Minsoo Rhu

Maximilian Lam, Jeff Johnson, Wenjie Xiong et al.

On-device machine learning (ML) inference can enable the use of private user data on user devices without revealing them to remote servers. However, a pure on-device solution to private ML inference is impractical for many applications that rely on embedding tables that are too large to be stored on-device. In particular, recommendation models typically use multiple embedding tables each on the order of 1-10 GBs of data, making them impractical to store on-device. To overcome this barrier, we propose the use of private information retrieval (PIR) to efficiently and privately retrieve embeddings from servers without sharing any private information. As off-the-shelf PIR algorithms are usually too computationally intensive to directly use for latency-sensitive inference tasks, we 1) propose novel GPU-based acceleration of PIR, and 2) co-design PIR with the downstream ML application to obtain further speedup. Our GPU acceleration strategy improves system throughput by more than $20 \times$ over an optimized CPU PIR implementation, and our PIR-ML co-design provides an over $5 \times$ additional throughput improvement at fixed model quality. Together, for various on-device ML applications such as recommendation and language modeling, our system on a single V100 GPU can serve up to $100,000$ queries per second -- a $>100 \times$ throughput improvement over a CPU-based baseline -- while maintaining model accuracy.

Yunjae Lee, Jinha Chung, Minsoo Rhu

Graph neural networks (GNNs) can extract features by learning both the representation of each objects (i.e., graph nodes) and the relationship across different objects (i.e., the edges that connect nodes), achieving state-of-the-art performance in various graph-based tasks. Despite its strengths, utilizing these algorithms in a production environment faces several challenges as the number of graph nodes and edges amount to several billions to hundreds of billions scale, requiring substantial storage space for training. Unfortunately, state-of-the-art ML frameworks employ an in-memory processing model which significantly hampers the productivity of ML practitioners as it mandates the overall working set to fit within DRAM capacity. In this work, we first conduct a detailed characterization on a state-of-the-art, large-scale GNN training algorithm, GraphSAGE. Based on the characterization, we then explore the feasibility of utilizing capacity-optimized NVM SSDs for storing memory-hungry GNN data, which enables large-scale GNN training beyond the limits of main memory size. Given the large performance gap between DRAM and SSD, however, blindly utilizing SSDs as a direct substitute for DRAM leads to significant performance loss. We therefore develop SmartSAGE, our software/hardware co-design based on an in-storage processing (ISP) architecture. Our work demonstrates that an ISP based large-scale GNN training system can achieve both high capacity storage and high performance, opening up opportunities for ML practitioners to train large GNN datasets without being hampered by the physical limitations of main memory size.

Ranggi Hwang, Minhoo Kang, Jiwon Lee et al.

Graph convolutional neural networks (GCNs) have emerged as a key technology in various application domains where the input data is relational. A unique property of GCNs is that its two primary execution stages, aggregation and combination, exhibit drastically different dataflows. Consequently, prior GCN accelerators tackle this research space by casting the aggregation and combination stages as a series of sparse-dense matrix multiplication. However, prior work frequently suffers from inefficient data movements, leaving significant performance left on the table. We present GROW, a GCN accelerator based on Gustavson's algorithm to architect a row-wise product based sparse-dense GEMM accelerator. GROW co-designs the software/hardware that strikes a balance in locality and parallelism for GCNs, achieving significant energy-efficiency improvements vs. state-of-the-art GCN accelerators.

Jehyeon Bang, Yujeong Choi, Myeongwoo Kim et al.

As large language models (LLMs) become widespread in various application domains, a critical challenge the AI community is facing is how to train these large AI models in a cost-effective manner. Existing LLM training plans typically employ a heuristic based parallel training strategy which is based on empirical observations rather than grounded upon a thorough examination of the search space of LLM parallelization. Such limitation renders existing systems to leave significant performance left on the table, wasting millions of dollars worth of training cost. This paper presents our profiling-driven simulator called vTrain, providing AI practitioners a fast yet accurate software framework to determine an efficient and cost-effective LLM training system configuration. We demonstrate vTrain's practicality through several case studies, e.g., effectively evaluating optimal training parallelization strategies that balances training time and its associated training cost, efficient multi-tenant GPU cluster schedulers targeting multiple LLM training jobs, and determining a compute-optimal LLM model architecture given a fixed compute budget.

Youngeun Kwon, Minsoo Rhu

Personalized recommendation models (RecSys) are one of the most popular machine learning workload serviced by hyperscalers. A critical challenge of training RecSys is its high memory capacity requirements, reaching hundreds of GBs to TBs of model size. In RecSys, the so-called embedding layers account for the majority of memory usage so current systems employ a hybrid CPU-GPU design to have the large CPU memory store the memory hungry embedding layers. Unfortunately, training embeddings involve several memory bandwidth intensive operations which is at odds with the slow CPU memory, causing performance overheads. Prior work proposed to cache frequently accessed embeddings inside GPU memory as means to filter down the embedding layer traffic to CPU memory, but this paper observes several limitations with such cache design. In this work, we present a fundamentally different approach in designing embedding caches for RecSys. Our proposed ScratchPipe architecture utilizes unique properties of RecSys training to develop an embedding cache that not only sees the past but also the "future" cache accesses. ScratchPipe exploits such property to guarantee that the active working set of embedding layers can "always" be captured inside our proposed cache design, enabling embedding layer training to be conducted at GPU memory speed.

Beomsik Park, Ranggi Hwang, Dongho Yoon et al.

The widespread deployment of machine learning (ML) is raising serious concerns on protecting the privacy of users who contributed to the collection of training data. Differential privacy (DP) is rapidly gaining momentum in the industry as a practical standard for privacy protection. Despite DP's importance, however, little has been explored within the computer systems community regarding the implication of this emerging ML algorithm on system designs. In this work, we conduct a detailed workload characterization on a state-of-the-art differentially private ML training algorithm named DP-SGD. We uncover several unique properties of DP-SGD (e.g., its high memory capacity and computation requirements vs. non-private ML), root-causing its key bottlenecks. Based on our analysis, we propose an accelerator for differentially private ML named DiVa, which provides a significant improvement in compute utilization, leading to 2.6x higher energy-efficiency vs. conventional systolic arrays.

Jehyeon Bang, Eunyeong Cho, Ranggi Hwang et al.

The Mixture-of-Experts (MoE) architecture has emerged as a promising approach to mitigate the rising computational costs of large language models (LLMs) by selectively activating parameters. However, its high memory requirements and sub-optimal parameter efficiency pose significant challenges for efficient deployment. Although CPU-offloaded MoE inference systems have been proposed in the literature, they offer limited efficiency, particularly for large batch sizes. In this work, we propose SpecMoE, a memory-efficient MoE inference system based on our self-assisted speculative decoding algorithm. SpecMoE demonstrates the effectiveness of applying speculative decoding to MoE inference without requiring additional model training or fine-tuning. Our system improves inference throughput by up to $4.30\times$, while significantly reducing bandwidth requirements of both memory and interconnect on memory-constrained systems.

Hyeseong Kim, Gwangoo Yeo, Minsoo Rhu

GPU-initiated I/O has emerged as a key mechanism for achieving high-throughput storage access by leveraging massive GPU thread-level parallelism, while recent industry trends point toward SSDs optimized for ultra-high random-read IOPS. Together, these trends are enabling the emergence of IOPS-optimized, GPU-centric storage systems. Despite this momentum, no existing framework enables quantitative end-to-end evaluation of storage systems optimized for GPU-initiated I/O. While conventional SSD emulators provide a promising path toward end-to-end modeling in traditional storage systems, they face three key challenges in this GPU-centric setting: limited frontend scalability for ingesting massive request streams, high software overhead in emulating GPU-initiated I/O control and data paths, and excessive timing-model maintenance overhead at extremely high I/O request rates. We propose SwarmIO, an SSD emulator for massively parallel, GPU-centric storage. SwarmIO faithfully models IOPS-optimized SSDs at target performance levels of up to 40 MIOPS, achieving a 303.9x speedup over the state-of-the-art baseline SSD emulator under GPU-initiated I/O. We further demonstrate its utility through a vector search case study, showing that increasing SSD IOPS from 2.5 MIOPS to 40 MIOPS yields an average end-to-end speedup of up to 9.7x.

Yujeong Choi, John Kim, Minsoo Rhu

While providing low latency is a fundamental requirement in deploying recommendation services, achieving high resource utility is also crucial in cost-effectively maintaining the datacenter. Co-locating multiple workers of a model is an effective way to maximize query-level parallelism and server throughput, but the interference caused by concurrent workers at shared resources can prevent server queries from meeting its SLA. Hera utilizes the heterogeneous memory requirement of multi-tenant recommendation models to intelligently determine a productive set of co-located models and its resource allocation, providing fast response time while achieving high throughput. We show that Hera achieves an average 37.3% improvement in effective machine utilization, enabling 26% reduction in required servers, significantly improving upon the baseline recommedation inference server.

Eunyeong Cho, Jehyeon Bang, Ranggi Hwang et al.

The emergence of reasoning-based LLMs leveraging Chain-of-Thought (CoT) inference introduces new serving challenges, as their extended reasoning phases delay user-visible output and inflate Time-To-First-Token (TTFT). Existing LLM serving frameworks fail to distinguish between reasoning and answering phases, leading to performance degradation under GPU memory constraints. We present PASCAL, a phase-aware scheduling algorithm that prioritizes reasoning to reduce TTFT while using controlled preemption and token pacing during answering to preserve Quality-of-Experience (QoE). Our hierarchical scheduler combines instance-level placement with intra-instance execution and enables dynamic migration at phase boundaries to balance load and reduce interference. Across benchmarks using DeepSeek-R1-Distill-Qwen-32B, PASCAL reduces tail TTFT by up to 72% while maintaining answering phase SLO attainment, demonstrating the importance of phase-aware scheduling for reasoning-based LLM deployment.

Jinha Chung, Byeongjun Shin, Jiin Kim et al.

LLM-based agents deliver state-of-the-art performance across tasks but incur high end-to-end latency on edge devices. We introduce Agent-X, a software-only, accuracy-preserving framework that accelerates both the prefill and decode stages of on-device agent workloads. Agent-X's two key components rewrite prompts to leverage prefix caching tailored to agent-specific input-token patterns and enable LLM-free speculative decoding for fast token generation with minimal overhead. On representative agentic workloads, Agent-X achieves a 1.61x end-to-end speedup in real systems with no accuracy loss and can be seamlessly integrated into existing on-device AI agents. To the best of our knowledge, ours is the first to systematically characterize and eliminate latency bottlenecks in on-device agents.

Jiin Kim, Byeongjun Shin, Jinha Chung et al.

Large-language-model (LLM)-based AI agents have recently showcased impressive versatility by employing dynamic reasoning, an adaptive, multi-step process that coordinates with external tools. This shift from static, single-turn inference to agentic, multi-turn workflows broadens task generalization and behavioral flexibility, but it also introduces serious concerns about system-level cost, efficiency, and sustainability. This paper presents the first comprehensive system-level analysis of AI agents, quantifying their resource usage, latency behavior, energy consumption, and datacenter-wide power consumption demands across diverse agent designs and test-time scaling strategies. We further characterize how AI agent design choices, such as few-shot prompting, reflection depth, and parallel reasoning, impact accuracy-cost tradeoffs. Our findings reveal that while agents improve accuracy with increased compute, they suffer from rapidly diminishing returns, widening latency variance, and unsustainable infrastructure costs. Through detailed evaluation of representative agents, we highlight the profound computational demands introduced by AI agent workflows, uncovering a looming sustainability crisis. These results call for a paradigm shift in agent design toward compute-efficient reasoning, balancing performance with deployability under real-world constraints.

Yunjae Lee, Juntaek Lim, Jehyeon Bang et al.

This paper presents a comprehensive evaluation of Intel Gaudi NPUs as an alternative to NVIDIA GPUs, which is currently the de facto standard in AI system design. First, we create a suite of microbenchmarks to compare Intel Gaudi-2 with NVIDIA A100, showing that Gaudi-2 achieves competitive performance not only in primitive AI compute, memory, and communication operations but also in executing several important AI workloads end-to-end. We then assess Gaudi NPU's programmability by discussing several software-level optimization strategies to employ for implementing critical FBGEMM operators and vLLM, evaluating their efficiency against GPU-optimized counterparts. Results indicate that Gaudi-2 achieves energy efficiency comparable to A100, though there are notable areas for improvement in terms of software maturity. Overall, we conclude that, with effective integration into high-level AI frameworks, Gaudi NPUs could challenge NVIDIA GPU's dominance in the AI server market, though further improvements are necessary to fully compete with NVIDIA's robust software ecosystem.

Gwangoo Yeo, Jiin Kim, Yujeong Choi et al.

NVIDIA's Multi-Instance GPU (MIG) is a feature that enables system designers to reconfigure one large GPU into multiple smaller GPU slices. This work characterizes this emerging GPU and evaluates its effectiveness in designing high-performance AI inference servers. Our study reveals that the data preprocessing stage of AI inference causes significant performance bottlenecks to MIG. To this end, we present PREBA, which is a hardware/software co-design targeting MIG inference servers. Our first proposition is an FPGA-based data preprocessing accelerator that unlocks the full potential of MIG with domain-specific acceleration of data preprocessing. The MIG inference server unleashed from preprocessing overheads is then augmented with our dynamic batching system that enables high-performance inference. PREBA is implemented end-to-end in real systems, providing a 3.7x improvement in throughput, 3.4x reduction in tail latency, 3.5x improvement in energy-efficiency, and 3.0x improvement in cost-efficiency.

Juntaek Lim, Youngeun Kwon, Ranggi Hwang et al.

Differential privacy (DP) is widely being employed in the industry as a practical standard for privacy protection. While private training of computer vision or natural language processing applications has been studied extensively, the computational challenges of training of recommender systems (RecSys) with DP have not been explored. In this work, we first present our detailed characterization of private RecSys training using DP-SGD, root-causing its several performance bottlenecks. Specifically, we identify DP-SGD's noise sampling and noisy gradient update stage to suffer from a severe compute and memory bandwidth limitation, respectively, causing significant performance overhead in training private RecSys. Based on these findings, we propose LazyDP, an algorithm-software co-design that addresses the compute and memory challenges of training RecSys with DP-SGD. Compared to a state-of-the-art DP-SGD training system, we demonstrate that LazyDP provides an average 119x training throughput improvement while also ensuring mathematically equivalent, differentially private RecSys models to be trained.

Yunjae Lee, Hyeseong Kim, Minsoo Rhu

Training recommendation systems (RecSys) faces several challenges as it requires the "data preprocessing" stage to preprocess an ample amount of raw data and feed them to the GPU for training in a seamless manner. To sustain high training throughput, state-of-the-art solutions reserve a large fleet of CPU servers for preprocessing which incurs substantial deployment cost and power consumption. Our characterization reveals that prior CPU-centric preprocessing is bottlenecked on feature generation and feature normalization operations as it fails to reap out the abundant inter-/intra-feature parallelism in RecSys preprocessing. PreSto is a storage-centric preprocessing system leveraging In-Storage Processing (ISP), which offloads the bottlenecked preprocessing operations to our ISP units. We show that PreSto outperforms the baseline CPU-centric system with a $9.6\times$ speedup in end-to-end preprocessing time, $4.3\times$ enhancement in cost-efficiency, and $11.3\times$ improvement in energyefficiency on average for production-scale RecSys preprocessing.

Yujeong Choi, Jiin Kim, Minsoo Rhu

With the increasing popularity of recommendation systems (RecSys), the demand for compute resources in datacenters has surged. However, the model-wise resource allocation employed in current RecSys model serving architectures falls short in effectively utilizing resources, leading to sub-optimal total cost of ownership. We propose ElasticRec, a model serving architecture for RecSys providing resource elasticity and high memory efficiency. ElasticRec is based on a microservice-based software architecture for fine-grained resource allocation, tailored to the heterogeneous resource demands of RecSys. Additionally, ElasticRec achieves high memory efficiency via our utility-based resource allocation. Overall, ElasticRec achieves an average 3.3x reduction in memory allocation size and 8.1x increase in memory utility, resulting in an average 1.6x reduction in deployment cost compared to state-of-the-art RecSys inference serving system.

Yunseong Kim, Yujeong Choi, Minsoo Rhu

In cloud machine learning (ML) inference systems, providing low latency to end-users is of utmost importance. However, maximizing server utilization and system throughput is also crucial for ML service providers as it helps lower the total-cost-of-ownership. GPUs have oftentimes been criticized for ML inference usages as its massive compute and memory throughput is hard to be fully utilized under low-batch inference scenarios. To address such limitation, NVIDIA's recently announced Ampere GPU architecture provides features to "reconfigure" one large, monolithic GPU into multiple smaller "GPU partitions". Such feature provides cloud ML service providers the ability to utilize the reconfigurable GPU not only for large-batch training but also for small-batch inference with the potential to achieve high resource utilization. In this paper, we study this emerging GPU architecture with reconfigurability to develop a high-performance multi-GPU ML inference server. Our first proposition is a sophisticated partitioning algorithm for reconfigurable GPUs that systematically determines a heterogeneous set of multi-granular GPU partitions, best suited for the inference server's deployment. Furthermore, we co-design an elastic scheduling algorithm tailored for our heterogeneously partitioned GPU server which effectively balances low latency and high GPU utilization.

Sangpyo Kim, Jongmin Kim, Michael Jaemin Kim et al.

Homomorphic encryption (HE) enables the secure offloading of computations to the cloud by providing computation on encrypted data (ciphertexts). HE is based on noisy encryption schemes in which noise accumulates as more computations are applied to the data. The limited number of operations applicable to the data prevents practical applications from exploiting HE. Bootstrapping enables an unlimited number of operations or fully HE (FHE) by refreshing the ciphertext. Unfortunately, bootstrapping requires a significant amount of additional computation and memory bandwidth as well. Prior works have proposed hardware accelerators for computation primitives of FHE. However, to the best of our knowledge, this is the first to propose a hardware FHE accelerator that supports bootstrapping as a first-class citizen. In particular, we propose BTS - Bootstrappable, Technologydriven, Secure accelerator architecture for FHE. We identify the challenges of supporting bootstrapping in the accelerator and analyze the off-chip memory bandwidth and computation required. In particular, given the limitations of modern memory technology, we identify the HE parameter sets that are efficient for FHE acceleration. Based on the insights gained from our analysis, we propose BTS, which effectively exploits the parallelism innate in HE operations by arranging a massive number of processing elements in a grid. We present the design and microarchitecture of BTS, including a network-on-chip design that exploits a deterministic communication pattern. BTS shows 5,556x and 1,306x improved execution time on ResNet-20 and logistic regression over a CPU, with a chip area of 373.6mm^2 and up to 163.2W of power.

Yujeong Choi, Yunseong Kim, Minsoo Rhu

In cloud ML inference systems, batching is an essential technique to increase throughput which helps optimize total-cost-of-ownership. Prior graph batching combines the individual DNN graphs into a single one, allowing multiple inputs to be concurrently executed in parallel. We observe that the coarse-grained graph batching becomes suboptimal in effectively handling the dynamic inference request traffic, leaving significant performance left on the table. This paper proposes LazyBatching, an SLA-aware batching system that considers both scheduling and batching in the granularity of individual graph nodes, rather than the entire graph for flexible batching. We show that LazyBatching can intelligently determine the set of nodes that can be efficiently batched together, achieving an average 15x, 1.5x, and 5.5x improvement than graph batching in terms of average response time, throughput, and SLA satisfaction, respectively.

Youngeun Kwon, Yunjae Lee, Minsoo Rhu

Personalized recommendations are one of the most widely deployed machine learning (ML) workload serviced from cloud datacenters. As such, architectural solutions for high-performance recommendation inference have recently been the target of several prior literatures. Unfortunately, little have been explored and understood regarding the training side of this emerging ML workload. In this paper, we first perform a detailed workload characterization study on training recommendations, root-causing sparse embedding layer training as one of the most significant performance bottlenecks. We then propose our algorithm-architecture co-design called Tensor Casting, which enables the development of a generic accelerator architecture for tensor gather-scatter that encompasses all the key primitives of training embedding layers. When prototyped on a real CPU-GPU system, Tensor Casting provides 1.9-21x improvements in training throughput compared to state-of-the-art approaches.

Ranggi Hwang, Taehun Kim, Youngeun Kwon et al.

Personalized recommendations are the backbone machine learning (ML) algorithm that powers several important application domains (e.g., ads, e-commerce, etc) serviced from cloud datacenters. Sparse embedding layers are a crucial building block in designing recommendations yet little attention has been paid in properly accelerating this important ML algorithm. This paper first provides a detailed workload characterization on personalized recommendations and identifies two significant performance limiters: memory-intensive embedding layers and compute-intensive multi-layer perceptron (MLP) layers. We then present Centaur, a chiplet-based hybrid sparse-dense accelerator that addresses both the memory throughput challenges of embedding layers and the compute limitations of MLP layers. We implement and demonstrate our proposal on an Intel HARPv2, a package-integrated CPU+FPGA device, which shows a 1.7-17.2x performance speedup and 1.7-19.5x energy-efficiency improvement than conventional approaches.

Bongjoon Hyun, Youngeun Kwon, Yujeong Choi et al.

To satisfy the compute and memory demands of deep neural networks, neural processing units (NPUs) are widely being utilized for accelerating deep learning algorithms. Similar to how GPUs have evolved from a slave device into a mainstream processor architecture, it is likely that NPUs will become first class citizens in this fast-evolving heterogeneous architecture space. This paper makes a case for enabling address translation in NPUs to decouple the virtual and physical memory address space. Through a careful data-driven application characterization study, we root-cause several limitations of prior GPU-centric address translation schemes and propose a memory management unit (MMU) that is tailored for NPUs. Compared to an oracular MMU design point, our proposal incurs only an average 0.06% performance overhead.

Yujeong Choi, Minsoo Rhu

To amortize cost, cloud vendors providing DNN acceleration as a service to end-users employ consolidation and virtualization to share the underlying resources among multiple DNN service requests. This paper makes a case for a "preemptible" neural processing unit (NPU) and a "predictive" multi-task scheduler to meet the latency demands of high-priority inference while maintaining high throughput. We evaluate both the mechanisms that enable NPUs to be preemptible and the policies that utilize them to meet scheduling objectives. We show that preemptive NPU multi-tasking can achieve an average 7.8x, 1.4x, and 4.8x improvement in latency, throughput, and SLA satisfaction, respectively.

Youngeun Kwon, Yunjae Lee, Minsoo Rhu

Recent studies from several hyperscalars pinpoint to embedding layers as the most memory-intensive deep learning (DL) algorithm being deployed in today's datacenters. This paper addresses the memory capacity and bandwidth challenges of embedding layers and the associated tensor operations. We present our vertically integrated hardware/software co-design, which includes a custom DIMM module enhanced with near-data processing cores tailored for DL tensor operations. These custom DIMMs are populated inside a GPU-centric system interconnect as a remote memory pool, allowing GPUs to utilize for scalable memory bandwidth and capacity expansion. A prototype implementation of our proposal on real DL systems shows an average 6.2-17.6x performance improvement on state-of-the-art recommender systems.

Youngeun Kwon, Minsoo Rhu

As the models and the datasets to train deep learning (DL) models scale, system architects are faced with new challenges, one of which is the memory capacity bottleneck, where the limited physical memory inside the accelerator device constrains the algorithm that can be studied. We propose a memory-centric deep learning system that can transparently expand the memory capacity available to the accelerators while also providing fast inter-device communication for parallel training. Our proposal aggregates a pool of memory modules locally within the device-side interconnect, which are decoupled from the host interface and function as a vehicle for transparent memory capacity expansion. Compared to conventional systems, our proposal achieves an average 2.8x speedup on eight DL applications and increases the system-wide memory capacity to tens of TBs.

Maohua Zhu, Jason Clemons, Jeff Pool et al.

Exploiting sparsity enables hardware systems to run neural networks faster and more energy-efficiently. However, most prior sparsity-centric optimization techniques only accelerate the forward pass of neural networks and usually require an even longer training process with iterative pruning and retraining. We observe that artificially inducing sparsity in the gradients of the gates in an LSTM cell has little impact on the training quality. Further, we can enforce structured sparsity in the gate gradients to make the LSTM backward pass up to 45% faster than the state-of-the-art dense approach and 168% faster than the state-of-the-art sparsifying method on modern GPUs. Though the structured sparsifying method can impact the accuracy of a model, this performance gap can be eliminated by mixing our sparse training method and the standard dense training method. Experimental results show that the mixed method can achieve comparable results in a shorter time span than using purely dense training.

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.

Minsoo Rhu, Mike O'Connor, Niladrish Chatterjee et al.

Popular deep learning frameworks require users to fine-tune their memory usage so that the training data of a deep neural network (DNN) fits within the GPU physical memory. Prior work tries to address this restriction by virtualizing the memory usage of DNNs, enabling both CPU and GPU memory to be utilized for memory allocations. Despite its merits, virtualizing memory can incur significant performance overheads when the time needed to copy data back and forth from CPU memory is higher than the latency to perform the computations required for DNN forward and backward propagation. We introduce a high-performance virtualization strategy based on a "compressing DMA engine" (cDMA) that drastically reduces the size of the data structures that are targeted for CPU-side allocations. The cDMA engine offers an average 2.6x (maximum 13.8x) compression ratio by exploiting the sparsity inherent in offloaded data, improving the performance of virtualized DNNs by an average 32% (maximum 61%).

Minsoo Rhu, Natalia Gimelshein, Jason Clemons et al.

The most widely used machine learning frameworks require users to carefully tune their memory usage so that the deep neural network (DNN) fits into the DRAM capacity of a GPU. This restriction hampers a researcher's flexibility to study different machine learning algorithms, forcing them to either use a less desirable network architecture or parallelize the processing across multiple GPUs. We propose a runtime memory manager that virtualizes the memory usage of DNNs such that both GPU and CPU memory can simultaneously be utilized for training larger DNNs. Our virtualized DNN (vDNN) reduces the average GPU memory usage of AlexNet by up to 89%, OverFeat by 91%, and GoogLeNet by 95%, a significant reduction in memory requirements of DNNs. Similar experiments on VGG-16, one of the deepest and memory hungry DNNs to date, demonstrate the memory-efficiency of our proposal. vDNN enables VGG-16 with batch size 256 (requiring 28 GB of memory) to be trained on a single NVIDIA Titan X GPU card containing 12 GB of memory, with 18% performance loss compared to a hypothetical, oracular GPU with enough memory to hold the entire DNN.