To accelerate CNN inference, existing deep learning frameworks focus on optimizing intra-operator parallelization. However, a single operator can no longer fully utilize the available parallelism given the rapid advances in high-performance hardware, resulting in a large gap between the peak performance and the real performance. This performance gap is more severe under smaller batch sizes. In this work, we extensively study the parallelism between operators and propose Inter-Operator Scheduler (IOS) to automatically schedule multiple operators' parallel execution through a novel dynamic programming algorithm. IOS consistently outperforms state-of-the-art libraries (e.g., TensorRT) by 1.1 to 1.5x on modern CNN benchmarks. The code to reproduce each experiment is available at: https://github.com/mit-han-lab/inter-operator-scheduler.

Engineering a top-notch deep learning model is an expensive procedure that involves collecting data, hiring human resources with expertise in machine learning, and providing high computational resources. For that reason, deep learning models are considered as valuable Intellectual Properties (IPs) of the model vendors. To ensure reliable commercialization of deep learning models, it is crucial to develop techniques to protect model vendors against IP infringements. One of such techniques that recently has shown great promise is digital watermarking. However, current watermarking approaches can embed very limited amount of information and are vulnerable against watermark removal attacks. In this paper, we present GradSigns, a novel watermarking framework for deep neural networks (DNNs). GradSigns embeds the owner's signature into the gradient of the cross-entropy cost function with respect to inputs to the model. Our approach has a negligible impact on the performance of the protected model and it allows model vendors to remotely verify the watermark through prediction APIs. We evaluate GradSigns on DNNs trained for different image classification tasks using CIFAR-10, SVHN, and YTF datasets. Experimental results show that GradSigns is robust against all known counter-watermark attacks and can embed a large amount of information into DNNs.

Deep neural network (DNN) accelerators received considerable attention in past years due to saved energy compared to mainstream hardware. Low-voltage operation of DNN accelerators allows to further reduce energy consumption significantly, however, causes bit-level failures in the memory storing the quantized DNN weights. In this paper, we show that a combination of robust fixed-point quantization, weight clipping, and random bit error training (RandBET) improves robustness against random bit errors in (quantized) DNN weights significantly. This leads to high energy savings from both low-voltage operation as well as low-precision quantization. Our approach generalizes across operating voltages and accelerators, as demonstrated on bit errors from profiled SRAM arrays. We also discuss why weight clipping alone is already a quite effective way to achieve robustness against bit errors. Moreover, we specifically discuss the involved trade-offs regarding accuracy, robustness and precision: Without losing more than 1% in accuracy compared to a normally trained 8-bit DNN, we can reduce energy consumption on CIFAR-10 by 20%. Higher energy savings of, e.g., 30%, are possible at the cost of 2.5% accuracy, even for 4-bit DNNs.

Quantization enables efficient acceleration of deep neural networks by reducing model memory footprint and exploiting low-cost integer math hardware units. Quantization maps floating-point weights and activations in a trained model to low-bitwidth integer values using scale factors. Excessive quantization, reducing precision too aggressively, results in accuracy degradation. When scale factors are shared at a coarse granularity across many dimensions of each tensor, effective precision of individual elements within the tensor are limited. To reduce quantization-related accuracy loss, we propose using a separate scale factor for each small vector of (~16-64) elements within a single dimension of a tensor. To achieve an efficient hardware implementation, the per-vector scale factors can be implemented with low-bitwidth integers when calibrated using a two-level quantization scheme. We find that per-vector scaling consistently achieves better inference accuracy at low precision compared to conventional scaling techniques for popular neural networks without requiring retraining. We also modify a deep learning accelerator hardware design to study the area and energy overheads of per-vector scaling support. Our evaluation demonstrates that per-vector scaled quantization with 4-bit weights and activations achieves 69% energy saving and 36% area saving over an 8-bit baseline while maintaining over 75% accuracy for ResNet50 on ImageNet. 4-bit weights and 8-bit activations achieve near-full-precision accuracy for both BERT-base and BERT-large on SQuAD while reducing area by 28% compared to an 8-bit baseline.

Distributed model training suffers from communication bottlenecks due to frequent model updates transmitted across compute nodes. To alleviate these bottlenecks, practitioners use gradient compression techniques like sparsification, quantization, low rank updates etc. The techniques usually require choosing a static compression ratio, often requiring users to balance the trade-off between model accuracy and per-iteration speedup. In this work, we show that such performance degradation due to choosing a high compression ratio is not fundamental and that an adaptive compression strategy can reduce communication while maintaining final test accuracy.Inspired by recent findings on critical learning regimes, in which small gradient errors can have irrecoverable impact on model performance, we propose ACCORDION a simple yet effective adaptive compression algorithm. While ACCORDION maintains a high enough compression rate on average, it avoids detrimental impact by not compressing gradients too much whenever in critical learning regimes, detected by a simple gradient-norm based criterion. Our extensive experimental study over a number of machine learning tasks in distributed environments indicates that ACCORDION, maintains similar model accuracy to uncompressed training, yet achieves up to 5.5×better compression and up to 4.1×end-to-end speedup over static approaches. We show that ACCORDION also works for adjusting the batch size, another popular strategy for alleviating communication bottlenecks. Our code is available at https://github.com/uw-mad-dash/Accordion

Modern deep neural networks increasingly make use of features such as control flow, dynamic data structures, and dynamic tensor shapes. Existing deep learning systems focus on optimizing and executing static neural networks which assume a pre-determined model architecture and input data shapes—assumptions that are violated by dynamic neural networks. Therefore, executing dynamic models with deep learning systems is currently both inflexible and sub-optimal, if not impossible. Optimizing dynamic neural networks is more challenging than static neural networks; optimizations must consider all possible execution paths and tensor shapes. This paper proposes Nimble, a high-performance and flexible system to optimize, compile, and execute dynamic neural networks on multiple platforms. Nimble handles model dynamism by introducing a dynamic type system, a set of dynamism-oriented optimizations, and a light-weight virtual machine runtime. Our evaluation demonstrates that Nimble outperforms existing solutions for dynamic neural networks by up to 20x on hardware platforms including Intel CPUs, ARM CPUs, and Nvidia GPUs.

Mixed precision DNN accelerators become more ubiquitous especially when both efficient training and inference are required. In this paper, we propose a mixed-precision convolution unit architecture which supports different integer and floating point~(FP) precisions. The proposed architecture is based on low-bit inner product units and realizes higher precision based on temporal decomposition. We illustrate how to integrate FP computations on integer-based architecture and evaluate overheads incurred by FP arithmetic support. We argue that alignment and addition overhead for FP inner product can be significant since the maximum exponent difference could be up to 58 bits, which results into a large alignment logic. To address this issue, we illustrate empirically that at least 8 bits of alignment logic are required to maintain inference accuracy. We present novel optimizations based on the above observations to reduce the FP arithmetic hardware overheads. Our empirical results, based on simulation and hardware implementation, show significant reduction in FP16 overhead. Over typical mixed precision implementation, the proposed architecture achieves area improvements of up to 25\% in TFLOPS/$mm^2$ and up to 46\% in TOPS/$mm^2$ with power efficiency improvements of up to 40\% in TFLOPS/W and up to 63\% in TOPS/W.

The recent many-fold increase in the size of deep neural networks makes efficient distributed training challenging. Many proposals exploit the compressibility of the gradients and propose lossy compression techniques to speed up the communication stage of distributed training. Nevertheless, compression comes at the cost of reduced model quality and extra computation overhead. In this work, we design an efficient compressor with minimal overhead. Noting the sparsity of the gradients, we propose to model the gradients as random variables distributed according to some sparsity-inducing distributions (SIDs). We empirically validate our assumption by studying the statistical characteristics of the evolution of gradient vectors over the training process. We then propose Sparsity-Inducing Distribution-based Compression (SIDCo), a threshold-based sparsification scheme that enjoys similar threshold estimation quality to deep gradient compression (DGC) while being faster by imposing lower compression overhead. Our extensive evaluation of popular machine learning benchmarks involving both recurrent neural network (RNN) and convolution neural network (CNN) models shows that SIDCo speeds up training by up to ~41.7X, 7.6X, and 1.9X compared to the no-compression baseline, Topk, and DGC compressors, respectively.

As the usage of machine learning techniques is becoming ubiquitous, the efficient execution of deep learning models is crucial to many applications. Frameworks such as Halide or TVM separate the algorithmic representation of the neural network from the schedule that determines its implementation. Finding good schedules, however, remains extremely challenging. Autotuning methods, which search the space of valid schedules and execute each candidate on the hardware, identify some of the best performing schedules, but the search can take hours, hampering the productivity of deep learning practitioners. What is needed is a method that achieves a similar performance without extensive search, delivering the needed efficiency quickly.

We model the scheduling process as a sequence of optimization choices, and present a new technique to accurately predict the expected performance of a partial schedule using a LSTM over carefully engineered features that describe each DNN operator and their current scheduling choices. Leveraging these predictions we are able to make these optimization decisions greedily, and without any executions on the target hardware, quickly identify an efficient schedule.

Our evaluation shows that our performance predictions are one order of magnitude more accurate than the state of the art. This enables us to find schedules that improve the execution performance of deep neural networks by 1.5x or more over the best autoschedulers. Moreover, our technique is two to three orders of magnitude faster than these tools, and completes in seconds instead of hours.

Accurate hardware performance models are critical to efficient code generation. They can be used by compilers to make heuristic decisions, by superoptimizers as a minimization objective, or by autotuners to find an optimal configuration for a specific program. However, they are difficult to develop because contemporary processors are complex, and the recent proliferation of deep learning accelerators has increased the development burden. We demonstrate a method of learning performance models from a corpus of tensor computation graph programs for Tensor Processing Unit (TPU) instances. We show that our learned model outperforms a heavily-optimized analytical performance model on two tasks—tile-size selection and operator fusion—and that it helps an autotuner discover faster programs in a setting where access to TPUs is limited or expensive.

Virtual machines (VM) form the foundation of modern cloud computing as they help logically abstract per-user compute from shared physical infrastructure. Users of these services require VMs of varying sizes and configurations, which the provider places on a set of physical machines (PMs). VMs on the same physical PM share memory and CPU resources so a bad packing directly impacts the quality of user experience. We consider the placement of Firecracker VMs (a form of Micro-VMs) -- lightweight VMs that are typically used for short lived tasks. Our objective is to place each VM as it arrives, so that the peak to average ratio of resource usage across PMs is minimized. Placement is challenging as we need to consider resource use in multiple dimensions, such as CPU and memory, and because resource use changes over time. Past approaches to similar problems have suggested that one could forecast VM resource use for placement. We see that in our production traffic, Micro-VM resource use is spiky and short lived, and that forecasting algorithms are not useful. We evaluate Reinforcement Learning (RL) approaches for this task, but find that off-the-shelf RL algorithms are not always performant. We present a forecasting free algorithm, called FirePlace, that learns the placement decision using a variant of hindsight optimization, which we call hindsight imitation. We evaluate our approach using a production traffic trace of Firecracker usage AWS Lambda. FirePlace improves upon baseline algorithms by 10% on a production data trace of 100K Firecracker VMs.

Convolutional Neural Networks (ConvNets) enable computers to excel on vision learning tasks such as image classification, object detection. Recently, real-time inference on live data is becoming more and more important. From a system perspective, it requires fast inference on each single, incoming data item (e.g. 1 image). Two main-stream distributed model serving paradigms – data parallelism and model parallelism – are not necessarily desirable here, because we cannot further split a single input data piece via data parallelism, and model parallelism introduces huge communication overhead. To achieve live data inference with low latency, we propose sensAI, a novel and generic approach that decouples a CNN model into disconnected subnets, each is responsible for predicting certain class(es). We call this new model distribution paradigm as class parallelism. Experimental results show that, sensAI achieves up to 18x faster inference on single input data item with no or negligible accuracy loss on CIFAR-10, CIFAR-100 and ImageNet-1K datasets.

We introduce Larq Compute Engine (LCE), a state-of-the-art Binarized Neural Network (BNN) inference engine, and use this framework to investigate several important questions about the efficiency of BNNs and to design a new leading BNN architecture. LCE provides highly optimized implementations of binary operations and accelerates binary convolutions by 8.5 - 18.5x compared to their full-precision counterparts on Pixel 1 phones. LCE's integration with Larq and a sophisticated MLIR-based converter allow users to move smoothly from training to deployment. By extending TensorFlow and TensorFlow Lite, LCE supports models which combine binary and full-precision layers, and can be easily integrated into existing applications. Using LCE, we analyze the performance of existing BNN computer vision architectures and develop QuickNet, a simple, easy-to-reproduce BNN that outperforms existing binary networks in terms of latency and accuracy on ImageNet. Furthermore, we investigate the impact of full-precision shortcuts and the relationship between number of multiply-accumulate operations and model latency. We are convinced that empirical performance should drive BNN architecture design and hope this work will facilitate others to design, benchmark and deploy binary models.

Strong empirical evidence that one machine-learning algorithm A outperforms another one B, ideally calls for multiple trials optimizing the learning pipeline over sources of variation such as data sampling, augmentation, parameter initialization, and hyperparameters choices. This is prohibitively expensive, and corners are cut to reach conclusions. We model the whole benchmarking process and all sources of variation, revealing that variance due to data sampling, parameter initialization and hyperparameter choice impact markedly machine learning benchmark. We analyze the predominant comparison methods used today in the light of this variance. We show a counter-intuitive result that a biased estimator with more source of variation will give better results, closer to the ideal estimator at a 51× reduction in compute cost. Using this we perform a detailed study on the error rate of detecting improvements, on five different deep-learning tasks/architectures. This study leads us to propose recommendations for future performance comparisons.

Deep reinforcement learning (RL) has made groundbreaking advancements in robotics, data center management and other applications. Unfortunately, system-level bottlenecks in RL workloads are poorly understood; we observe fundamental structural differences in RL workloads that make them inherently less GPU-bound than supervised learning (SL). To explain where training time is spent in RL workloads, we propose RL-Scope, a cross-stack profiler that scopes low-level CPU/GPU resource usage to high-level algorithmic operations, and provides accurate insights by correcting for profiling overhead. Using RL-Scope, we survey RL workloads across its major dimensions including ML backend, RL algorithm, and simulator. For ML backends, we explain a 2.3× difference in runtime between equivalent PyTorch and TensorFlow algorithm implementations, and identify a bottleneck rooted in overly abstracted algorithm implementations. For RL algorithms and simulators, we show that on-policy algorithms are at least 3.5× more simulation-bound than off-policy algorithms. Finally, we profile a scale-up workload and demonstrate that GPU utilization metrics reported by commonly used tools dramatically inflate GPU usage, whereas RL-Scope reports true GPU-bound time. RL-Scope is an open-source tool available at https://github.com/UofT-EcoSystem/rlscope.

Training of large scale models on distributed clusters is a critical component of the machine learning pipeline. However, this training can easily be made to fail if some workers behave in an adversarial (Byzantine) fashion whereby they return arbitrary results to the parameter server (PS). A plethora of existing papers consider a variety of attack models and propose robust aggregation and/or computational redundancy to alleviate the effects of these attacks. In this work we consider an omniscient attack model where the adversary has full knowledge about the gradient computation assignments of the workers and can choose to attack (up to) any q out of K worker nodes to induce maximal damage. Our redundancy-based method ByzShield leverages the properties of bipartite expander graphs for the assignment of tasks to workers; this helps to effectively mitigate the effect of the Byzantine behavior. Specifically, we demonstrate an upper bound on the worst case fraction of corrupted gradients based on the eigenvalues of our constructions which are based on mutually orthogonal Latin squares and Ramanujan graphs. Our numerical experiments indicate over a 36% reduction on average in the fraction of corrupted gradients compared to the state of the art. Likewise, our experiments on training followed by image classification on the CIFAR-10 dataset show that ByzShield has on average a 20% advantage in accuracy under the most sophisticated attacks. ByzShield also tolerates a much larger fraction of adversarial nodes compared to prior work.

Deep neural networks are widely used in personalized recommendation systems. Unlike regular DNN inference workloads, recommendation inference is memory-bound due to the many random memory accesses needed to lookup the embedding tables. The inference is also heavily constrained in terms of latency because producing a recommendation for a user must be done in about tens of milliseconds. In this paper, we propose MicroRec, a high-performance inference engine for recommendation systems. MicroRec accelerates recommendation inference by (1) redesigning the data structures involved in the embeddings to reduce the number of lookups needed and (2) taking advantage of the availability of High-Bandwidth Memory (HBM) in FPGA accelerators to tackle the latency by enabling parallel lookups. We have implemented the resulting design on an FPGA board including the embedding lookup step as well as the complete inference process. Compared to the optimized CPU baseline (16 vCPU, AVX2-enabled), MicroRec achieves 13.8~14.7x speedup on embedding lookup alone and 2.5~5.4x speedup for the entire recommendation inference in terms of throughput. As for latency, CPU-based engines needs milliseconds for inferring a recommendation while MicroRec only takes microseconds, a significant advantage in real-time recommendation systems.

This paper proposes a range-bound-aware convolution layer that accelerates the inference of rectified linear unit (ReLU)-based convolutional neural networks (CNNs) for analyzing video streams. Since video analysis systems require to process each video frame in real-time, the computational cost of inference of CNNs must be reduced. Several techniques heuristically skip the computation for the current frame and reuse the results of the previous frame when the current and previous frames are sufficiently similar. However, for critical applications such as surveillance systems, their accuracy can be unsatisfactory because they sacrifice accuracy for efficiency. In contrast, our method reduces the computational cost of convolution layers accompanied by ReLU while producing exactly the same inference results as an original model. We utilize both temporal similarity of video frames and activation sparsity in ReLU-based CNNs to guarantee to skip truly redundant computations. We experimentally confirm that our method can accelerate widely used pre-trained CNNs with both CPU and GPU implementations.