies have been proposed. Prior efforts improve core utilization through better data-flows and computation sequencing, but little effort has thus far been devoted to systematically programming DNN accelerators to extract best possible system utilization, particularly for DNN training, which can be parallelized across peta-scale systems. In this work, we address the hitherto open challenge of systematically mapping computations onto Peta-scale accelerator systems, comprising many (thousands of) processing cores spanning many chips, while maximizing overall system performance. We achieve this by characterizing the design space of possible mapping configurations, building a detailed performance model that incorporates every computation and data-transfer involved in DNN training, and using a design space exploration tool called DEEPSPATIALMATRIX to identify the performance optimal configuration. We highlight 4 key optimizations built within DEEPSPATIALMATRIX - hybrid data-model parallelism, inter-layer memory reuse, time-step pipelining, and dynamic spatial minibatching - each of which improve system utilization by carefully managing the available memory capacity and interconnect bandwidth to balance the compute vs. communication costs. On a 8-peta-FLOP accelerator system, we demonstrate 1.36×-32× improvement in training performance through our design space exploration and optimizations across image recognition (VGG16, ResNet50) and machine translation (GNMT) DNN models.

Reducing the numerical precision of data and computation is extremely effective in accelerating deep learning training workloads. Towards this end, 8-bit floating point representations (FP8) were recently proposed for DNN training. However, its applicability was demonstrated on a few selected models only and significant degradation is observed when popular networks such as MobileNet and Transformer are trained using FP8. This degradation is due to the inherent precision requirement difference in the forward and backward passes of DNN training. Using theoretical insights, we propose a hybrid FP8 (HFP8) format and DNN end-to-end distributed training procedure. We demonstrate, using HFP8, the successful training of deep learning models across a whole spectrum of applications including Image Classification, Object Detection, Language and Speech without accuracy degradation. Finally, we demonstrate that, by using the new 8 bit format, we can directly quantize a pre-trained model down to 8-bits without losing accuracy by simply fine-tuning batch normalization statistics. These novel techniques enable a new generations of 8-bit hardware that are robust for building and deploying neural network models.

Highly distributed training of Deep Neural Networks (DNNs) on future compute platforms (offering 100 of TeraOps/s of computational capacity) is expected to be severely communication constrained. To overcome this limitation, new gradient compression techniques are needed that are computationally friendly, applicable to a wide variety of layers seen in Deep Neural Networks and adaptable to variations in network architectures as well as their hyper-parameters. In this paper we introduce a novel technique - the Adaptive Residual Gradient Compression (AdaComp) scheme. AdaComp is based on localized selection of gradient residues and automatically tunes the compression rate depending on local activity. We show excellent results on a wide spectrum of state of the art Deep Learning models in multiple domains (vision, speech, language), datasets (MNIST, CIFAR10, ImageNet, BN50, Shakespeare), optimizers (SGD with momentum, Adam) and network parameters (number of learners, minibatch-size etc.). Exploiting both sparsity and quantization, we demonstrate end-to-end compression rates of ∼200× for fully-connected and recurrent layers, and ∼40× for convolutional layers, without any noticeable degradation in model accuracies.

Deep Neural Networks (DNNs) have emerged as a powerful and versatile set of techniques to address challenging artificial intelligence (AI) problems. Applications in domains such as image/video processing, natural language processing, speech synthesis and recognition, genomics and many others have embraced deep learning as the foundational technique. DNNs achieve superior accuracy for these applications using very large models which require 100s of MBs of data storage, ExaOps of computation and high bandwidth for data movement. Despite advances in computing systems, training state-of-the-art DNNs on large datasets takes several days/weeks, directly limiting the pace of innovation and adoption. In this paper, we discuss how these challenges can be addressed via approximate computing. Based on our earlier studies demonstrating that DNNs are resilient to numerical errors from approximate computing, we present techniques to reduce communication overhead of distributed deep learning training via adaptive residual gradient compression (AdaComp), and computation cost for deep learning inference via Prameterized clipping ACTivation (PACT) based network quantization. Experimental evaluation demonstrates order of magnitude savings in communication overhead for training and computational cost for inference while not compromising application accuracy.

With the ever growing popularity of deep learning, the tremendous complexity of deep neural networks is becoming problematic when one considers inference on resource constrained platforms. Binary networks have emerged as a potential solution, however, they exhibit a fundamentallimi-tation in realizing gradient-based learning as their activations are non-differentiable. Current work has so far relied on approximating gradients in order to use the back-propagation algorithm via the straight through estimator (STE). Such approximations harm the quality of the training procedure causing a noticeable gap in accuracy between binary neural networks and their full precision baselines. We present a novel method to train binary activated neural networks using true gradient-based learning. Our idea is motivated by the similarities between clipping and binary activation functions. We show that our method has minimal accuracy degradation with respect to the full precision baseline. Finally, we test our method on three benchmarking datasets: MNIST, CIFAR-10, and SVHN. For each benchmark, we show that continuous binarization using true gradient-based learning achieves an accuracy within 1.5% of the floating-point baseline, as compared to accuracy drops as high as 6% when training the same binary activated network using the STE.

The combination of growth in compute capabilities and availability of large datasets has led to a re-birth of deep learning. Deep Neural Networks (DNNs) have become state-of-the-art in a variety of machine learning tasks spanning domains across vision, speech, and machine translation. Deep Learning (DL) achieves high accuracy in these tasks at the expense of 100s of ExaOps of computation; posing significant challenges to efficient large-scale deployment in both resource-constrained environments and data centers.

No abstract available.

The state-of-the-art hardware platforms for training deep neural networks are moving from traditional single precision (32-bit) computations towards 16 bits of precision - in large part due to the high energy efficiency and smaller bit storage associated with using reduced-precision representations. However, unlike inference, training with numbers represented with less than 16 bits has been challenging due to the need to maintain fidelity of the gradient computations during back-propagation. Here we demonstrate, for the first time, the successful training of deep neural networks using 8-bit floating point numbers while fully maintaining the accuracy on a spectrum of deep learning models and datasets. In addition to reducing the data and computation precision to 8 bits, we also successfully reduce the arithmetic precision for additions (used in partial product accumulation and weight updates) from 32 bits to 16 bits through the introduction of a number of key ideas including chunk-based accumulation and floating point stochastic rounding. The use of these novel techniques lays the foundation for a new generation of hardware training platforms with the potential for 2-4 times improved throughput over today's systems.

The growing prominence and computational challenges imposed by Deep Neural Networks (DNNs) has fueled the design of specialized accelerator architectures and associated dataflows to improve their implementation efficiency. Each of these solutions serve as a datapoint on the throughput vs. energy trade-offs for a given DNN and a set of architectural constraints. In this paper, we set out to explore whether it is possible to systematically explore the design space so as to estimate a given DNN's (both inference and training) performance on an shared memory architecture specification using a variety of data-flows. To this end, we have developed a framework, DEEPMATRIX, which given a description of a DNN and a hardware architecture, automatically identifies how the computations of the DNN's layers need to partitioned and mapped on to the architecture such that the overall performance is maximized, while meeting the constraints imposed by the hardware (processing power, memory capacity, bandwidth etc.) We demonstrate DEEPMATRIX's effectiveness for the VGG DNN benchmark, showing the trade-offs and sensitivity of utilization based on different architecture constraints.

Deep Neural Networks (DNNs) have emerged as a powerful and versatile set of techniques showing successes on challenging artificial intelligence (AI) problems. Applications in domains such as image/video processing, autonomous cars, natural language processing, speech synthesis and recognition, genomics and many others have embraced deep learning as the foundation. DNNs achieve superior accuracy for these applications with high computational complexity using very large models which require 100s of MBs of data storage, exaops of computation and high bandwidth for data movement. In spite of these impressive advances, it still takes days to weeks to train state of the art Deep Networks on large datasets - which directly limits the pace of innovation and adoption. In this paper, we present a multi-pronged approach to address the challenges in meeting both the throughput and the energy efficiency goals for DNN training.

Training of large-scale deep neural networks is often constrained by the available computational resources. We study the effect of limited precision data representation and computation on neural network training. Within the context of low-precision fixed-point computations, we observe the rounding scheme to play a crucial role in determining the network’s behavior during training. Our results show that deep networks can be trained using only 16-bit wide fixed-point number representation when using stochastic rounding, and incur little to no degradation in the classification accuracy. We also demonstrate an energy-efficient hardware accelerator that implements low-precision fixed-point arithmetic with stochastic rounding

The memory capacity, computational power, communication bandwidth, energy consumption, and physical size of the brain all tend to scale with the number of synapses, which outnumber neurons by a factor of 10,000. Although progress in cortical simulations using modern digital computers has been rapid, the essential disparity between the classical von Neumann computer architecture and the computational fabric of the nervous system makes large-scale simulations expensive, power hungry, and time consuming. Over the last three decades, CMOS-based neuromorphic implementations of “electronic cortex” have emerged as an energy efficient alternative for modeling neuronal behavior. However, the key ingredient for electronic implementation of any self-learning system—programmable, plastic Hebbian synapses scalable to biological densities—has remained elusive. We demonstrate the viability of implementing such electronic synapses using nanoscale phase change devices. We introduce novel programming schemes for modulation of device conductance to closely mimic the phenomenon of Spike Timing Dependent Plasticity (STDP) observed biologically, and verify through simulations that such plastic phase change devices should support simple correlative learning in networks of spiking neurons. Our devices, when arranged in a crossbar array architecture, could enable the development of synaptronic systems that approach the density (∼10 11 synapses per sq cm) and energy efficiency (consuming ∼1pJ per synaptic programming event) of the human brain.