William J. Dally

Charbel Sakr, Steve Dai, Rangharajan Venkatesan et al.

Data clipping is crucial in reducing noise in quantization operations and improving the achievable accuracy of quantization-aware training (QAT). Current practices rely on heuristics to set clipping threshold scalars and cannot be shown to be optimal. We propose Optimally Clipped Tensors And Vectors (OCTAV), a recursive algorithm to determine MSE-optimal clipping scalars. Derived from the fast Newton-Raphson method, OCTAV finds optimal clipping scalars on the fly, for every tensor, at every iteration of the QAT routine. Thus, the QAT algorithm is formulated with provably minimum quantization noise at each step. In addition, we reveal limitations in common gradient estimation techniques in QAT and propose magnitude-aware differentiation as a remedy to further improve accuracy. Experimentally, OCTAV-enabled QAT achieves state-of-the-art accuracy on multiple tasks. These include training-from-scratch and retraining ResNets and MobileNets on ImageNet, and Squad fine-tuning using BERT models, where OCTAV-enabled QAT consistently preserves accuracy at low precision (4-to-6-bits). Our results require no modifications to the baseline training recipe, except for the insertion of quantization operations where appropriate.

Huizi Mao, Sibo Zhu, Song Han et al.

Object recognition is a fundamental problem in many video processing tasks, accurately locating seen objects at low computation cost paves the way for on-device video recognition. We propose PatchNet, an efficient convolutional neural network to match objects in adjacent video frames. It learns the patchwise correlation features instead of pixel features. PatchNet is very compact, running at just 58MFLOPs, $5\times$ simpler than MobileNetV2. We demonstrate its application on two tasks, video object detection and visual object tracking. On ImageNet VID, PatchNet reduces the flops of R-FCN ResNet-101 by 5x and EfficientDet-D0 by 3.4x with less than 1% mAP loss. On OTB2015, PatchNet reduces SiamFC and SiamRPN by 2.5x with no accuracy loss. Experiments on Jetson Nano further demonstrate 2.8x to 4.3x speed-ups associated with flops reduction. Code is open sourced at https://github.com/RalphMao/PatchNet.

Yujun Lin, Song Han, Huizi Mao et al.

Large-scale distributed training requires significant communication bandwidth for gradient exchange that limits the scalability of multi-node training, and requires expensive high-bandwidth network infrastructure. The situation gets even worse with distributed training on mobile devices (federated learning), which suffers from higher latency, lower throughput, and intermittent poor connections. In this paper, we find 99.9% of the gradient exchange in distributed SGD is redundant, and propose Deep Gradient Compression (DGC) to greatly reduce the communication bandwidth. To preserve accuracy during compression, DGC employs four methods: momentum correction, local gradient clipping, momentum factor masking, and warm-up training. We have applied Deep Gradient Compression to image classification, speech recognition, and language modeling with multiple datasets including Cifar10, ImageNet, Penn Treebank, and Librispeech Corpus. On these scenarios, Deep Gradient Compression achieves a gradient compression ratio from 270x to 600x without losing accuracy, cutting the gradient size of ResNet-50 from 97MB to 0.35MB, and for DeepSpeech from 488MB to 0.74MB. Deep gradient compression enables large-scale distributed training on inexpensive commodity 1Gbps Ethernet and facilitates distributed training on mobile. Code is available at: https://github.com/synxlin/deep-gradient-compression.

Forrest N. Iandola, Song Han, Matthew W. Moskewicz et al.

Recent research on deep neural networks has focused primarily on improving accuracy. For a given accuracy level, it is typically possible to identify multiple DNN architectures that achieve that accuracy level. With equivalent accuracy, smaller DNN architectures offer at least three advantages: (1) Smaller DNNs require less communication across servers during distributed training. (2) Smaller DNNs require less bandwidth to export a new model from the cloud to an autonomous car. (3) Smaller DNNs are more feasible to deploy on FPGAs and other hardware with limited memory. To provide all of these advantages, we propose a small DNN architecture called SqueezeNet. SqueezeNet achieves AlexNet-level accuracy on ImageNet with 50x fewer parameters. Additionally, with model compression techniques we are able to compress SqueezeNet to less than 0.5MB (510x smaller than AlexNet). The SqueezeNet architecture is available for download here: https://github.com/DeepScale/SqueezeNet

Swapnil Gandhi, Siva Hari, William J. Dally et al.

Recent methods expose intra-request parallelism in LLM outputs, allowing independent branches to decode concurrently. Existing serving systems execute these branches eagerly or under fixed caps. We show that both are brittle: eager admission inflates the shared decode step, degrading co-batched requests in serial stages, while conservative fixed caps forgo the throughput that motivated exposing branches in the first place. We call the excess step latency caused by admitted branches the branch externality and show that the safe width depends on batch composition, context lengths, and accumulated slack, all of which change continuously over a workload trace. We introduce TAPER, a per-step admission controller that treats extra branches as opportunistic work, admitted only when the predicted branch externality fits within the batch's current slack budget. Per-step regulation is practical because branch-level scheduling decouples compute from memory: branches share the request's prefix KV, so expanding or contracting width requires no memory reclamation. On Qwen3-32B, TAPER improves goodput by $1.77\times$ over IRP-Off and by $1.48\times$ over IRP-Eager, while maintaining over $95\%$ SLO attainment.

Steve Dai, Rangharajan Venkatesan, Haoxing Ren et al.

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 ($\approx$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 37% area saving and 24% energy saving 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 26% compared to an 8-bit baseline.

Huizi Mao, Xiaodong Yang, William J. Dally

Average precision (AP) is a widely used metric to evaluate detection accuracy of image and video object detectors. In this paper, we analyze object detection from videos and point out that AP alone is not sufficient to capture the temporal nature of video object detection. To tackle this problem, we propose a comprehensive metric, average delay (AD), to measure and compare detection delay. To facilitate delay evaluation, we carefully select a subset of ImageNet VID, which we name as ImageNet VIDT with an emphasis on complex trajectories. By extensively evaluating a wide range of detectors on VIDT, we show that most methods drastically increase the detection delay but still preserve AP well. In other words, AP is not sensitive enough to reflect the temporal characteristics of a video object detector. Our results suggest that video object detection methods should be additionally evaluated with a delay metric, particularly for latency-critical applications such as autonomous vehicle perception.

Huizi Mao, Taeyoung Kong, William J. Dally

Detecting objects in a video is a compute-intensive task. In this paper we propose CaTDet, a system to speedup object detection by leveraging the temporal correlation in video. CaTDet consists of two DNN models that form a cascaded detector, and an additional tracker to predict regions of interests based on historic detections. We also propose a new metric, mean Delay(mD), which is designed for latency-critical video applications. Experiments on the KITTI dataset show that CaTDet reduces operation count by 5.1-8.7x with the same mean Average Precision(mAP) as the single-model Faster R-CNN detector and incurs additional delay of 0.3 frame. On CityPersons dataset, CaTDet achieves 13.0x reduction in operations with 0.8% mAP loss.

Xingyu Liu, Jeff Pool, Song Han et al.

Convolutional Neural Networks (CNNs) are computationally intensive, which limits their application on mobile devices. Their energy is dominated by the number of multiplies needed to perform the convolutions. Winograd's minimal filtering algorithm (Lavin, 2015) and network pruning (Han et al., 2015) can reduce the operation count, but these two methods cannot be directly combined $-$ applying the Winograd transform fills in the sparsity in both the weights and the activations. We propose two modifications to Winograd-based CNNs to enable these methods to exploit sparsity. First, we move the ReLU operation into the Winograd domain to increase the sparsity of the transformed activations. Second, we prune the weights in the Winograd domain to exploit static weight sparsity. For models on CIFAR-10, CIFAR-100 and ImageNet datasets, our method reduces the number of multiplications by $10.4\times$, $6.8\times$ and $10.8\times$ respectively with loss of accuracy less than $0.1\%$, outperforming previous baselines by $2.0\times$-$3.0\times$. We also show that moving ReLU to the Winograd domain allows more aggressive pruning.

Huizi Mao, Song Han, Jeff Pool et al.

Sparsity helps reduce the computational complexity of deep neural networks by skipping zeros. Taking advantage of sparsity is listed as a high priority in next generation DNN accelerators such as TPU. The structure of sparsity, i.e., the granularity of pruning, affects the efficiency of hardware accelerator design as well as the prediction accuracy. Coarse-grained pruning creates regular sparsity patterns, making it more amenable for hardware acceleration but more challenging to maintain the same accuracy. In this paper we quantitatively measure the trade-off between sparsity regularity and prediction accuracy, providing insights in how to maintain accuracy while having more a more structured sparsity pattern. Our experimental results show that coarse-grained pruning can achieve a sparsity ratio similar to unstructured pruning without loss of accuracy. Moreover, due to the index saving effect, coarse-grained pruning is able to obtain a better compression ratio than fine-grained sparsity at the same accuracy threshold. Based on the recent sparse convolutional neural network accelerator (SCNN), our experiments further demonstrate that coarse-grained sparsity saves about 2x the memory references compared to fine-grained sparsity. Since memory reference is more than two orders of magnitude more expensive than arithmetic operations, the regularity of sparse structure leads to more efficient hardware design.

Angshuman Parashar, Minsoo Rhu, Anurag Mukkara et al.

Convolutional Neural Networks (CNNs) have emerged as a fundamental technology for machine learning. High performance and extreme energy efficiency are critical for deployments of CNNs in a wide range of situations, especially mobile platforms such as autonomous vehicles, cameras, and electronic personal assistants. This paper introduces the Sparse CNN (SCNN) accelerator architecture, which improves performance and energy efficiency by exploiting the zero-valued weights that stem from network pruning during training and zero-valued activations that arise from the common ReLU operator applied during inference. Specifically, SCNN employs a novel dataflow that enables maintaining the sparse weights and activations in a compressed encoding, which eliminates unnecessary data transfers and reduces storage requirements. Furthermore, the SCNN dataflow facilitates efficient delivery of those weights and activations to the multiplier array, where they are extensively reused. In addition, the accumulation of multiplication products are performed in a novel accumulator array. Our results show that on contemporary neural networks, SCNN can improve both performance and energy by a factor of 2.7x and 2.3x, respectively, over a comparably provisioned dense CNN accelerator.

Chenzhuo Zhu, Song Han, Huizi Mao et al.

Deep neural networks are widely used in machine learning applications. However, the deployment of large neural networks models can be difficult to deploy on mobile devices with limited power budgets. To solve this problem, we propose Trained Ternary Quantization (TTQ), a method that can reduce the precision of weights in neural networks to ternary values. This method has very little accuracy degradation and can even improve the accuracy of some models (32, 44, 56-layer ResNet) on CIFAR-10 and AlexNet on ImageNet. And our AlexNet model is trained from scratch, which means it's as easy as to train normal full precision model. We highlight our trained quantization method that can learn both ternary values and ternary assignment. During inference, only ternary values (2-bit weights) and scaling factors are needed, therefore our models are nearly 16x smaller than full-precision models. Our ternary models can also be viewed as sparse binary weight networks, which can potentially be accelerated with custom circuit. Experiments on CIFAR-10 show that the ternary models obtained by trained quantization method outperform full-precision models of ResNet-32,44,56 by 0.04%, 0.16%, 0.36%, respectively. On ImageNet, our model outperforms full-precision AlexNet model by 0.3% of Top-1 accuracy and outperforms previous ternary models by 3%.

Song Han, Junlong Kang, Huizi Mao et al.

Long Short-Term Memory (LSTM) is widely used in speech recognition. In order to achieve higher prediction accuracy, machine learning scientists have built larger and larger models. Such large model is both computation intensive and memory intensive. Deploying such bulky model results in high power consumption and leads to high total cost of ownership (TCO) of a data center. In order to speedup the prediction and make it energy efficient, we first propose a load-balance-aware pruning method that can compress the LSTM model size by 20x (10x from pruning and 2x from quantization) with negligible loss of the prediction accuracy. The pruned model is friendly for parallel processing. Next, we propose scheduler that encodes and partitions the compressed model to each PE for parallelism, and schedule the complicated LSTM data flow. Finally, we design the hardware architecture, named Efficient Speech Recognition Engine (ESE) that works directly on the compressed model. Implemented on Xilinx XCKU060 FPGA running at 200MHz, ESE has a performance of 282 GOPS working directly on the compressed LSTM network, corresponding to 2.52 TOPS on the uncompressed one, and processes a full LSTM for speech recognition with a power dissipation of 41 Watts. Evaluated on the LSTM for speech recognition benchmark, ESE is 43x and 3x faster than Core i7 5930k CPU and Pascal Titan X GPU implementations. It achieves 40x and 11.5x higher energy efficiency compared with the CPU and GPU respectively.

Song Han, Jeff Pool, Sharan Narang et al.

Modern deep neural networks have a large number of parameters, making them very hard to train. We propose DSD, a dense-sparse-dense training flow, for regularizing deep neural networks and achieving better optimization performance. In the first D (Dense) step, we train a dense network to learn connection weights and importance. In the S (Sparse) step, we regularize the network by pruning the unimportant connections with small weights and retraining the network given the sparsity constraint. In the final D (re-Dense) step, we increase the model capacity by removing the sparsity constraint, re-initialize the pruned parameters from zero and retrain the whole dense network. Experiments show that DSD training can improve the performance for a wide range of CNNs, RNNs and LSTMs on the tasks of image classification, caption generation and speech recognition. On ImageNet, DSD improved the Top1 accuracy of GoogLeNet by 1.1%, VGG-16 by 4.3%, ResNet-18 by 1.2% and ResNet-50 by 1.1%, respectively. On the WSJ'93 dataset, DSD improved DeepSpeech and DeepSpeech2 WER by 2.0% and 1.1%. On the Flickr-8K dataset, DSD improved the NeuralTalk BLEU score by over 1.7. DSD is easy to use in practice: at training time, DSD incurs only one extra hyper-parameter: the sparsity ratio in the S step. At testing time, DSD doesn't change the network architecture or incur any inference overhead. The consistent and significant performance gain of DSD experiments shows the inadequacy of the current training methods for finding the best local optimum, while DSD effectively achieves superior optimization performance for finding a better solution. DSD models are available to download at https://songhan.github.io/DSD.

Song Han, Xingyu Liu, Huizi Mao et al.

State-of-the-art deep neural networks (DNNs) have hundreds of millions of connections and are both computationally and memory intensive, making them difficult to deploy on embedded systems with limited hardware resources and power budgets. While custom hardware helps the computation, fetching weights from DRAM is two orders of magnitude more expensive than ALU operations, and dominates the required power. Previously proposed 'Deep Compression' makes it possible to fit large DNNs (AlexNet and VGGNet) fully in on-chip SRAM. This compression is achieved by pruning the redundant connections and having multiple connections share the same weight. We propose an energy efficient inference engine (EIE) that performs inference on this compressed network model and accelerates the resulting sparse matrix-vector multiplication with weight sharing. Going from DRAM to SRAM gives EIE 120x energy saving; Exploiting sparsity saves 10x; Weight sharing gives 8x; Skipping zero activations from ReLU saves another 3x. Evaluated on nine DNN benchmarks, EIE is 189x and 13x faster when compared to CPU and GPU implementations of the same DNN without compression. EIE has a processing power of 102GOPS/s working directly on a compressed network, corresponding to 3TOPS/s on an uncompressed network, and processes FC layers of AlexNet at 1.88x10^4 frames/sec with a power dissipation of only 600mW. It is 24,000x and 3,400x more energy efficient than a CPU and GPU respectively. Compared with DaDianNao, EIE has 2.9x, 19x and 3x better throughput, energy efficiency and area efficiency.

Song Han, Huizi Mao, William J. Dally

Neural networks are both computationally intensive and memory intensive, making them difficult to deploy on embedded systems with limited hardware resources. To address this limitation, we introduce "deep compression", a three stage pipeline: pruning, trained quantization and Huffman coding, that work together to reduce the storage requirement of neural networks by 35x to 49x without affecting their accuracy. Our method first prunes the network by learning only the important connections. Next, we quantize the weights to enforce weight sharing, finally, we apply Huffman coding. After the first two steps we retrain the network to fine tune the remaining connections and the quantized centroids. Pruning, reduces the number of connections by 9x to 13x; Quantization then reduces the number of bits that represent each connection from 32 to 5. On the ImageNet dataset, our method reduced the storage required by AlexNet by 35x, from 240MB to 6.9MB, without loss of accuracy. Our method reduced the size of VGG-16 by 49x from 552MB to 11.3MB, again with no loss of accuracy. This allows fitting the model into on-chip SRAM cache rather than off-chip DRAM memory. Our compression method also facilitates the use of complex neural networks in mobile applications where application size and download bandwidth are constrained. Benchmarked on CPU, GPU and mobile GPU, compressed network has 3x to 4x layerwise speedup and 3x to 7x better energy efficiency.

Song Han, Jeff Pool, John Tran et al.

Neural networks are both computationally intensive and memory intensive, making them difficult to deploy on embedded systems. Also, conventional networks fix the architecture before training starts; as a result, training cannot improve the architecture. To address these limitations, we describe a method to reduce the storage and computation required by neural networks by an order of magnitude without affecting their accuracy by learning only the important connections. Our method prunes redundant connections using a three-step method. First, we train the network to learn which connections are important. Next, we prune the unimportant connections. Finally, we retrain the network to fine tune the weights of the remaining connections. On the ImageNet dataset, our method reduced the number of parameters of AlexNet by a factor of 9x, from 61 million to 6.7 million, without incurring accuracy loss. Similar experiments with VGG-16 found that the number of parameters can be reduced by 13x, from 138 million to 10.3 million, again with no loss of accuracy.