Hadi Esmaeilzadeh

Hadi Esmaeilzadeh, Soroush Ghodrati, Andrew B. Kahng et al.

Parameterizable machine learning (ML) accelerators are the product of recent breakthroughs in ML. To fully enable their design space exploration (DSE), we propose a physical-design-driven, learning-based prediction framework for hardware-accelerated deep neural network (DNN) and non-DNN ML algorithms. It adopts a unified approach that combines backend power, performance, and area (PPA) analysis with frontend performance simulation, thereby achieving a realistic estimation of both backend PPA and system metrics such as runtime and energy. In addition, our framework includes a fully automated DSE technique, which optimizes backend and system metrics through an automated search of architectural and backend parameters. Experimental studies show that our approach consistently predicts backend PPA and system metrics with an average 7% or less prediction error for the ASIC implementation of two deep learning accelerator platforms, VTA and VeriGOOD-ML, in both a commercial 12 nm process and a research-oriented 45 nm process.

Hadi Esmaeilzadeh, Soroush Ghodrati, Andrew B. Kahng et al.

Today's performance analysis frameworks for deep learning accelerators suffer from two significant limitations. First, although modern convolutional neural network (CNNs) consist of many types of layers other than convolution, especially during training, these frameworks largely focus on convolution layers only. Second, these frameworks are generally targeted towards inference, and lack support for training operations. This work proposes a novel performance analysis framework, SimDIT, for general ASIC-based systolic hardware accelerator platforms. The modeling effort of SimDIT comprehensively covers convolution and non-convolution operations of both CNN inference and training on a highly parameterizable hardware substrate. SimDIT is integrated with a backend silicon implementation flow and provides detailed end-to-end performance statistics (i.e., data access cost, cycle counts, energy, and power) for executing CNN inference and training workloads. SimDIT-enabled performance analysis reveals that on a 64X64 processing array, non-convolution operations constitute 59.5% of total runtime for ResNet-50 training workload. In addition, by optimally distributing available off-chip DRAM bandwidth and on-chip SRAM resources, SimDIT achieves 18X performance improvement over a generic static resource allocation for ResNet-50 inference.

Zheng Li, Soroush Ghodrati, Amir Yazdanbakhsh et al.

Self-attention is a key enabler of state-of-art accuracy for various transformer-based Natural Language Processing models. This attention mechanism calculates a correlation score for each word with respect to the other words in a sentence. Commonly, only a small subset of words highly correlates with the word under attention, which is only determined at runtime. As such, a significant amount of computation is inconsequential due to low attention scores and can potentially be pruned. The main challenge is finding the threshold for the scores below which subsequent computation will be inconsequential. Although such a threshold is discrete, this paper formulates its search through a soft differentiable regularizer integrated into the loss function of the training. This formulation piggy backs on the back-propagation training to analytically co-optimize the threshold and the weights simultaneously, striking a formally optimal balance between accuracy and computation pruning. To best utilize this mathematical innovation, we devise a bit-serial architecture, dubbed LeOPArd, for transformer language models with bit-level early termination microarchitectural mechanism. We evaluate our design across 43 back-end tasks for MemN2N, BERT, ALBERT, GPT-2, and Vision transformer models. Post-layout results show that, on average, LeOPArd yields 1.9x and 3.9x speedup and energy reduction, respectively, while keeping the average accuracy virtually intact (<0.2% degradation)

Annabelle Sujun Tang, Christopher Priebe, Lianhui Qin et al.

Model serving costs dominate AI systems, making compiler optimization essential for scalable deployment. Recent works show that a large language model (LLM) can guide compiler search by reasoning over program structure and optimization history. However, using a single large model throughout the search is expensive, while smaller models are less reliable when used alone. Thus, this paper seeks to answer whether multi-LLM collaborative reasoning relying primarily on small LLMs can match or exceed the performance of a single large model. As such, we propose a lightweight collaborative multi-LLM framework, dubbed COLT, for compiler optimization that enables coordinated reasoning across multiple models within a single Monte Carlo tree search (MCTS) process. A key contribution is the use of a single shared MCTS tree as the collaboration substrate across LLMs, enabling the reuse of transformation prefixes and cross-model value propagation. Hence, we circumvent both heavy internal reasoning mechanisms and conventional agentic machinery that relies on external planners, multiple concurrent LLMs, databases, external memory/versioning of intermediate results, and controllers by simply endogenizing model selection within the lightweight MCTS optimization loop. Every iteration, the acting LLM proposes a joint action: (compiler transformation, model to be queried next). We also introduce a model-aware tree policy that biases search toward smaller models while preserving exploration, and a course-alteration mechanism that escalates to the largest model when the search exhibits persistent regressions attributable to smaller models.

Sujun Tang, Christopher Priebe, Rohan Mahapatra et al.

While model serving has unlocked unprecedented capabilities, the high cost of serving large-scale models continues to be a significant barrier to widespread accessibility and rapid innovation. Compiler optimizations have long driven substantial performance improvements, but existing compilers struggle with neural workloads due to the exponentially large and highly interdependent space of possible transformations. Although existing stochastic search techniques can be effective, they are often sample-inefficient and fail to leverage the structural context underlying compilation decisions. We set out to investigate the research question of whether reasoning with large language models (LLMs), without any retraining, can leverage the context-aware decision space of compiler optimizations to significantly improve sample efficiency. To that end, we introduce a novel compilation framework (dubbed Reasoning Compiler) that formulates optimization as a sequential, context-aware decision process guided by a large language model and structured Monte Carlo tree search (MCTS). The LLM acts as a proposal mechanism, suggesting hardware-informed transformations that reflect the current program state and accumulated performance feedback. MCTS incorporates the LLM-generated proposals to balance exploration and exploitation, facilitating structured, context-sensitive traversal of the expansive compiler optimization space. By achieving substantial speedups with markedly fewer samples than leading neural compilers, our approach demonstrates the potential of LLM-guided reasoning to transform the landscape of compiler optimization.

Hadi Esmaeilzadeh, Reza Vaezi

Recent advances in artificial intelligence (AI) have achieved human-scale speed and accuracy for classification tasks. In turn, these capabilities have made AI a viable replacement for many human activities that at their core involve classification, such as basic mechanical and analytical tasks in low-level service jobs. Current systems do not need to be conscious to recognize patterns and classify them. However, for AI to progress to more complicated tasks requiring intuition and empathy, it must develop capabilities such as metathinking, creativity, and empathy akin to human self-awareness or consciousness. We contend that such a paradigm shift is possible only through a fundamental shift in the state of artificial intelligence toward consciousness, a shift similar to what took place for humans through the process of natural selection and evolution. As such, this paper aims to theoretically explore the requirements for the emergence of consciousness in AI. It also provides a principled understanding of how conscious AI can be detected and how it might be manifested in contrast to the dominant paradigm that seeks to ultimately create machines that are linguistically indistinguishable from humans.

Fatemehsadat Mireshghallah, Mohammadkazem Taram, Praneeth Vepakomma et al.

The ever-growing advances of deep learning in many areas including vision, recommendation systems, natural language processing, etc., have led to the adoption of Deep Neural Networks (DNNs) in production systems. The availability of large datasets and high computational power are the main contributors to these advances. The datasets are usually crowdsourced and may contain sensitive information. This poses serious privacy concerns as this data can be misused or leaked through various vulnerabilities. Even if the cloud provider and the communication link is trusted, there are still threats of inference attacks where an attacker could speculate properties of the data used for training, or find the underlying model architecture and parameters. In this survey, we review the privacy concerns brought by deep learning, and the mitigating techniques introduced to tackle these issues. We also show that there is a gap in the literature regarding test-time inference privacy, and propose possible future research directions.

Soroush Ghodrati, Hardik Sharma, Cliff Young et al.

Conventional neural accelerators rely on isolated self-sufficient functional units that perform an atomic operation while communicating the results through an operand delivery-aggregation logic. Each single unit processes all the bits of their operands atomically and produce all the bits of the results in isolation. This paper explores a different design style, where each unit is only responsible for a slice of the bit-level operations to interleave and combine the benefits of bit-level parallelism with the abundant data-level parallelism in deep neural networks. A dynamic collection of these units cooperate at runtime to generate bits of the results, collectively. Such cooperation requires extracting new grouping between the bits, which is only possible if the operands and operations are vectorizable. The abundance of Data Level Parallelism and mostly repeated execution patterns, provides a unique opportunity to define and leverage this new dimension of Bit-Parallel Vector Composability. This design intersperses bit parallelism within data-level parallelism and dynamically interweaves the two together. As such, the building block of our neural accelerator is a Composable Vector Unit that is a collection of Narrower-Bitwidth Vector Engines, which are dynamically composed or decomposed at the bit granularity. Using six diverse CNN and LSTM deep networks, we evaluate this design style across four design points: with and without algorithmic bitwidth heterogeneity and with and without availability of a high-bandwidth off-chip memory. Across these four design points, Bit-Parallel Vector Composability brings (1.4x to 3.5x) speedup and (1.1x to 2.7x) energy reduction. We also comprehensively compare our design style to the Nvidia RTX 2080 TI GPU, which also supports INT-4 execution. The benefits range between 28.0x and 33.7x improvement in Performance-per-Watt.

Fatemehsadat Mireshghallah, Mohammadkazem Taram, Ali Jalali et al.

When receiving machine learning services from the cloud, the provider does not need to receive all features; in fact, only a subset of the features are necessary for the target prediction task. Discerning this subset is the key problem of this work. We formulate this problem as a gradient-based perturbation maximization method that discovers this subset in the input feature space with respect to the functionality of the prediction model used by the provider. After identifying the subset, our framework, Cloak, suppresses the rest of the features using utility-preserving constant values that are discovered through a separate gradient-based optimization process. We show that Cloak does not necessarily require collaboration from the service provider beyond its normal service, and can be applied in scenarios where we only have black-box access to the service provider's model. We theoretically guarantee that Cloak's optimizations reduce the upper bound of the Mutual Information (MI) between the data and the sifted representations that are sent out. Experimental results show that Cloak reduces the mutual information between the input and the sifted representations by 85.01% with only a negligible reduction in utility (1.42%). In addition, we show that Cloak greatly diminishes adversaries' ability to learn and infer non-conducive features.

Byung Hoon Ahn, Jinwon Lee, Jamie Menjay Lin et al.

Recent advances demonstrate that irregularly wired neural networks from Neural Architecture Search (NAS) and Random Wiring can not only automate the design of deep neural networks but also emit models that outperform previous manual designs. These designs are especially effective while designing neural architectures under hard resource constraints (memory, MACs, . . . ) which highlights the importance of this class of designing neural networks. However, such a move creates complication in the previously streamlined pattern of execution. In fact one of the main challenges is that the order of such nodes in the neural network significantly effects the memory footprint of the intermediate activations. Current compilers do not schedule with regard to activation memory footprint that it significantly increases its peak compared to the optimum, rendering it not applicable for edge devices. To address this standing issue, we present a memory-aware compiler, dubbed SERENITY, that utilizes dynamic programming to find a sequence that finds a schedule with optimal memory footprint. Our solution also comprises of graph rewriting technique that allows further reduction beyond the optimum. As such, SERENITY achieves optimal peak memory, and the graph rewriting technique further improves this resulting in 1.68x improvement with dynamic programming-based scheduler and 1.86x with graph rewriting, against TensorFlow Lite with less than one minute overhead.

Ahmed T. Elthakeb, Prannoy Pilligundla, Fatemehsadat Mireshghallah et al.

As deep neural networks make their ways into different domains, their compute efficiency is becoming a first-order constraint. Deep quantization, which reduces the bitwidth of the operations (below 8 bits), offers a unique opportunity as it can reduce both the storage and compute requirements of the network super-linearly. However, if not employed with diligence, this can lead to significant accuracy loss. Due to the strong inter-dependence between layers and exhibiting different characteristics across the same network, choosing an optimal bitwidth per layer granularity is not a straight forward. As such, deep quantization opens a large hyper-parameter space, the exploration of which is a major challenge. We propose a novel sinusoidal regularization, called SINAREQ, for deep quantized training. Leveraging the sinusoidal properties, we seek to learn multiple quantization parameterization in conjunction during gradient-based training process. Specifically, we learn (i) a per-layer quantization bitwidth along with (ii) a scale factor through learning the period of the sinusoidal function. At the same time, we exploit the periodicity, differentiability, and the local convexity profile in sinusoidal functions to automatically propel (iii) network weights towards values quantized at levels that are jointly determined. We show how SINAREQ balance compute efficiency and accuracy, and provide a heterogeneous bitwidth assignment for quantization of a large variety of deep networks (AlexNet, CIFAR-10, MobileNet, ResNet-18, ResNet-20, SVHN, and VGG-11) that virtually preserves the accuracy. Furthermore, we carry out experimentation using fixed homogenous bitwidths with 3- to 5-bit assignment and show the versatility of SINAREQ in enhancing quantized training algorithms (DoReFa and WRPN) with about 4.8% accuracy improvements on average, and then outperforming multiple state-of-the-art techniques.

Byung Hoon Ahn, Prannoy Pilligundla, Amir Yazdanbakhsh et al.

Achieving faster execution with shorter compilation time can foster further diversity and innovation in neural networks. However, the current paradigm of executing neural networks either relies on hand-optimized libraries, traditional compilation heuristics, or very recently genetic algorithms and other stochastic methods. These methods suffer from frequent costly hardware measurements rendering them not only too time consuming but also suboptimal. As such, we devise a solution that can learn to quickly adapt to a previously unseen design space for code optimization, both accelerating the search and improving the output performance. This solution dubbed Chameleon leverages reinforcement learning whose solution takes fewer steps to converge, and develops an adaptive sampling algorithm that not only focuses on the costly samples (real hardware measurements) on representative points but also uses a domain-knowledge inspired logic to improve the samples itself. Experimentation with real hardware shows that Chameleon provides 4.45x speed up in optimization time over AutoTVM, while also improving inference time of the modern deep networks by 5.6%.

Ahmed T. Elthakeb, Prannoy Pilligundla, Alex Cloninger et al.

The deep layers of modern neural networks extract a rather rich set of features as an input propagates through the network. This paper sets out to harvest these rich intermediate representations for quantization with minimal accuracy loss while significantly reducing the memory footprint and compute intensity of the DNN. This paper utilizes knowledge distillation through teacher-student paradigm (Hinton et al., 2015) in a novel setting that exploits the feature extraction capability of DNNs for higher-accuracy quantization. As such, our algorithm logically divides a pretrained full-precision DNN to multiple sections, each of which exposes intermediate features to train a team of students independently in the quantized domain. This divide and conquer strategy, in fact, makes the training of each student section possible in isolation while all these independently trained sections are later stitched together to form the equivalent fully quantized network. Our algorithm is a sectional approach towards knowledge distillation and is not treating the intermediate representation as a hint for pretraining before one knowledge distillation pass over the entire network (Romero et al., 2015). Experiments on various DNNs (AlexNet, LeNet, MobileNet, ResNet-18, ResNet-20, SVHN and VGG-11) show that, this approach -- called DCQ (Divide and Conquer Quantization) -- on average, improves the performance of a state-of-the-art quantized training technique, DoReFa-Net (Zhou et al., 2016) by 21.6% and 9.3% for binary and ternary quantization, respectively. Additionally, we show that incorporating DCQ to existing quantized training methods leads to improved accuracies as compared to previously reported by multiple state-of-the-art quantized training methods.

Byung Hoon Ahn, Prannoy Pilligundla, Hadi Esmaeilzadeh

Achieving faster execution with shorter compilation time can enable further diversity and innovation in neural networks. However, the current paradigm of executing neural networks either relies on hand-optimized libraries, traditional compilation heuristics, or very recently, simulated annealing and genetic algorithms. Our work takes a unique approach by formulating compiler optimizations for neural networks as a reinforcement learning problem, whose solution takes fewer steps to converge. This solution, dubbed ReLeASE, comes with a sampling algorithm that leverages clustering to focus the costly samples (real hardware measurements) on representative points, subsuming an entire subspace. Our adaptive sampling not only reduces the number of samples, but also improves the quality of samples for better exploration in shorter time. As such, experimentation with real hardware shows that reinforcement learning with adaptive sampling provides 4.45x speed up in optimization time over AutoTVM, while also improving inference time of the modern deep networks by 5.6%. Further experiments also confirm that our adaptive sampling can even improve AutoTVM's simulated annealing by 4.00x.

Fatemehsadat Mireshghallah, Mohammadkazem Taram, Prakash Ramrakhyani et al.

A wide variety of deep neural applications increasingly rely on the cloud to perform their compute-heavy inference. This common practice requires sending private and privileged data over the network to remote servers, exposing it to the service provider and potentially compromising its privacy. Even if the provider is trusted, the data can still be vulnerable over communication channels or via side-channel attacks in the cloud. To that end, this paper aims to reduce the information content of the communicated data with as little as possible compromise on the inference accuracy by making the sent data noisy. An undisciplined addition of noise can significantly reduce the accuracy of inference, rendering the service unusable. To address this challenge, this paper devises Shredder, an end-to-end framework, that, without altering the topology or the weights of a pre-trained network, learns additive noise distributions that significantly reduce the information content of communicated data while maintaining the inference accuracy. The key idea is finding the additive noise distributions by casting it as a disjoint offline learning process with a loss function that strikes a balance between accuracy and information degradation. The loss function also exposes a knob for a disciplined and controlled asymmetric trade-off between privacy and accuracy. Experimentation with six real-world DNNs from text processing and image classification shows that Shredder reduces the mutual information between the input and the communicated data to the cloud by 74.70% compared to the original execution while only sacrificing 1.58% loss in accuracy. On average, Shredder also offers a speedup of 1.79x over Wi-Fi and 2.17x over LTE compared to cloud-only execution when using an off-the-shelf mobile GPU (Tegra X2) on the edge.

Ahmed T. Elthakeb, Prannoy Pilligundla, Hadi Esmaeilzadeh

Deep quantization of neural networks (below eight bits) offers significant promise in reducing their compute and storage cost. Albeit alluring, without special techniques for training and optimization, deep quantization results in significant accuracy loss. To further mitigate this loss, we propose a novel sinusoidal regularization, called SinReQ1, for deep quantized training. SinReQ adds a periodic term to the original objective function of the underlying training algorithm. SinReQ exploits the periodicity, differentiability, and the desired convexity profile in sinusoidal functions to automatically propel weights towards values that are inherently closer to quantization levels. Since, this technique does not require invasive changes to the training procedure, SinReQ can harmoniously enhance quantized training algorithms. SinReQ offers generality and flexibility as it is not limited to a certain bitwidth or a uniform assignment of bitwidths across layers. We carry out experimentation using the AlexNet, CIFAR-10, ResNet-18, ResNet-20, SVHN, and VGG-11 DNNs with three to five bits for quantization and show the versatility of SinReQ in enhancing multiple quantized training algorithms, DoReFa [32] and WRPN [24]. Averaging across all the bit configurations shows that SinReQ closes the accuracy gap between these two techniques and the full-precision runs by 32.4% and 27.5%, respectively. That is improving the absolute accuracy of DoReFa and WRPN by 2.8% and 2.1%, respectively.

Ahmed T. Elthakeb, Prannoy Pilligundla, FatemehSadat Mireshghallah et al.

Deep Neural Networks (DNNs) typically require massive amount of computation resource in inference tasks for computer vision applications. Quantization can significantly reduce DNN computation and storage by decreasing the bitwidth of network encodings. Recent research affirms that carefully selecting the quantization levels for each layer can preserve the accuracy while pushing the bitwidth below eight bits. However, without arduous manual effort, this deep quantization can lead to significant accuracy loss, leaving it in a position of questionable utility. As such, deep quantization opens a large hyper-parameter space (bitwidth of the layers), the exploration of which is a major challenge. We propose a systematic approach to tackle this problem, by automating the process of discovering the quantization levels through an end-to-end deep reinforcement learning framework (ReLeQ). We adapt policy optimization methods to the problem of quantization, and focus on finding the best design decisions in choosing the state and action spaces, network architecture and training framework, as well as the tuning of various hyperparamters. We show how ReLeQ can balance speed and quality, and provide an asymmetric general solution for quantization of a large variety of deep networks (AlexNet, CIFAR-10, LeNet, MobileNet-V1, ResNet-20, SVHN, and VGG-11) that virtually preserves the accuracy (=< 0.3% loss) while minimizing the computation and storage cost. With these DNNs, ReLeQ enables conventional hardware to achieve 2.2x speedup over 8-bit execution. Similarly, a custom DNN accelerator achieves 2.0x speedup and energy reduction compared to 8-bit runs. These encouraging results mark ReLeQ as the initial step towards automating the deep quantization of neural networks.

Amir Yazdanbakhsh, Hajar Falahati, Philip J. Wolfe et al.

Generative Adversarial Networks (GANs) are one of the most recent deep learning models that generate synthetic data from limited genuine datasets. GANs are on the frontier as further extension of deep learning into many domains (e.g., medicine, robotics, content synthesis) requires massive sets of labeled data that is generally either unavailable or prohibitively costly to collect. Although GANs are gaining prominence in various fields, there are no accelerators for these new models. In fact, GANs leverage a new operator, called transposed convolution, that exposes unique challenges for hardware acceleration. This operator first inserts zeros within the multidimensional input, then convolves a kernel over this expanded array to add information to the embedded zeros. Even though there is a convolution stage in this operator, the inserted zeros lead to underutilization of the compute resources when a conventional convolution accelerator is employed. We propose the GANAX architecture to alleviate the sources of inefficiency associated with the acceleration of GANs using conventional convolution accelerators, making the first GAN accelerator design possible. We propose a reorganization of the output computations to allocate compute rows with similar patterns of zeros to adjacent processing engines, which also avoids inconsequential multiply-adds on the zeros. This compulsory adjacency reclaims data reuse across these neighboring processing engines, which had otherwise diminished due to the inserted zeros. The reordering breaks the full SIMD execution model, which is prominent in convolution accelerators. Therefore, we propose a unified MIMD-SIMD design for GANAX that leverages repeated patterns in the computation to create distinct microprograms that execute concurrently in SIMD mode.

Divya Mahajan, Joon Kyung Kim, Jacob Sacks et al.

The data revolution is fueled by advances in machine learning, databases, and hardware design. Programmable accelerators are making their way into each of these areas independently. As such, there is a void of solutions that enables hardware acceleration at the intersection of these disjoint fields. This paper sets out to be the initial step towards a unifying solution for in-Database Acceleration of Advanced Analytics (DAnA). Deploying specialized hardware, such as FPGAs, for in-database analytics currently requires hand-designing the hardware and manually routing the data. Instead, DAnA automatically maps a high-level specification of advanced analytics queries to an FPGA accelerator. The accelerator implementation is generated for a User Defined Function (UDF), expressed as a part of an SQL query using a Python-embedded Domain-Specific Language (DSL). To realize an efficient in-database integration, DAnA accelerators contain a novel hardware structure, Striders, that directly interface with the buffer pool of the database. Striders extract, cleanse, and process the training data tuples that are consumed by a multi-threaded FPGA engine that executes the analytics algorithm. We integrate DAnA with PostgreSQL to generate hardware accelerators for a range of real-world and synthetic datasets running diverse ML algorithms. Results show that DAnA-enhanced PostgreSQL provides, on average, 8.3x end-to-end speedup for real datasets, with a maximum of 28.2x. Moreover, DAnA-enhanced PostgreSQL is, on average, 4.0x faster than the multi-threaded Apache MADLib running on Greenplum. DAnA provides these benefits while hiding the complexity of hardware design from data scientists and allowing them to express the algorithm in =30-60 lines of Python.

Hardik Sharma, Jongse Park, Naveen Suda et al.

Fully realizing the potential of acceleration for Deep Neural Networks (DNNs) requires understanding and leveraging algorithmic properties. This paper builds upon the algorithmic insight that bitwidth of operations in DNNs can be reduced without compromising their classification accuracy. However, to prevent accuracy loss, the bitwidth varies significantly across DNNs and it may even be adjusted for each layer. Thus, a fixed-bitwidth accelerator would either offer limited benefits to accommodate the worst-case bitwidth requirements, or lead to a degradation in final accuracy. To alleviate these deficiencies, this work introduces dynamic bit-level fusion/decomposition as a new dimension in the design of DNN accelerators. We explore this dimension by designing Bit Fusion, a bit-flexible accelerator, that constitutes an array of bit-level processing elements that dynamically fuse to match the bitwidth of individual DNN layers. This flexibility in the architecture enables minimizing the computation and the communication at the finest granularity possible with no loss in accuracy. We evaluate the benefits of BitFusion using eight real-world feed-forward and recurrent DNNs. The proposed microarchitecture is implemented in Verilog and synthesized in 45 nm technology. Using the synthesis results and cycle accurate simulation, we compare the benefits of Bit Fusion to two state-of-the-art DNN accelerators, Eyeriss and Stripes. In the same area, frequency, and process technology, BitFusion offers 3.9x speedup and 5.1x energy savings over Eyeriss. Compared to Stripes, BitFusion provides 2.6x speedup and 3.9x energy reduction at 45 nm node when BitFusion area and frequency are set to those of Stripes. Scaling to GPU technology node of 16 nm, BitFusion almost matches the performance of a 250-Watt Titan Xp, which uses 8-bit vector instructions, while BitFusion merely consumes 895 milliwatts of power.