Zhengyi Li, Yakai Wang, Jingwen Leng, Kang Yang, Yu Yu, Jiaping Gui, Yu Feng, Ning Liu, Minyi Guo. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Xing Ma, Yangjie Zhou, Wu Sun, Zihan Liu, Jingwen Leng, Yun Lin, Shixuan Sun, Minyi Guo, Jin Song Dong. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Existing low-bit Microscaling (MX) formats, such as MXFP4, often suffer from substantial accuracy degradation due to the use of a shared scaling factor with the Power-of-Two format. In this work, we explore strategies that introduce minimal metadata to recover accuracy lost during quantization while maintaining high bit efficiency across a wide range of large language models. We propose a complete algorithm-hardware co-design based on flexible metadata, featuring an online quantization with simple encoding. To support the proposed method efficiently, we implement a lightweight hardware unit and integrate it into the accelerator. Evaluation results demonstrate that our method substantially narrows the accuracy gap, achieving on average a 70.63% reduction in accuracy loss compared to MXFP4 and a 37.30% reduction relative to the latest NVFP4 on LLM benchmarks. Furthermore, our design delivers up to 1.91× speedup and 1.75× energy savings over state-of-the-art accelerators.

3D Gaussian splatting (3DGS) has drawn significant attention in the architectural community recently. However, current architectural designs often overlook the 3DGS scalability, making them fragile for extremely large-scale 3DGS. Meanwhile, the VR bandwidth requirement makes it impossible to deliver high-fidelity and smooth VR content from the cloud.

The scaling of computation throughput continues to outpace improvements in memory bandwidth, making many deep learning workloads memory-bound. Kernel fusion is a key technique to alleviate this problem, but the fusion strategies of existing compilers and frameworks are limited to using local scratchpad memory. When the intermediate results exceed the limited capacity (such as FFN), the fusion fails. Although modern GPUs (like the NVIDIA H100) now incorporate an inter-core connection mechanism known as Distributed Shared Memory (DSM)—providing a larger, high-bandwidth, and low-latency on-chip memory pool—this hardware potential has yet to be exploited by software frameworks. To bridge this gap, we present FlashFuser, the first compiler framework to utilize inter-core connection for kernel fusion on modern GPUs. FlashFuser extends established fusion techniques to the DSM domain through three core contributions. First, we propose a powerful DSM-based communication abstraction that formalizes complex cluster-based data exchange patterns, such as reduce, shuffle and multiply. Second, we introduce a dataflow analyzer that generalizes loop scheduling, resource mapping, and tile selection to the distributed memory hierarchy; it determines the optimal execution order and tile sizes by quantifying data movement across memory levels. Finally, FlashFuser integrates these components into a unified search engine that employs analytical cost modeling and DSM-aware pruning strategies to efficiently discover the optimal execution plan. Our evaluation on an NVIDIA H100 GPU shows that FlashFuser reduces memory access by 58 % and delivers kernel speedups of$3.3 x$against highly-tuned libraries and 4.1x against state-of-the-art compilers, resulting in a$1.24 \times$end-to-end speedup.

3D Gaussian splatting (3DGS) has emerged as a promising direction for SLAM due to its high-fidelity reconstruction and rapid convergence. However, 3DGS-SLAM algorithms remain impractical for mobile platforms due to their high computational cost, especially for their tracking process. This work introduces Splatonic, a sparse and efficient realtime 3DGS-SLAM algorithm-hardware co-design for resourceconstrained devices. Inspired by classical SLAMs, we propose an adaptive sparse pixel sampling algorithm that reduces the number of rendered pixels by up to$256 \times$while retaining accuracy. To unlock this performance potential on mobile GPUs, we design a novel pixel-based rendering pipeline that improves hardware utilization via Gaussian-parallel rendering and preemptive$\alpha$-checking. Together, these optimizations yield up to$121.7 \times$speedup on the bottleneck stages and$14.6 \times$end-toend speedup on off-the-shelf GPUs. To further address new bottlenecks introduced by our rendering pipeline, we propose a pipelined architecture that simplifies the overall design while addressing newly emerged bottlenecks in projection and aggregation. Evaluated across four 3DGS-SLAM algorithms, Splatonic achieves up to$274.9 \times$speedup and$4738.5 \times$energy savings over mobile GPUs and up to$25.2 \times$speedup and$241.1 \times$energy savings over state-of-the-art accelerators, all with comparable accuracy.

Important applications like quantum search and quantum simulation rely on control flow implemented by oracle circuits, which may account for up to 99% of the total gate count [1] as oracle circuits consist of expensive multi-controlled (MC) quantum operations that formulate the control flow. Existing efforts optimize them by simplifying Boolean expressions associated with MC gates. They merge MC gates, if associated Boolean expressions are adjacent, to reduce the gate cost. Being more aggressive, our work introduces control line encoding (CLE), a technique that even allows merging non-adjacent MC gates by modifying their control (line) polarity. We further develop a plug-in compiler framework that efficiently applies this technique. The framework is grounded in the following observations: 1) Control flows often commute with one another and can be globally reorganized to maximize their merging by CLE; 2) numerous CLE schemes exist to allow gate merge and we can co-design CLE applications to diminish the overhead of CLE itself. Experiments demonstrate that, compared to the baseline, our framework further reduces the CX and T cost, on average by 54.7% and 56.8%, respectively.

Tensor parallelism (TP) in large-scale LLM inference and training introduces frequent collective operations that dominate inter-GPU communication. While in-switch computing, exemplified by NVLink SHARP (NVLS), accelerates collective operations by reducing redundant data transfer, its communication-centric design philosophy introduces the mismatch between its communication mode and the memory semantic requirement of LLM's computation kernel. Such a mismatch isolates the compute and communication phases, resulting in underutilized resources and limited overlap in multi-GPU systems. To address the limitation, we propose CAIS, the first ComputeAware In-Switch computing framework that aligns communication modes with computation's memory semantics requirement. CAIS consists of three integral techniques: (1) compute-aware ISA and microarchitecture extension to enable compute-aware in-switch computing. (2) merging-aware TB (Thread Block) coordination to improve the temporal alignment for efficient request merging. (3) graph-level dataflow optimizer to achieve a tight cross-kernel overlap. Evaluations on LLM workloads show that CAIS achieves$1.38 \times$average end-to-end training speedup over the SOTA NVLS-enabled solution, and$1.61 \times$over T3, the SOTA compute-communicate overlap solutions but do not leverage NVLS, demonstrating its effectiveness in accelerating TP on multi-GPU systems.

Zero-knowledge succinct non-interactive arguments of knowledge (zkSNARKs) are a key technology to privacy-preserving applications today. The complexity of proof generation, however, heavily constrains throughput in latency-sensitive environments. The computational burden primarily stems from two fundamental algorithms: Multi-Scalar Multiplication (MSM) and the Number Theoretic Transform (NTT). We propose a series of optimizations for these two kernels, including computation-transfer pipelining, load balancing, and memory access fusion, achieving 1.97 × to 2.16 × proof generation speedup over a state-of-the-art open source GPU acceleration library. Our design also supports out-of-core computation, enabling the generation of large-scale ZKP proofs.

Mixture-of-Experts (MoE) has been adopted by many leading large models to reduce computational requirements. However, frequent inter-GPU communication in MoE expert parallelism (EP) becomes a performance challenge. We observe substantial redundant inter-GPU data transfers in MoE that can be potentially addressed by in-switch computing. Unfortunately, the existing solution, NVLink SHARP (NVLS), can only support static collectives with regular patterns, incapable of dynamic communication with irregular patterns in MoE. To bridge the functionality gap, we propose DySHARP, an integral dynamic in-switch computing solution to accelerate MoE, encompassing both communication primitives and communication-aware scheduling: 1) Dynamic multimem addressing co-designs ISA, architecture, and runtime, as a dynamic extension to NVLS, reducing redundant traffic. However, the resulting traffic reduction is inherently asymmetric between two directions, preventing it from directly translating into speedup. 2) Token-centric kernel fusion deeply fuses the dispatch-computation-combine pipeline, resolving this asymmetry to translate traffic reduction into actual speedup. Compared with the state-of-the-art solution, DySHARP achieves up to 1.79× speedup.

In the current NISQ era, the performance of QNN models is strictly hindered by the limited qubit number and inevitable noise. A natural idea to improve the robustness of QNN is the implementation of a distributed system. Nevertheless, due to the heterogeneity and instability of quantum chips (e.g., noise, frequent online/offline), training and inference on distributed quantum devices may even destroy the accuracy. In this paper, we propose HeteroQNN, a comprehensive QNN framework designed for efficient and high-accuracy distributed training and inference. The main innovation of HeteroQNN is it decouples the QNN circuit into two uniform representations: model vector and behavioral vector. The model vector specifies the gate parameters in the QNN model, while the behavioral vector captures the hardware features when implementing the QNN circuit. To handle the architectural heterogeneity, we introduce personalized QNN models in each QPU and share the gradient among QPUs with homogeneous behavioral vectors. We propose shot-oriented distributed inference, which is much more fine-grained scheduling that can improve accuracy and balance the workload. Finally, by leveraging the hidden homogeneity in the model vector, we present the maintenance for QPU variability. The experiments show that accelerates the training process by 4.03× with 7.87% loss reduction, compared with the previous distributed QNN framework.

Graph Neural Networks (GNNs) have achieved remarkable success across domains such as recommendation and scientific computing, yet their practical deployment remains constrained by high execution cost. The diversity of GNN model structures and the sparsity of real-world graphs pose two fundamental challenges for hardware acceleration: supporting heterogeneous operator patterns and achieving high resource utilization under irregular data access. Existing accelerators often address only one aspect, either targeting specific models with hardwired pipelines or applying general architectures with limited efficiency. To address these challenges, we propose SWITCHBLADE, a full-stack framework for GNN acceleration through the coordinated design of partitioning, compilation, and architecture. SWITCHBLADE addresses these challenges through three key components. First, a phase-based intermediate representation unifies diverse GNN models by abstracting computation stages for model-independent code generation. Second, a fine-grained graph partitioner enhances data locality and reduces memory traffic by adapting to graph topology and model semantics. Third, the hardware architecture supports stream-level parallelism and decoupled execution to exploit cross-shard and inter-phase concurrency. Evaluation on representative models and datasets shows that SWITCHBLADE achieves up to 1.85× speedup and 19.03× energy savings over an NVIDIA V100 GPU, while outperforming state-of-the-art GNN accelerators across diverse full-graph workloads, demonstrating both high efficiency and broad model generality.

RAG systems rely on rerankers to identify relevant documents. However, fine-tuning these models remains challenging due to the scarcity of annotated query-document pairs. Existing distillation-based approaches suffer from training-inference misalignment and fail to capture interdependencies among candidate documents. To overcome these limitations, we reframe the reranking process as an attention-mask problem and propose Gumbel Reranking, an end-to-end training framework for rerankers aimed at minimizing the training-inference gap. In our approach, reranker optimization is reformulated as learning a stochastic, document-wise Top-k attention mask using the Gumbel Trick and Relaxed Top-k Sampling. This formulation enables end-to-end optimization by minimizing the overall language loss. Experiments across various settings consistently demonstrate performance gains, including a 10.4% improvement in recall on HotpotQA for distinguishing indirectly relevant documents.

Graph neural networks (GNNs) are gaining popularity in diverse application domains and growing in complexity.As a result, it is crucial to achieve high-performance GNN execution.Among various techniques, algebraic transformations, including operator reordering and operator fusion, have been successfully applied to improve the computation and memory access efficiencies of DNN models.However,

Point clouds are increasingly important in intelligent applications, but frequent off-chip memory traffic in accelerators causes pipeline stalls and leads to high energy consumption. While conventional line buffer techniques can eliminate off-chip traffic, they cannot be directly applied to point clouds due to their inherent computation patterns. To address this, we introduce two techniques: compulsory splitting and deterministic termination, enabling fully-streaming processing. We further propose StreamGrid, a framework that integrates these techniques and automatically optimizes on-chip buffer sizes. Our evaluation shows StreamGrid reduces on-chip memory by 61.3% and energy consumption by 40.5% with marginal accuracy loss compared to the baselines without our techniques. Additionally, we achieve 10.0× speedup and 3.9× energy efficiency over state-of-the-art accelerators.

In the current NISQ era, the performance of QNN models is strictly hindered by the limited qubit number and inevitable noise. A natural idea to improve the robustness of QNN is to involve multiple quantum devices. Nevertheless, due to the heterogeneity and instability of quantum devices (e.g., noise, frequent online/offline), training and inference on distributed quantum devices may even destroy the accuracy. In this paper, we propose ArbiterQ, a comprehensive QNN framework designed for efficient and high-accuracy training and inference on heterogeneous QPUs. The main innovation of ArbiterQ is it applies personalized models for each QPU via two uniform QNN representations: model vector and behavioral vector. The model vector specifies the logical-level parameters in the QNN model, while the behavioral vector captures the hardware-level features when implementing the QNN circuit. In this manner, by sharing the gradient among QPUs with similar behavioral vectors, we can effectively leverage parallelism while considering heterogeneity. We also propose shot-oriented inference scheduling, which is a much more fine-grained scheduling that can improve accuracy and balance the workload. The experiments show that ArbiterQ accelerates the training process by $4.03 \times$ with $7.87 \%$ loss reduction, compared with the previous distributed QNN framework EQC [1].

Vector quantization (VQ), which treats a vector as a compression unit, gains increasing research interests for its potential to accelerate large language models (LLMs). Compared to conventional element-wise quantization methods, VQ algorithms can compress weight and KV cache tensors in LLMs with a greater ratio while maintaining the high model accuracy. However, translating a VQ algorithm’s memory reduction into the actual latency improvement is challenging. We profile and analyze the current approach of integrating VQ into computation kernels and show that its major inefficiency lies in the poor access efficiency of codebooks in VQ algorithms and uncoordinated computation dataflow. Meanwhile, the diversity of VQ algorithms (e.g., different vector sizes and entry counts) and LLMs, computation kernels (e.g matrix-matrix/vector multiplication and attention computation) makes it impractical to manually craft efficient kernel implementations for each specific case. In this work, we design and implement VQ-LLM, an efficient fused VQ kernel generation framework. We first introduce a software abstraction called codebook cache to optimize codebook access efficiency and support the integration of VQ with various computations. The codebook cache adaptively stores different entries across the GPU’s memory hierarchy, including off-chip global memory, on-chip shared memory, and registers. Centered around the codebook cache, we design an efficient computation engine that optimizes memory traffic during computations involving codebooks. This compute engine adopts the codebook-centric dataflow and fusion optimizations. Additionally, we provide adaptive heuristics to tailor parameter selection in our optimizations to diverse VQ configurations. Our optimizations achieve the latency reduction of $\mathbf{6 4. 3 6 \%}$ to $\mathbf{9 9. 1 \%}$ compared to existing open-source implementations. A final comparison with state-of-the-art element-wise quantization methods like AWQ and QoQ shows that our VQ-LLM is practically viable, achieving latencies close or even better latencies to those at equivalent bit-widths, potentially offering greater accuracy.

Large language models (LLMs) are one of the most important killer computer applications. The recent algorithmic advancement proposes a fine-grained group-wise quantization for LLMs, which treats a small set (e.g., 64) of values in a tensor as a compression unit. It effectively preserves the model accuracy without retraining, and has become the standard approach to efficiently deploy LLMs. On the other hand, there are works that propose various adaptive data types to better adapt to different distributions and further reduce the required bit length for LLMs. In this work, our detailed analysis unveils a key finding that while different tensors exhibit similar distributions, small groups can have markedly different distributions. As such, the group-level diversity requires a new level of adaptivity for which existing adaptive data types fail to provide.In this paper, we propose MANT, a mathematically adaptive numeric type, featuring a more flexible encoding paradigm with a wider range of data distribution and more efficient decoding-computation fusion mechanism to address these challenges. Based on MANT, we develop a supporting framework to assign the appropriate data type for each group adaptively. Meanwhile, the dynamically generated Key-Value (KV) caches in LLMs introduce further complexity for real-time quantization. To tackle this, we propose an efficient real-time quantization mechanism. Besides, we implement a specific processing element (PE) to efficiently support MANT and incorporate a real-time quantization unit. By integrating these components into a systolic array, MANT unifies the group-wise weight and KV cache quantization and addresses the associated challenges. Our evaluation shows achieving, on average, 2.99 × (up to 4.46 ×) speedup and 2.81 × (up to 4.10 ×) energy reduction to the state-of-the-art LLM accelerator.

Large Language Models (LLMs) use key-value (KV) cache to reduce redundant computation in autoregressive generation. However, the KV cache size increases linearly during generation, leading to excessive memory usage, especially for long texts. Most KV cache compression methods evict the unimportant KV pairs to maintain a fixed cache size, which leads to the permanent loss of tokens during generation. However, singular value decomposition shows that values do not exhibit a strong low-rank property as keys do, suggesting that information is distributed more evenly across values, in contrast to its more redundant distribution within keys. Therefore, methods that evict both keys and values risk losing crucial information and compromise context integrity, ultimately degrading the output quality. To address this problem, we propose WeightedKV, a novel, training-free approach that discards the keys of less important tokens, while merging their values into neighboring tokens via a convex combination weighted by their average attention scores. In this way, the retained keys serve as anchors that guide the generation process, while the merged values provide a rich contextual backdrop. We assess our method on four widely used language modeling datasets, demonstrating superior performance compared to all baseline methods, particularly with a lower budget ratio.

Rendering is critical in fields like 3D modeling, AR/VR, and autonomous driving, where high-quality, real-time output is essential. Point-based neural rendering (PBNR) offers a photorealistic and efficient alternative to conventional methods, yet it is still challenging to achieve real-time rendering on mobile platforms. We pinpoint two major bottlenecks in PBNR pipelines: LoD search and splatting. LoD search suffers from workload imbalance and irregular memory access, making it inefficient on off-the-shelf GPUs. Meanwhile, splatting introduces severe warp divergence across GPU threads due to its inherent sparsity.To tackle these challenges, we propose SLTarch, an algorithm-architecture co-designed framework. At its core, SLTarch introduces SLTree, a dedicated subtree-based data structure, and LTcore, a specialized hardware architecture tailored for efficient LoD search. Additionally, we co-design a divergence-free splatting algorithm with our simple yet principled hardware augmentation, SPcore, to existing PBNR accelerators. Compared to a mobile GPU, SLTarch achieves 3.9× speedup and 98% energy savings with negligible architecture overhead. Compared to existing accelerator designs, SLTarch achieves 1.8× speedup with 54% energy savings.

The wide deployment of the generative pre-trained transformer (GPT) has raised privacy concerns for both clients and servers. While cryptographic primitives can be employed for secure GPT inference to protect the privacy of both parties, they introduce considerable performance overhead. To accelerate secure inference, this study proposes a public decoding and secure verification approach that utilizes public GPT models, motivated by the observation that securely decoding one and multiple tokens takes a similar latency. The client uses the public model to generate a set of tokens, which are then securely verified by the private model for acceptance. The efficiency of our approach depends on the acceptance ratio of tokens proposed by the public model, which we improve from two aspects: (1) a private sampling protocol optimized for cryptographic primitives and (2) model alignment using knowledge distillation. Our approach improves the efficiency of secure decoding while maintaining the same level of privacy and generation quality as standard secure decoding. Experiments demonstrate a $2.1\times \sim 6.0\times$ speedup compared to standard decoding across three pairs of public-private models and different network conditions.

Efficient key-value (KV) cache compression is critical for scaling transformer-based Large Language Models (LLMs) in long sequences and resource-limited settings. Existing methods evict tokens based on their positions or importance, but position-based strategies can miss crucial information outside predefined regions, while those relying on global importance scores resulting in strong regional biases, limiting the KV cache's overall context retention and potentially impairing the performance of LLMs on complex tasks. Our wavelet analysis reveals that as tokens approach the end of sequence, their contributions to generation gradually increase and tends to diverge more from neighboring tokens, indicating a smooth transition with increasing complexity and variability from distant to nearby context. Motivated by this observation, we propose TreeKV, an intuitive, training-free method that employs a tree structure for smooth cache compression. TreeKV maintains a fixed cache size, allowing LLMs to deliver high-quality output in long text scenarios and is applicable during both the generation and prefilling stages. TreeKV consistently surpasses all baseline models in language modeling tasks on PG19 and OpenWebText2, allowing LLMs trained with short context window to generalize to longer window with a 16x cache reduction. On the Longbench benchmark, TreeKV achieves the best performance with only 6% of the budget at optimal efficiency.

3D Gaussian Splatting (3DGS) has vastly advanced the pace of neural rendering, but it remains computationally demanding on today's mobile SoCs.To address this challenge, we propose Lumina, a hardware-algorithm co-designed system, which integrates two principal optimizations: a novel algorithm, S 2 , and a radiance caching mechanism, RC, to improve the efficiency of neural rendering.S 2 algorithm exploits temporal coherence in rendering to reduce the computational overhead, while RC leverages the color integration process of 3DGS to decrease the frequency of intensive rasterization computations.Coupled with these techniques, we propose an accelerator architecture, LuminCore, to further accelerate cache lookup and address the fundamental inefficiencies in Rasterization.We show that Lumina achieves 4.5× speedup and 5.3× energy reduction against a mobile Volta GPU, with a marginal quality loss (< 0.2 dB peak signal-to-noise ratio reduction) across synthetic and real-world datasets.

The growing adoption of NVIDIA Jetson devices in edge-AI applications highlights the need for accurate architecture simulation tools on their integrated GPUs. Existing cycle-accurate GPU simulators primarily target traditional discrete GPUs and exhibit significant inaccuracies when applied to Jetson integrated GPUs. While Accel-Sim serves as the most widely used academic simulator for NVIDIA GPU research, its lack of support for the latest Jetson integrated GPUs severely hinders architectural exploration for next generation edge-AI devices.We propose Accel-Sim-J, which bridges the gap by repurposing Accel-Sim simulation framework to NVIDIA Jetson GPUs. We refine three major Accel-Sim framework components by applying tuner modifications, GPGPU-Sim performance model enhancements, and correlator adjustments. These improvements enable precise Jetson GPU simulation support, reducing simulation cycle errors from 29.0% to 22.7% on the Rodinia benchmark and from 26.1% to 16.1% on a transformer block. Furthermore, our enhanced architectural support for Ampere GPUs achieves a considerable reduction in simulation error (from 140.1% to 50.2%) for GEMM kernels.Based on Accel-Sim-J, we conduct a case study investigating the architecture design difference between an edge GPU and a traditional one. Specifically, we compare the optimal Compute-to-Cache (C2C) ratio by changing the L2 cache size of Jetson AGX Orin and RTX 3090. We conclude that Jetson GPUs demonstrate a higher optimal C2C ratio than discrete GPUs for the same workloads. We suggest that designers reduce the on-chip area proportion of the L2 cache in the next generation Jetson GPU design for better performance and efficiency.

Quantum Low-Density Parity-Check (qLDPC) codes are a promising class of quantum error-correcting codes that exhibit constantrate encoding and high error thresholds, thereby facilitating scalable fault-tolerant quantum computation.However, real-time decoding of qLDPC codes remains a significant challenge due to the high connectivity of their check matrices, which typically requires solving large-scale linear systems with sparse structures.In particular, off-the-shelf qLDPC decoders are often subject to a tradeoff between accuracy and latency, thus yielding no accurate and realtime decoding.This paper presents Vegapunk, a software-hardware co-design framework that enables real-time qLDPC decoding with high accuracy.To improve decoding accuracy, we design an offline decoupling strategy leveraging Satisfiability Modulo Theories (SMT) optimizations to mitigate quantum degeneracy.To enable fast decoding, we introduce an online hierarchical decoding algorithm employing a greedy strategy.Furthermore, we show that our SMT-optimized strategy suffices to produce decoupled matrices with maximized sparsity, thus admitting a dedicated accelerator to fully exploit the sparsity and parallelism to achieve real-time qLDPC decoding.Experimental results demonstrate that Vegapunk enables real-time decoding (< 1𝜇𝑠) for the Bivariate Bicycle (BB) code up to [[784,24,24]] while exhibiting logical error rates on par with the state-of-the-art decoder, i.e., BP+OSD.

Speculative decoding improves LLM inference by generating and verifying multiple tokens in parallel, but existing systems suffer from suboptimal performance due to a mismatch between dynamic speculation and static runtime assumptions. We present Yggdrasil, a co-designed system that enables latency-optimal speculative decoding through context-aware tree drafting and compiler-friendly execution. Yggdrasil introduces an equal-growth tree structure for static graph compatibility, a latency-aware optimization objective for draft selection, and stage-based scheduling to reduce overhead. Yggdrasil supports unmodified LLMs and achieves up to $3.98\times$ speedup over state-of-the-art baselines across multiple hardware setups.

Large language model (LLM) decoding suffers from high latency due to fragmented execution across operators and heavy reliance on off-chip memory for data exchange and reduction. This execution model limits opportunities for fusion and incurs significant memory traffic and kernel launch overhead. While modern architectures such as NVIDIA Hopper provide distributed shared memory and low-latency intra-cluster interconnects, they expose only low-level data movement instructions, lacking structured abstractions for collective on-chip communication. To bridge this software-hardware gap, we introduce two cluster-level communication primitives, ClusterReduce and ClusterGather, which abstract common communication patterns and enable structured, high-speed data exchange and reduction between thread blocks within a cluster, allowing intermediate results to be on-chip without involving off-chip memory. Building on these abstractions, we design ClusterFusion, an execution framework that schedules communication and computation jointly to expand operator fusion scope by composing decoding stages such as QKV Projection, Attention, and Output Projection into a single fused kernels. Evaluations on H100 GPUs show that ClusterFusion outperforms state-of-the-art inference frameworks by $1.61\times$ on average in end-to-end latency across different models and configurations.

Dynamic-shape tensor computation poses challenges for shape-specific compilation due to variable input dimensions. Existing compilers rely on shape samples, incurring high tuning costs and performance degradation on unseen inputs. We present Helix, a dynamic tensor compilation framework with sample-free compilation and architecture-guided optimization to achieve both compilation efficiency and shape-general performance. To avoid shape sampling, Helix constructs shape-agnostic compilation by decomposing computations across architectural layers. A bidirectional strategy combines top-down abstraction to align tensor computations with architectural hierarchies, and bottom-up kernel construction to build efficient execution strategies from reusable, architecture-aligned micro-kernels. A hybrid analyzer ensures accuracy through profiling at lower architectural levels, and achieves scalability through architecture-informed modeling at higher levels and runtime. This hierarchical design eliminates shape-specific tuning and enables shape-adaptive execution. Evaluations conducted on x86 CPUs, ARM CPUs, and NVIDIA GPUs demonstrate that Helix reduces compilation time by 174 × over the existing compilers and delivers 2.26 × and 3.29 × execution speedups over vendor libraries and dynamic-shape compilers, respectively.

With the increasing practicality of autonomous vehicles and drones, the importance of reliability requirements has escalated substantially. In many instances, traditional system designs tend to overlook reliability issues, emphasizing primarily on performance constraints. However, certain designers may opt for a lock-step (redundant) system design, duplicating every component, which in turn can result in significant performance, energy, and cost overheads. In software for autonomous machines, such as self-driving vehicles, performance degradation can increase reaction time, posing safety risks and reducing mission success rates. This article introduces a novel multi-domain lock-step system design, Kindred , which places a strong emphasis on maximizing reliability while minimizing performance overhead. The proposed approach capitalizes on the inherent diversity in fault tolerance among various tasks within autonomous machine software, intelligently scheduling only the vulnerable nodes in the lock-domain. The primary challenge addressed in this study involves the intelligent task scheduling across different domains, complemented by efficient error detection and correction in the lock-domain. In a real system demonstration, we illustrate the effectiveness of Kindred , showcasing its ability to attain the same level of reliability as a full lock-step system while incurring only a mere 2.8% overhead, as opposed to a fully split system, indicating the advantages and potential of our multi-domain lock-step system design in achieving high reliability without compromising performance.

Integrating GPUs into serverless computing platforms is crucial for improving efficiency. Many GPU functions, such as DNN inferences and scientific services, benefit from GPU usage, which requires only tens to hundreds of milliseconds for pure computation. Under these circumstances, fast data loading is imperative for function performance. However, existing GPU serverless systems face significant data stall issues, leading to extremely low GPU efficiency. Faced with the above problems, we observe opportunities to optimize data loading, such as data preloading and deduplicated data loading. However, these optimizations are impossible in existing GPU serverless systems due to the lack of insights into data information, such as data sizes and read-write attributes of function inputs. To address this, we propose a novel GPU serverless system, EDAS. EDAS first enhances user request specifications, allowing users to annotate data retrieved by GPU functions from the database with additional attributes. Based on this, EDAS takes over data loading from GPU functions and proposes two innovative data loading management schemes: a parallelized data loading scheme and a multi-stage resource exit scheme. Our experimental results show that EDAS reduces function duration by 16.2× and improves system throughput by 1.91× compared with the state-of-the-art serverless platform.

Leveraging sparsity in deep neural network (DNN) models holds significant promise for accelerating model inference. However, current GPUs can only harness sparsity in model weights, leaving activations unutilized due to their dynamic and unpredictable nature, which poses a considerable challenge for exploitation. In our research, we introduce a novel architectural approach aimed at effectively leveraging dual-side sparsity, encompassing both weight and activation sparsity. Our methodology involves a systematic examination of previous sparsity-related architectures, and culminating in the proposal of an uncharted paradigm that combines outer-product computation primitive and bitmap-based encoding format. Our approach showcases feasibility through minimal modifications to existing production-scale inner-product-based Tensor Cores. We introduce a set of innovative ISA extensions and carefully co-design matrix-matrix multiplication and convolution algorithms, the two predominant computation patterns in contemporary DNN models, to exploit our novel dual-side sparse Tensor Core. Our evaluation demonstrates the efficacy of our design, unlocking the full potential of dual-side DNN sparsity and delivering performance enhancements of up to an order of magnitude while incurring only modest hardware overhead.

Large-scale deep neural networks (DNNs), such as large language models (LLMs), have revolutionized the artificial intelligence (AI) field and become increasingly popular. However, training or fine-tuning such models requires substantial computational power and resources, where the memory capacity of a single acceleration device like a GPU is one of the most important bottlenecks. Owing to the prohibitively large overhead (e.g., 10×) of GPUs' native memory allocator, DNN frameworks like PyTorch and TensorFlow adopt a caching allocator that maintains a memory pool with a splitting mechanism for fast memory (de)allocation. Unfortunately, the caching allocator's efficiency degrades quickly for popular memory reduction techniques such as re-computation, offloading, distributed training, and low-rank adaptation. The primary reason is that those memory reduction techniques introduce frequent and irregular memory (de)allocation requests, leading to severe fragmentation problems for the splitting-based caching allocator. To mitigate this fragmentation problem, we propose a novel memory allocation framework based on low-level GPU virtual memory management called GPU memory lake (GMLake). GMLake employs a novel virtual memory stitching (VMS) mechanism, which can fuse or combine non-contiguous memory blocks with a virtual memory address mapping. GMLake can reduce average of 9.2 GB (up to 25 GB) GPU memory usage and 15% (up to 33%) fragmentation among eight LLM models on GPU A100 with 80 GB memory. GMLake is completely transparent to the DNN models and memory reduction techniques and ensures the seamless execution of resource-intensive deep-learning tasks. We have open-sourced GMLake at https://github.com/intelligent-machine-learning/glake/tree/main/GMLake.

Recently, memory consumption of Deep Neural Network (DNN) rapidly increases, mainly due to long lifetimes and large shapes of tensors. Graph scheduling has emerged as an effective memory optimization technique, which determines the optimal execution, re-computation, swap-out, and swap-in timings for each operator/tensor. However, it often hurts performance significantly and can only manipulate tensors' lifetimes but not shapes, limiting the optimization space. We find that graph transformation, which can change the tensor shapes and graph structure, creates a new trade-off space between memory and performance. Nevertheless, graph transformation are applied separately so far, with primary focus on optimizing performance and not memory.

The success of deep neural networks (DNNs) has sparked efforts to analyze (e.g., tracing) and optimize (e.g., pruning) them. These tasks have specific requirements and ad-hoc implementations in current execution backends like TensorFlow/PyTorch, which require developers to manage fragmented interfaces and adapt their codes to diverse models. In this study, we propose a new framework called Amanda to streamline the development of these tasks. We formalize the implementation of these tasks as neural network instrumentation, which involves introducing instrumentation into the operator level of DNNs. This allows us to abstract DNN analysis and optimization tasks as instrumentation tools on various DNN models. We build Amanda with two levels of APIs to achieve a unified, extensible, and efficient instrumentation design. The user-level API provides a unified operator-grained instrumentation API for different backends. Meanwhile, internally, we design a set of callback-centric APIs for managing and optimizing the execution of original and instrumentation codes in different backends. Through these design principles, the Amanda framework can accommodate a broad spectrum of use cases, such as tracing, profiling, pruning, and quantization, across different backends (e.g., TensorFlow/PyTorch) and execution modes (graph/eager mode). Moreover, our efficient execution management ensures that the performance overhead is typically kept within 5%.

Model pruning, which eliminates redundant parameters and reduces computational complexity, emerges as a viable strategy for efficient deep neural network (DNN) deployment. Owing to the irregular memory access and computation patterns in the sparse DNN models after pruning, existing arts have suggested various structured sparse patterns to enhance sparse DNN performance. In this work, we propose a unique perspective of understanding existing sparse pattern design as computation-skipping after tiling the tensor computation into multi-level hierarchies. This unified perspective opens up a new design space of multi-level sparse tiling to maximize the sparsity benefits of DNNs, as opposed to the single-level choice in current practices. We present Fractal, an auto-tuning system for sparse patterns that identifies the optimal multi-level sparse tiling pattern. We introduce PatternIR, a novel high-level intermediate representation (IR), to express a diverse range of multi-level sparse patterns. By leveraging insights from prior dense operator optimizations, we translate PatternIR into low-level compiler IRs, facilitating further operator optimization and code generation. Our evaluations demonstrate that Fractal yields substantial speedups of up to on average 3.16× on CUDA Core, 2.52× on TensorCore of GPUs compared to the state-of-art dense baseline under 75% sparsity while upholding minimal accuracy degradation compared to prior sparse operator libraries.

Approximate nearest neighbor (ANN) search is a widely applied technique in modern intelligent applications, such as recommendation systems and vector databases. Therefore, efficient and high-throughput execution of ANN search has become increasingly important. In this paper, we first characterize the state-of-the-art product quantization-based method of ANN search and identify a significant source of inefficiency in the form of unnecessary pairwise distance calculations and accumulations. To improve efficiency, we propose Juno, an end-to-end ANN search system that adopts a carefully designed sparsity- and locality-aware search algorithm. We also present an efficient hardware mapping that utilizes ray tracing cores in modern GPUs with pipelined execution on tensor cores to execute our sparsity-aware ANN search algorithm. Our evaluations on four datasets from 1 to 100 million search points demonstrate 2.2×-8.5× improvements in search throughput. Moreover, our algorithmic enhancements alone achieve a maximal 2.6× improvement on the hardware without the acceleration of the RT core.

The quadratic complexity of self-attention in Transformers has hindered the processing of long text. To alleviate this problem, previous works have proposed to sparsify the attention matrix, taking advantage of the observation that crucial information about a token can be derived from its neighbors. These methods typically combine one or another form of local attention and global attention. Such combinations introduce abrupt changes in contextual granularity when going from local to global, which may be undesirable. We believe that a smoother transition could potentially enhance model’s ability to capture long-context dependencies. In this study, we introduce Fovea Transformer, a long-context focused transformer that addresses the challenges of capturing global dependencies while maintaining computational efficiency. To achieve this, we construct a multi-scale tree from the input sequence, and use representations of context tokens with a progressively coarser granularity in the tree, as their distance to the query token increases. We evaluate our model on three long-context summarization tasks1. It achieves state-of-the-art performance on two of them, and competitive results on the third with mixed improvement and setback of the evaluation metrics.

The emerging trend of autonomous embedded systems minimizing human intervention has raised new questions about continuously maximizing system energy efficiency faced with stochastic runtime variance, which is costly for resource-constrained autonomous embedded systems. Considering heterogeneous hardware and variable software, we envision opportunities for vertical and horizontal shadow cycles within the AES pipeline for management facilities. This paper introduces SHEEO, a continuous energy efficiency optimizer that exploits underutilized heterogeneous computing resources to pursue variability-aware power management. To achieve this, SHEEO constantly monitors inner and outer variances and customizes reinforcement learning into two phases for stochastic runtime variance. We implement and deploy SHEEO on a commercial edge platform. The evaluation results show that SHEEO harvests up to 88% shadow cycles and improves up to 39% energy efficiency compared to state-of-the-art power management techniques with negligible overheads.

Neural Radiance Field (NeRF) is widely seen as an alternative to traditional physically-based rendering. However, NeRF has not yet seen its adoption in resource-limited mobile systems such as Virtual and Augmented Reality (VR/AR), because it is simply extremely slow. On a mobile Volta GPU, even the state-of-the-art NeRF models generally execute only at 0.8 FPS. We show that the main performance bottlenecks are both algorithmic and architectural. We introduce, Cicero, to tame both forms of inefficiencies. We first introduce two algorithms, one fundamentally reduces the amount of work any NeRF model has to execute, and the other eliminates irregular DRAM accesses. We then describe an on-chip data layout strategy that eliminates SRAM bank conflicts. A pure software implementation of Cicero offers an $8.0 \times$ speed-up and $7.9 \times$ energy saving over a mobile Volta GPU. When compared to a baseline with a dedicated DNN accelerator, our speed-up and energy reduction increase to $28.2 \times$ and $37.8 \times$, respectively - all with minimal quality loss (less than 1.0 dB peak signal-to-noise ratio reduction).

Autonomous Things (AuT) refers to a collection of self-sufficient tiny devices capable of performing intelligent computations. Looking ahead, AuT promises to enable ubiquitous deployment of intelligence on many emerging consumer electronics and mission-critical infrastructures. Nevertheless, there is an important research gap to date: architecting efficient AuT systems requires both energy autonomy (EA) and inference autonomy (IA). In other words, practical AuT application scenarios necessitate tailored architectures with significantly expanded inference performance and more efficient use of energy.We present CHRYSALIS, a novel automated EA/IA co-design methodology for autonomous things. It aims to guide the transition from a traditional EA-only and IA-only design approach to a truly AuT-oriented architecture design. To fully understand the interrelationship between the EA domain and the IA domain, CHRYSALIS first introduces an architectural modeling framework encompassing every key AuT module involving energy harvesting, intermittent execution, and accelerator control. Based on the holistic system model, we design an intelligent architecture generation tool that can help find the ideal design for targeted AuT scenarios adhering to different SWaP (Size, Weight and Power) constraints. To validate our work, we use CHRYSALIS for fast construction and exploration of efficient AuT design and pre-RTL design in representative AuT scenarios. Extensive evaluation shows that CHRYSALIS outperforms state-of-the-art designs and our proposed technique shows 56.4% better performance on average. We believe that the methodology and tools developed in this paper will foster the development of more performant and practical architectures in the upcoming AuT era.

Transformer models have gained significant attention due to their power in machine learning tasks. Their extensive deployment has raised concerns about the potential leakage of sensitive information during inference. However, when being applied to Transformers, existing approaches based on secure two-party computation (2PC) bring about efficiency limitations in two folds: (1) resource-intensive matrix multiplications in linear layers, and (2) complex non-linear activation functions like $\mathsf{GELU}$ and $\mathsf{Softmax}$. This work presents a new two-party inference framework $\mathsf{Nimbus}$ for Transformer models. Specifically, we propose a new 2PC paradigm to securely compute matrix multiplications based on an outer-product insight, which achieves $2.9\times \sim 12.5\times$ performance improvements compared to the state-of-the-art (SOTA) protocol. Furthermore, through a new observation of utilizing the input distribution, we propose an approach of low-degree polynomial approximation for $\mathsf{GELU}$ and $\mathsf{Softmax}$, which improves the performance of the SOTA polynomial approximation by $2.9\times \sim 4.0\times$, where the average accuracy loss of our approach is 0.08\% compared to the non-2PC inference without privacy. Compared with the SOTA two-party inference, $\mathsf{Nimbus}$ improves the end-to-end performance of $BERT_{base}$ inference by $2.7\times \sim 4.7\times$ across different network settings.

Neural Radiance Field (NeRF) has emerged as a promising alternative for photorealistic rendering. Despite recent algorithmic advancements, achieving real-time performance on today’s resource-constrained devices remains challenging. In this article, we identify the primary bottlenecks in current NeRF algorithms and introduce a unified algorithm-architecture co-design, Potamoi , designed to accommodate various NeRF algorithms. Specifically, we introduce a runtime system featuring a plug-and-play algorithm, SpaRW , which significantly reduces the per-frame computational workload and alleviates compute inefficiencies. Furthermore, our unified streaming pipeline coupled with customized hardware support effectively tames both SRAM and DRAM inefficiencies by minimizing repetitive DRAM access and completely eliminating SRAM bank conflicts. When evaluated against a baseline utilizing a dedicated DNN accelerator, our framework demonstrates a speedup and energy reduction of 53.1× and 67.7×, respectively, all while maintaining high visual quality with less than a 1.0 dB reduction in peak signal-to-noise ratio.

Network pruning can reduce the computation cost of deep neural network (DNN) models. However, sparse models often produce randomly-distributed weights to maintain accuracy, leading to irregular computations. Consequently, unstructured sparse models cannot achieve meaningful speedup on commodity hardware built for dense matrix computations. Accelerators are usually modified or designed with structured sparsity-optimized architectures for exploiting sparsity. For example, the Ampere architecture introduces a sparse tensor core, which adopts the 2:4 sparsity pattern.We propose a pruning method that builds upon the insight that matrix multiplication generally breaks the large matrix into multiple smaller tiles for parallel execution. We present the “tile-wise” sparsity pattern, which maintains a structured sparsity pattern at the tile level for efficient execution but allows for irregular pruning at the global scale to maintain high accuracy. In addition, the tile-wise sparsity is implemented at the global memory level, and the 2:4 sparsity executes at the register level inside the sparse tensor core. We can combine these two patterns into a “tile-vector-wise” (TVW) sparsity pattern to explore more fine-grained sparsity and further accelerate the sparse DNN models. We evaluate the TVW on the GPU, achieving averages of 1:85×, 2:75×, and 22:18× speedups over the dense model, block sparsity, and unstructured sparsity.

As graph neural networks (GNNs) have achieved great success in many graph learning problems, it is of paramount importance to support their efficient execution. Different graphs and different operators present different patterns during execution. However, there is still a gap in the existing GNN acceleration research to explore adaptive parallelism. We show that existing GNN frameworks rely on handwritten static kernels, which fail to achieve the best performance across different graph operators and input graph structures. In this work, we propose uGrapher, a unified interface that achieves general high performance for different graph operators and datasets. The existing GNN frameworks can easily integrate our design for its simple and unified API. We take a principled approach that decouples a graph operator’s computation and schedule to achieve that. We first build a GNN-specific operator abstraction that incorporates the semantics of graph tensors and graph loops. We explore various schedule strategies based on the abstraction that can balance the well-established trade-off relationship between parallelism, locality, and efficiency. Our evaluation shows that uGrapher can bring up to 29.1× (3.5× on average) performance improvement over the state-of-the-art baselines on two studied NVIDIA GPUs.

Graph neural networks (GNNs) are powerful tools for exploring and learning from graph structures and features. As such, achieving high-performance execution for GNNs becomes crucially important. Prior works have proposed to explore the sparsity (i.e., low density) in the input graph to accelerate GNNs, which uses the full-graph-level or block-level sparsity format. We show that they fail to balance the sparsity benefit and kernel execution efficiency. In this paper, we propose a novel system, referred to as AdaptGear, that addresses the challenge of optimizing GNNs performance by leveraging kernels tailored to the density characteristics at the subgraph level. Meanwhile, we also propose a method that dynamically chooses the optimal set of kernels for a given input graph. Our evaluation shows that AdaptGear can achieve a significant performance improvement, up to 6.49× (1.87× on average), over the state-of-the-art works on two mainstream NVIDIA GPUs across various datasets.

With the ever-increasing computational demand of DNN training workloads, distributed training has been widely adopted. A combination of data, model and pipeline parallelism strategy, called hybrid parallelism distributed training, is imported to tackle the problem of deploying large-scale models. However, how to evaluate the hybrid strategy and the utilization of each device remains a challenge since existing works either profile on a real large-scale cluster with high time and money costs or only analyze a specific type of parallelism without considering the hybrid parallelism. In this work, we proposed DistSim, an event-based performance model to accurately analyze each device's computation and communication activities with low profiling costs. DistDim breaks down the model into events according to the given distributed strategy, which can be profiled on two nodes. Then DistSim leverages the hierarchy of different parallel strategies to generate the computation and communication event-flow from layer level to model level and finally the activity timeline of each device participating in training. Experiment shows that DistSim can reach <4% errors when predicting distributing training batch time and <5% errors when predicting a single device's activity time in various hybrid strategy settings. We also provide a use-case of DistSim, automatically evaluate and search the best distributed training strategy, and find a hybrid strategy with at most 7.37× throughput improvement.

With the rapid development of cloud computing, the increasing scale of clusters and task parallelism put forward higher requirements on the scheduling capability at scale. To this end, the shared-state scheduler architecture has emerged as the popular solution for large-scale scheduling due to its high scalability and utilization. In such an architecture, a central resource state view periodically updates the global cluster status to distributed schedulers for parallel scheduling. However, the schedulers obtain broader resource views at the cost of intermittently stale states, rendering resources released invisible to schedulers until the next view update. These fleeting resource fragments are referred to as shadow resources in this paper. Current shared-state solutions overlook or fail to systematically utilize the shadow resources, leaving a void in fully exploiting these invisible resources.

Machine learning models with various tensor operators are becoming ubiquitous in recent years. There are two types of operators in machine learning: compute-intensive operators (e.g., GEMM and convolution) and memory-intensive operators (e.g., ReLU and softmax). In emerging machine learning models, compute-intensive operators are usually organized in a chain structure. With the continual specialization of hardware, the gap between computing performance and memory bandwidth has become more prominent. Consequently, the implementations of many compute-intensive operator chains are bounded by memory bandwidth, and generating fused kernels to improve locality for these compute-intensive operators becomes necessary. But in existing machine learning compilers, there lack both precise analysis and efficient optimization for compute-intensive operator chains on different accelerators. As a result, they usually produce sub-optimal performance for these operator chains.In this paper, we propose Chimera, an optimizing framework that can efficiently improve the locality of compute-intensive operator chains on different hardware accelerators. In Chimera, each compute-intensive operator is composed of a series of computation blocks. To generate efficient fused kernels for the operator chains, optimizations for both inter-block and intra-block are required. For inter-block optimization, Chimera decides the optimized block execution order by minimizing the data movement volume among blocks using an analytical model. For intra-block optimization, Chimera uses unified replaceable micro kernels to apply hardware-specific optimizations for different accelerators. Finally, Chimera generates fused kernels for compute-intensive operator chains. Evaluation of batch GEMM chains and convolution chains on CPU, GPU, and NPU shows that Chimera achieves up to 2.87×, 2.29×, and 2.39× speedups to hand-tuned libraries. Compared to state-of-the-art compilers, the speedups are up to 2.29×, 1.64×, and 1.14× for CPU, GPU, and NPU.

Latency-critical applications directly interact with end users and often experience the diurnal load pattern. In production, best-effort applications are often co-located with them to utilize the idle cores at the low load. Meanwhile, modern computers are evolving towards heterogeneous NUMA architecture, where the cores have different computation abilities, memory access latencies and network communication delays. Prior co-location scheduling work did not consider the NUMA architecture, and failed to maximize the throughput of best-effort applications while ensuring the required QoS of latency-critical applications. Our investigation shows that NUMA effect has complex impacts on the latency of latency-critical applications and the throughput of best-effort applications. We therefore propose PAC, a preference-aware co-location scheduling scheme that considers the NUMA effect for heterogeneous NUMA architectures. PAC has a performance monitor and a core scheduler. Specifically, the performance monitor identifies the "dangerous" latency-critical applications that require upgrading core allocations. We propose two low-overhead scheduling strategies for the scheduler. The strategies identify the bottlenecks of applications and adjust core allocations accordingly. Experimental result shows that PAC improves the throughput of best-effort applications by 3.87× while ensuring the required QoS of latency-critical applications.

Emerging cloud-native development models raise new challenges for managing server performance and power at microsecond scale. Compared with traditional cloud workloads, serverless functions exhibit unprecedented heterogeneity, variability, and dynamicity. Designing cloud-native power management schemes for serverless functions requires significant engineering effort. Current solutions remain sub-optimal since their orchestration process is often one-sided, lacking a systematic view. A key obstacle to truly efficient function deployment is the fundamental wide abstraction gap between the upper-layer request scheduling and the low-level hardware execution.In this work, we show that the optimal operating point (OOP) for energy efficiency cannot be attained without synthesizing the multi-dimensional attributes of functions. We present FIRST, a novel mechanism that enables servers to better orchestrate serverless functions. The key feature of FIRST is that it leverages a lightweight Internal Representation and meta-Scheduling (IRS) layer for collecting the maximum potential revenue from the servers. Specifically, FIRST follows a pipeline-style workflow. Its frontend components aim to analyze functions from different angles and expose their key features to the system. Meanwhile, its backend components are able to make informed function assignment decisions to avoid OOP divergence. We further demonstrate the way to create extensions based on FIRST to enable versatile cloud-native power management. In total, our design constitutes a flexible management layer that supports power-aware function deployment. We show that FIRST could allow 94% functions to be processed under the OOP, which brings up to 24% energy efficiency improvements.

Transformer-based large language models (LLMs) have achieved great success with the growing model size. LLMs' size grows by 240× every two years, which outpaces the hardware progress and makes model inference increasingly costly. Model quantization is a promising approach to mitigate the widening gap between LLM size and hardware capacity. However, the existence of outliers, values with significant magnitudes, in LLMs makes existing quantization methods less effective. Prior outlier-aware quantization schemes adopt sparsity encoding techniques to separate outliers from normal values where the process requires global coordination (e.g., a global sparsity coordination list). This incurs complex encoding/decoding hardware logics and an extra orchestration controller for the computation between outlier and normal values. As such, it is not hardware-efficient and hence only achieves sub-optimal quantization benefits.

Image processing algorithms are prime targets for hardware acceleration as they are commonly used in resource- and power-limited applications. Today's image processing accelerator designs make rigid assumptions about the algorithm structures and/or on-chip memory resources. As a result, they either have narrow applicability or result in inefficient designs.

While deep neural network (DNN) models are mainly trained using GPUs, many companies and research institutions build shared GPU clusters. These clusters host DNN training jobs, DNN inference jobs, and CPU jobs (jobs in traditional areas). DNN training jobs require GPU for main computation and CPU for auxiliary computation. Some DNN inference jobs could rely solely on CPU, while others must utilize both CPU and GPU. Our investigation demonstrates that the number of cores allocated to a training job significantly impacts its performance, and that DNN inference jobs can make use of the limited CPU cores on the GPU nodes. To accomplish this, we characterize representative deep learning models in terms of their CPU core requirements for their training jobs and inference jobs, and investigate their sensitivity to other CPU-side resource contention. Based on the characterization, we propose SODA, a scheduling system comprised of an adaptive CPU allocator, a multi-array job scheduler, a hardware-aware inference job placer, and a real-time contention eliminator. The experimental results indicate that SODA increases GPU utilization by an average of 19.9%, while maintaining the quality of service target for all DNN inference jobs and the queuing performance of CPU jobs.

Transformer models have achieved promising results on natural language processing (NLP) tasks including extractive question answering (QA). Common Transformer encoders used in NLP tasks process the hidden states of all input tokens in the context paragraph throughout all layers. However, different from other tasks such as sequence classification, answering the raised question does not necessarily need all the tokens in the context paragraph. Following this motivation, we propose Block-skim, which learns to skim unnecessary context in higher hidden layers to improve and accelerate the Transformer performance. The key idea of Block-Skim is to identify the context that must be further processed and those that could be safely discarded early on during inference. Critically, we find that such information could be sufficiently derived from the self-attention weights inside the Transformer model. We further prune the hidden states corresponding to the unnecessary positions early in lower layers, achieving significant inference-time speedup. To our surprise, we observe that models pruned in this way outperform their full-size counterparts. Block-Skim improves QA models' accuracy on different datasets and achieves 3 times speedup on BERT-base model.

Transformer architecture has become the defacto model for many machine learning tasks from natural language processing and computer vision.As such, improving its computational efficiency becomes paramount.One of the major computational inefficiency of Transformer-based models is that they spend the identical amount of computation throughout all layers.Prior works have proposed to augment the Transformer model with the capability of skimming tokens to improve its computational efficiency.However, they suffer from not having effectual and end-to-end optimization of the discrete skimming predictor.To address the above limitations, we propose the Transkimmer architecture, which learns to identify hidden state tokens that are not required by each layer.The skimmed tokens are then forwarded directly to the final output, thus reducing the computation of the successive layers.The key idea in Transkimmer is to add a parameterized predictor before each layer that learns to make the skimming decision.We also propose to adopt reparameterization trick and add skim loss for the end-to-end training of Transkimmer.Transkimmer achieves 10.97× average speedup on GLUE benchmark compared with vanilla BERT base baseline with less than 1% accuracy degradation.

Deep learning (DL) models have achieved great success in many application domains. As such, many industrial companies such as Google and Facebook have acknowledged the importance of multi-tenant DL services. Although the multi-tenant service has been studied in conventional workloads, it is not been deeply studied on deep learning service, especially on general-purpose hardware.

Multi-stage user-facing applications on GPUs are widely-used nowa- days, and are often implemented to be microservices. Prior re- search works are not applicable to ensuring QoS of GPU-based microservices due to the different communication patterns and shared resource contentions. We propose Astraea to manage GPU microservices considering the above factors. In Astraea, a microser- vice deployment policy is used to maximize the supported peak service load while ensuring the required QoS. To adaptively switch the communication methods between microservices according to different deployments, we propose an auto-scaling GPU communi- cation framework. The framework automatically scales based on the currently used hardware topology and microservice location, and adopts global memory-based techniques to reduce intra-GPU communication. Astraea increases the supported peak load by up to 82.3% while achieving the desired 99%-ile latency target compared with state-of-the-art solutions.

Spatial accelerators provide massive parallelism with an array of homogeneous PEs, and enable efficient data reuse with PE array dataflow and on-chip memory. Many previous works have studied the dataflow architecture of spatial accelerators, including performance analysis and automatic generation. However, existing accelerator generators fail to exploit the entire memory-level reuse opportunities, and generate suboptimal designs with data duplication and inefficient interconnection.

The attention mechanisms of transformers effectively extract pertinent information from the input sequence. However, the quadratic complexity of self-attention w.r.t the sequence length incurs heavy computational and memory burdens, especially for tasks with long sequences. Existing accelerators face performance degradation in these tasks. To this end, we propose SALO to enable hybrid sparse attention mechanisms for long sequences. SALO contains a data scheduler to map hybrid sparse attention patterns onto hardware and a spatial accelerator to perform the efficient attention computation. We show that SALO achieves 17.66x and 89.33x speedup on average compared to GPU and CPU implementations, respectively, on typical workloads, i.e., Longformer and ViL.

The proliferation of machine learning applications has promoted both CUDA Cores and Tensor Cores’ integration to meet their acceleration demands. While studies have shown that co-locating multiple tasks on the same GPU can effectively improve system throughput and resource utilization, existing schemes focus on scheduling the resources of traditional CUDA Cores and thus lack the ability to exploit the parallelism between Tensor Cores and CUDA Cores.In this paper, we propose Tacker, a static kernel fusion and scheduling approach to improve GPU utilization of both types of cores while ensuring the QoS (Quality-of-Service) of co-located tasks. Tacker consists of a Tensor-CUDA Core kernel fuser, a duration predictor for fused kernels, and a runtime QoS-aware kernel manager. The kernel fuser enables the flexible fusion of kernels that use Tensor Cores and CUDA Cores, respectively. The duration predictor precisely predicts the duration of the fused kernels. Finally, the kernel manager invokes the fused kernel or the original kernel based on the QoS headroom of latency-critical tasks to improve the system throughput. Our experimental results show that Tacker improves the throughput of best-effort applications compared with state-of-the-art solutions by 18.6% on average, while ensuring the QoS of latency-critical tasks.

An activation function is an element-wise mathematical function and plays a crucial role in deep neural networks (DNN). Many novel and sophisticated activation functions have been proposed to improve the DNN accuracy but also consume massive memory in the training process with back-propagation. In this study, we propose the nested forward automatic differentiation (Forward-AD), specifically for the element-wise activation function for memory-efficient DNN training. We deploy nested Forward-AD in two widely-used deep learning frameworks, TensorFlow and PyTorch, which support the static and dynamic computation graph, respectively. Our evaluation shows that nested Forward-AD reduces the memory footprint by up to 1.97× than the baseline model and outperforms the recomputation by 20% under the same memory reduction ratio.

In emerging DNN serving systems, queries are usually batched to fully leverage hardware resources, and all the queries in a batch run through the complete model and return at the same time. According to our findings, some queries only need to pass through a portion of the DNN model to attain sufficient precision in a DNN service. These queries can have shorter latencies if they can return early in the middle of a model. Therefore, we propose precision-aware multi-exit inference serving, PAME, to achieve the above purpose. PAME provides a holistic scheme to build a multi-exit DNN model and a corresponding system-level design of the inference engine. We use representative CV and NLP benchmarks to evaluate PAME. PAME is adaptive to various DNN tasks and service loads. Experimental results show that PAME reduces 39.9% average latency without increasing the tail latency, while maintaining 99.68% precision of the original single-exit DNN models on average.

Autonomous machines, such as Autonomous Vehicles (AV), are vulnerable to a variety of different faults such as radiation-induced soft/transient errors, adversarial attacks, and software bugs, which all jeopardize the reliability of autonomous machines. How vulnerable the AV software stack is to different error sources, however, remains an open question. This paper performs comprehensively fault injections to study how the AV software stack behaves under different error sources. We show that algorithms in an AV software stack inherently possess different forms of masking mechanisms. Based on the characteristic of the inherent fault tolerance mechanisms, we formalize the notion of Fault Tolerance Level (FTL), which quantifies how faults in an algorithm can be masked and/or attenuated without affecting the actuator commands, providing opportunities to relax fault protection. Leveraging the FTL formulation, we propose a dynamic protection system, which, at the high level, spends the limited protection budget (e.g., spatial/temporal redundancy) on the most vulnerable parts of the AV software (i.e., with the lowest FTL). Using Autoware as a case study, we show that our system reduces the error rate of AV software stack by more than 90% with negliaible performance overhead.

Quantization is a technique to reduce the computation and memory cost of DNN models, which are getting increasingly large. Existing quantization solutions use fixed-point integer or floating-point types, which have limited benefits, as both require more bits to maintain the accuracy of original models. On the other hand, variable-length quantization uses low-bit quantization for normal values and high-precision for a fraction of outlier values. Even though this line of work brings algorithmic benefits, it also introduces significant hardware overheads due to variable-length encoding and decoding.In this work, we propose a fixed-length a daptive n umerical data t ype called ANT to achieve low-bit quantization with tiny hardware overheads. Our data type ANT leverages two key innovations to exploit the intra-tensor and inter-tensor adaptive opportunities in DNN models. First, we propose a particular data type, flint, that combines the advantages of float and int for adapting to the importance of different values within a tensor. Second, we propose an adaptive framework that selects the best type for each tensor according to its distribution characteristics. We design a unified processing element architecture for ANT and show its ease of integration with existing DNN accelerators. Our design results in $2.8\times $ speedup and $2.5\times $ energy efficiency improvement over the state-of-the-art quantization accelerators.

With the development of deep neural networks (DNNs), more complex accelerators have been designed for more sophisticated networks. Naturally, the complexity of accelerators makes them vulnerable to transient errors. Also, some DNN accelerators are widely used the safety-critical systems, such as autonomous vehicles. Therefore, the susceptibility to transient errors makes research on mitigation techniques more significant, and errors of accelerators should be limited to none. Some researchers proposed the modular redundancy method, which offers a highly reliable way but also considerably increases overhead. In this regard, algorithm-based solutions offer cheaper solutions. However, their implementation is primarily observed in software-based error injections. In this study, we propose a novel approach that focuses on implementing algorithm-based error detection (ABED) for RTL-level (hardware-based) error injections. Previous studies generally focused on the impact of soft errors in memory structures of embedded system-based accelerators. However, the main goal of this research is to study the impact of soft errors in processing elements and how to mitigate them. We implement an algorithm-based error detection that utilizes checksums for verifying convolution operations with low overhead. We first explain how to overcome the challenges of implementing ABED on FPGA-based accelerators, then how to implement it. We implement and evaluate our solution on an industry-level DNN accelerator called NVIDIA deep learning accelerator (NVDLA). In this study, our error injection method is constructed to test the most common soft error scenarios in processing units. The results of the research show that algorithm-based fault tolerance can detect all silent data corruptions (SDC) while maintaining a very low overhead (6-23%) on runtime.

Graph sampling and random walk operations, capturing the structural properties of graphs, are playing an important role today as we cannot directly adopt computing-intensive algorithms on large-scale graphs. Existing system frameworks for these tasks are not only spatially and temporally inefficient, but many also lead to biased results. This paper presents Skywalker, a high-throughput, quality-preserving random walk and sampling framework based on GPUs. Skywalker makes three key contributions: first, it takes the first step to realize efficient biased sampling with the alias method on a GPU. Second, it introduces well-crafted load-balancing techniques to effectively utilize the massive parallelism of GPUs. Third, it accelerates alias table construction and reduce the GPU memory requirement with efficient memory management scheme. We show that Skywalker greatly outperforms the state-of-the-art CPU-based and GPU-based baselines, in a wide spectrum of workload scenarios.

Emerging GPUs have multiple Streaming Multiprocessors (SM), while each SM is comprised of CUDA Cores and Tensor Cores. While CUDA Cores do the general computation, Tensor Cores are designed to speed up matrix multiplication for deep learning applications. However, a GPU kernel often either uses CUDA Cores or Tensor Cores, leaving the other processing units idle. Although many prior research works have been proposed to co-locate kernels to improve GPU utilization, they cannot leverage the Intra-SM CUDA Core-Tensor Core Parallelism. We therefore propose Plasticine to exploit the intra-SM parallelism for maximizing the GPU throughput. Plasticine involves compilation and runtime schedule to achieve the above purpose. Experimental results on an Nvidia 2080Ti GPU show that Plasticine improves the system-wide throughput by 15.3% compared with prior co-location work.

Federated learning (FL) is a distributed machine learning paradigm that allows clients to collaboratively train a model over their own local data. FL promises the privacy of clients and its security can be strengthened by cryptographic methods such as additively homomorphic encryption (HE). However, the efficiency of FL could seriously suffer from the statistical heterogeneity in both the data distribution discrepancy among clients and the global distribution skewness. We mathematically demonstrate the cause of performance degradation in FL and examine the performance of FL over various datasets. To tackle the statistical heterogeneity problem, we propose a pluggable system-level client selection method named Dubhe, which allows clients to proactively participate in training, meanwhile preserving their privacy with the assistance of HE. Experimental results show that Dubhe is comparable with the optimal greedy method on the classification accuracy, with negligible encryption and communication overhead.

The microservice architecture is a hot trend which proposes to transform the traditional monolith application into massive dynamic and irregular small services. To boost the overall throughput and ensure the guaranteed latency, it is desirable to process massive service requests in parallel with efficient resource sharing in data centers. However, the disaggregation nature of microservice unavoidably upscales the design space of resource management and increases its complexity. In this paper, we propose AlphaR, a learning-powered resource management system tailored to the microservice environment. The basic idea of AlphaR is to generate microservice-specific resource management policies for improving efficiency. Specifically, we take the first step to use bipartite graph as a convenient abstraction for application built with microservices. Based on this, we devise a bipartite feature inference approach named Bi-GNN to extract the temporal characteristics of microservices. Furthermore, we implement a policy network to select appropriate resource allocation choices for maximizing the performance in resource-constrained data centers. AlphaR can improve the mean and p95 response time by up to 80% and 77.5% respectively compared with conventional schemes.

Leveraging sparsity in deep neural network (DNN) models is promising for accelerating model inference. Yet existing GPUs can only leverage the sparsity from weights but not activations, which are dynamic, unpredictable, and hence challenging to exploit. In this work, we propose a novel architecture to efficiently harness the dual-side sparsity (i.e., weight and activation sparsity). We take a systematic approach to understand the (dis)advantages of previous sparsity-related architectures and propose a novel, unexplored paradigm that combines outer-product computation primitive and bitmap-based encoding format. We demonstrate the feasibility of our design with minimal changes to the existing production-scale inner-product-based Tensor Core. We propose a set of novel ISA extensions and co-design the matrix-matrix multiplication and convolution algorithms, which are the two dominant computation patterns in today’s DNN models, to exploit our new dual-side sparse Tensor Core. Our evaluation shows that our design can fully unleash the dual-side DNN sparsity and improve the performance by up to one order of magnitude with small hardware overhead.

While user-facing services experience diurnal load patterns, co-locating services improve hardware utilization. Prior work on co-locating services on GPUs run queries sequentially, as the latencies of the queries are neither stable nor predictable when running simultaneously. The input sensitiveness and the non-deterministic operator overlap are two primary factors of the latency unpredictability. Hence, We propose Abacus, a runtime system that runs multiple services simultaneously. Abacus enables deterministic operator overlap to enforce latency predictability. Abacus composes of an overlap-aware latency predictor, a headroom-based query controller, and segmental model executors. The predictor predicts the latencies of the deterministic operator overlap. The controller determines the appropriate operator overlap for the QoS guarantee of all the services. The executors run the operators as needed to support the deterministic operator overlap. Our evaluation shows that Abacus reduces 51.3% of the QoS violation and improves the throughput by 29.8% on average compared with state-of-the-art solutions.

The energy efficiency of GPU architectures has emerged as an essential aspect of computer system design. In this article, we explore the energy benefits of reducing the GPU chip's voltage to the safe limit, i.e., Vminpoint, using predictive software techniques. We perform such a study on several commercial off-the-shelf GPU cards. We find that there exists about 20% voltage guardband on those GPUs spanning two architectural generations, which, if “eliminated” entirely, can result in up to 25% energy savings on one of the studied GPU cards. Our measurement results unveil a program dependent Vminbehavior across the studied applications, and the exact improvement magnitude depends on the program's available guardband. We make fundamental observations about the program-dependent Vminbehavior. We experimentally determine that the voltage noise has a more substantial impact on Vmincompared to the process and temperature variation, and the activities during the kernel execution cause large voltage droops. From these findings, we show how to use kernels' microarchitectural performance counters to predict its Vminvalue accurately. The average and maximum prediction errors are 0.5% and 3%, respectively. The accurate Vminprediction opens up new possibilities of a crosslayer dynamic guardbanding scheme for GPUs, in which software predicts and manages the voltage guardband, while the functional correctness is ensured by a hardware safety net mechanism.

In[1],Table IVand Table II are the same. The correctTable IVis meant for comparison against prior work.

Continuous vision is the cornerstone of a diverse range of intelligent applications found on emerging computing platforms such as autonomous machines and Augmented Reality glasses. A critical issue in today's continuous vision systems is their long end-to-end frame latency, which significantly impacts the system agility and user experience. We find that the long latency is fundamentally caused by the serialized execution model of today's continuous vision pipeline, whose key stages, including sensing, imaging, and vision computations, execute sequentially, leading to long frame latency.

Recent research on the multi-head attention mechanism, especially that in pre-trained models such as BERT, has shown us heuristics and clues in analyzing various aspects of the mechanism.As most of the research focus on probing tasks or hidden states, previous works have found some primitive patterns of attention head behavior by heuristic analytical methods, but a more systematic analysis specific on the attention patterns still remains primitive.In this work, we clearly cluster the attention heatmaps into significantly different patterns through unsupervised clustering on top of a set of proposed features, which corroborates with previous observations.We further study their corresponding functions through analytical study.In addition, our proposed features can be used to explain and calibrate different attention heads in Transformer models.

The research interest in specialized hardware accelerators for deep neural networks (DNN) spikes recently owing to their superior performance and efficiency. However, today’s DNN accelerators primarily focus on accelerating specific "kernels" such as convolution and matrix multiplication, which are vital but only part of an end-to-end DNN-enabled application. Meaningful speedups over the entire application often require supporting computations that are, while massively parallel, ill-suited to DNN accelerators. Integrating a general-purpose processor such as a CPU or a GPU incurs significant data movement overhead and leads to resource under-utilization on the DNN accelerators.We propose Simultaneous Multi-mode Architecture (SMA), a novel architecture design and execution model that offers general-purpose programmability on DNN accelerators in order to accelerate end-to-end applications. The key to SMA is the temporal integration of the systolic execution model with the GPU-like SIMD execution model. The SMA exploits the common components shared between the systolic-array accelerator and the GPU, and provides lightweight reconfiguration capability to switch between the two modes in-situ. The SMA achieves up to 63% performance improvement while consuming 23% less energy than the baseline Volta architecture with TensorCore.

Accelerators make the task of building systems that are re-silient against transient errors like voltage noise and soft errors hard. Architects integrate accelerators into the system as black box third-party IP components. So a fault in one or more accelerators may threaten the system's reliability if there are no established failure semantics for how an error propagates from the accelerator to the main CPU. Existing solutions that assure system reliability come at the cost of sacrificing accelerator generality, efficiency, and incur significant overhead, even in the absence of errors. To over-come these drawbacks, we examine reliability management of accelerator systems via hardware-software co-design, coupling an efficient architecture design with compiler and run-time support, to cope with transient errors. We introduce asymmetric resilience that architects reliability at the system level, centered around a hardened CPU, rather than at the accelerator level. At runtime, the system exploits task-level idempotency to contain accelerator errors and use memory protection instead of taking checkpoints to mitigate over-heads. We also leverage the fact that errors rarely occur in systems, and exploit the trade-off between error recovery performance and improved error-free performance to enhance system efficiency. Using GPUs, which are at the fore-front of accelerator systems, we demonstrate how our system architecture manages reliability in both integrated and discrete systems, under voltage-noise and soft-error related faults, leading to extremely low overhead (less than 1%) and substantial gains (20% energy savings on average).

While deep neural network (DNN) models are often trained on GPUs, many companies and research institutes build GPU clusters that are shared by different groups. On such GPU cluster, DNN training jobs also require CPU cores to run pre-processing, gradient synchronization. Our investigation shows that the number of cores allocated to a training job significantly impact its performance. To this end, we characterize representative deep learning models on their requirement for CPU cores under different GPU resource configurations, and study the sensitivity of these models to other CPU-side shared resources. Based on the characterization, we propose CODA, a scheduling system that is comprised of an adaptive CPU allocator, a real-time contention eliminator, and a multi-array job scheduler. Experimental results show that CODA improves GPU utilization by 20.8% on average without increasing the queuing time of CPU jobs.

Database platform-as-a-service (dbPaaS) is developing rapidly and a large number of databases have been migrated to run on the Clouds for the low cost and flexibility. Emerging Clouds rely on the tenants to provide the resource specification for their database workloads. However, they tend to over-estimate the resource requirement of their databases, resulting in the unnecessarily high cost and low Cloud utilization. A methodology that automatically suggests the “just-enough” resource specification that fulfills the performance requirement of every database workload is profitable.

Multi-tenant (or colocation) data centers are good solutions to support edge computing, since each enterprise or organization usually has limited servers at an edge site. When any data center tenant faces a burst of workload, renting resources from the other tenants in the same data center can provide the required resources while keeping the merits of edge computing, but its challenges of reliability and performance are daunting. In this paper, we propose OVERSEE, an outsourcing verification mechanism that enables resource sharing in multi-tenant data centers, fully exploiting the benefits of edge computing. OVERSEE addresses the above two challenges by making skillful use of Intel SGX suite. OVERSEE consists of two sub schemes, the Report-Proof mechanism and Sampling-Challenging mechanism. The Report-Proof mechanism guarantees a task outsourced by a tenant can be executed correctly, i.e., completely and without modification, in the operating environment provided by another tenant. The Sampling-Challenging mechanism can be used to verify that sufficient computing capacity is provided to achieve the required QoS according to the resource lease agreement between the tenants. The theoretical analysis shows the effectiveness of OVERSEE and the experimental results show that it brings minimal overhead.

Large-scale datacenters often host latency-sensitive services that have stringent Quality-of-Service requirement and experience diurnal load pattern. Co-locating best-effort applications that have no QoS requirement with latency-sensitive services has been widely used to improve the resource utilization with careful shared resource management. However, existing co-location techniques tend to result in the power overload problem on power constrained computers due to the ignorance of the power consumption. To this end, we propose Sturgeon, a runtime system proactively manages resources between colocated applications in a power constrained environment, to ensure the QoS of latency-sensitive services while maximizing the resource utilization. Our investigation shows that, at a given load, there are multiple feasible resource configurations to meet both QoS requirement and power budget, while one of them yields the maximum throughput of best-effort applications. To find such a configuration, we establish models to accurately predict the performance and power consumption of the colocated applications. Sturgeon monitors the QoS periodically in order to eliminate the potential QoS violation caused by the unpredictable interference. The experimental results show that Sturgeon improves the throughput of best-effort applications by 24.96% compared to the state-of-the-art technique, while guaranteeing the 95%-ile latency within the QoS target.

Deep learning is vulnerable to adversarial attacks, where carefully-crafted input perturbations could mislead a well-trained Deep Neural Network (DNN) to produce incorrect results. Adversarial attacks jeopardize the safety, security, and privacy of DNN-enabled systems. Today's countermeasures to adversarial attacks either do not have the capability to detect adversarial samples at inference-time, or introduce prohibitively high overhead to be practical at inference-time.We propose Ptolemy, an algorithm-architecture co-designed system that detects adversarial attacks at inference time with low overhead and high accuracy. We exploit the synergies between DNN inference and imperative program execution: an input to a DNN uniquely activates a set of neurons that contribute significantly to the inference output, analogous to the sequence of basic blocks exercised by an input in a conventional program. Critically, we observe that adversarial samples tend to activate distinctive paths from those of benign inputs. Leveraging this insight, we propose an adversarial sample detection framework, which uses canary paths generated from offline profiling to detect adversarial samples at runtime. The Ptolemy compiler along with the co-designed hardware enable efficient execution by exploiting the unique algorithmic characteristics. Extensive evaluations show that Ptolemy achieves higher or similar adversarial sample detection accuracy than today's mechanisms with a much lower (as low as 2%) runtime overhead.

Network pruning can reduce the high computation cost of deep neural network (DNN) models. However, to maintain their accuracies, sparse models often carry randomly-distributed weights, leading to irregular computations. Consequently, sparse models cannot achieve meaningful speedup on commodity hardware (e.g., GPU) built for dense matrix computations. As such, prior works usually modify or design completely new sparsity-optimized architectures for exploiting sparsity. We propose an algorithm-software co-designed pruning method that achieves latency speedups on existing dense architectures. Our work builds upon the insight that the matrix multiplication generally breaks the large matrix into multiple smaller tiles for parallel execution. We propose a tiling-friendly “tile-wise” sparsity pattern, which maintains a regular pattern at the tile level for efficient execution but allows for irregular, arbitrary pruning at the global scale to maintain the high accuracy. We implement and evaluate the sparsity pattern on GPU tensor core, achieving a 1.95× speedup over the dense model.

In today's manycore processors, the energy loss of more than 20% may result from inherent inefficiencies of conventional power delivery system (PDS) design. By stacking multiple voltage domains in series to lower the step-down conversion ratio of the off-chip voltage regulator module (VRM) and reduce the energy loss along the path of the power delivery network (PDN), voltage stacking (VS) offers a novel alternative power delivery technique to fundamentally improve power delivery efficiency (PDE). However, VS suffers from aggravated supply voltage noise from the current imbalance, which hinders its adoption. In this article, we investigate practical VS implementation in manycore processors to improve PDE and achieve reliable performance, while maintaining compatibility with advanced power management techniques. We first present the system configuration of a voltage-stacked manycore processor. We then systematically characterize supply voltage noise in VS, identify global, and residual differential currents as its dominant contributors, and calculate the possible worst supply voltage noise. We next propose a hybrid voltage regulation solution, based on a charge-recycling off-chip voltage regulator and distributed integrated voltage regulators, to mitigate supply voltage noise effectively. We also study the compatibility of VS with higher-level power management techniques. Finally, the performance of a voltage-stacked GPU system is comprehensively evaluated. The simulation results show that our approach can achieve 93.5% PDE, reducing the power loss by 13.6% compared to conventional single-layer PDS.

Recently, researchers have started decomposing deep neural network models according to their semantics or functions. Recent work has shown the effectiveness of decomposed functional blocks for defending adversarial attacks, which add small input perturbation to the input image to fool the DNN models. This work proposes a profiling-based method to decompose the DNN models to different functional blocks, which lead to the effective path as a new approach to exploring DNNs' internal organization. Specifically, the per-image effective path can be aggregated to the class-level effective path, through which we observe that adversarial images activate effective path different from normal images. We propose an effective path similarity-based method to detect adversarial images with an interpretable model, which achieve better accuracy and broader applicability than the state-of-the-art technique.

GPUs have been widely adopted to serve online deep learning-based services that have stringent QoS requirements. However, emerging deep learning serving systems often result in long latency and low throughput of the inference requests that damage user experience and increase the number of GPUs required to host an online service. Our investigation shows that the poor batching operation and the lacking of data transfer-computation overlap are the root causes of the long latency and low throughput. To this end, we propose Ebird, a deep learning serving system that is comprised of a GPU-resident memory pool, a multi-granularity inference engine, and an elastic batch scheduler. The memory pool eliminates the unnecessary waiting of the batching operation and enables data transfer-computation overlap. The inference engine enables concurrent execution of different batches, improving the GPUs resource utilization. The batch scheduler organizes inference requests elastically. Our experimental results on an Nvidia Titan RTX GPU show that Ebird reduces the response latency of inferences by up to 70.9% and improves the throughput by up to 49.3% while guaranteeing the QoS target compared with TensorFlow Serving.

Existing techniques for improving datacenter utilization while guaranteeing the QoS are based on the assumption that queries have similar behaviors. However, user queries in emerging compute demanding services demonstrate significantly diverse behavior and require adaptive parallelism. Our study shows that the end-to-end latency of the compute demanding query is determined together by the system-wide load, its workload, its parallelism, contention on shared cache, and memory bandwidth. When hosting such new services, the current cross-query resource allocation results in either severe QoS violation or significant resource under-utilization.

Modern large-scale computing systems (data centers, supercomputers, cloud and edge setups and high-end cyber-physical systems) employ heterogeneous architectures that consist of multicore CPUs, general-purpose many-core GPUs, and programmable FPGAs. The effective utilization of these architectures poses several challenges, among which a primary one is power consumption. Voltage reduction is one of the most efficient methods to reduce power consumption of a chip. With the galloping adoption of hardware accelerators (i.e., GPUs and FPGAs) in large datacenters and other large-scale computing infrastructures, a comprehensive evaluation of the safe voltage reduction levels for each different chip can be employed for efficient reduction of the total power. We present a survey of recent studies in voltage margins reduction at the system level for modern CPUs, GPUs and FPGAs. The pessimistic voltage guardbands inserted by the silicon vendors can be exploited in all devices for significant power savings. Voltage reduction can reach 12% in multicore CPUs, 20% in manycore GPUs and 39% in FPGAs.

Predicting performance degradation of a GPU application when it is co-located with other applications on a spatial multitasking GPU without prior application knowledge is essential in public Clouds. Prior work mainly targets CPU co-location, and is inaccurate and/or inefficient for predicting performance of applications at co-location on spatial multitasking GPUs. Our investigation shows that hardware event statistics caused by co-located applications, which can be collected with negligible overhead, strongly correlate with their slowdowns. Based on this observation, we present Themis, an online slowdown predictor that can precisely and efficiently predict application slowdown without prior application knowledge. We first train a precise slowdown model offline using hardware event statistics collected from representative co-locations. When new applications co-run, Themis collects event statistics and predicts their slowdowns simultaneously. Our evaluation shows that Themis has negligible runtime overhead and can precisely predict application-level slowdown with prediction error smaller than 9.5%. Based on Themis, we also implement an SM allocation engine to rein in application slowdown at co-location. Case studies show that the engine successfully enforces fair sharing and QoS.

Parallel computers now start to adopt Bandwidth-Asymmetric Memory architecture that consists of traditional DRAM memory and new High Bandwidth Memory (HBM) for high memory bandwidth. However, existing task schedulers suffer from low bandwidth usage and poor data locality problems in bandwidth-asymmetric memory architectures. To solve the two problems, we propose a Bandwidth and Locality Aware Task-stealing (BATS) system, which consists of an HBM-aware data allocator, a bandwidth-aware traffic balancer, and a hierarchical task-stealing scheduler. Leveraging compile-time code transformation and run-time data distribution, the data allocator enables HBM usage automatically without user interference. According to data access hotness, the traffic balancer migrates data to balance memory traffic across memory nodes proportional to their bandwidth. The hierarchical scheduler improves data locality at runtime without a priori program knowledge. Experiments on an Intel Knights Landing server that adopts bandwidth-asymmetric memory show that BATS reduces the execution time of memory-bound programs up to 83.5% compared with traditional task-stealing schedulers.

Emerging data analytic workloads such as graph processing, neural network and edge data preprocesing desire efficient memory read operations. Unfortunately, due to the necessity of dynamic refresh, modern DRAM systems have to stall access during refresh cycles. As DRAM device density continues to grow, refresh operations can be a crucial throughput bottleneck. To fully unleash memory access performance, we revisit conventional refresh mechanism and DRAM architecture. We propose DR refresh, a specific refresh mechanism that enable read and refresh operations to be done simultaneously. We devise DR DRAM, a specific memory hardware system that can efficiently deploy DR refresh. Unlike traditional refresh, DR explores device refresh that only refreshes a designated device at a time. Meanwhile, DR increases read efficiency by recovering the inaccessible data that resides on a device under refreshing. We also propose Hybrid Refresh Main Memory (HRMM) which can designate refresh schemes (DR or traditional refresh) in a specific memory space. We expect that our design can benefit many real-life tasks such as SPEC CPU2006, CNN, LLT and PageRank.

Voltage stacking (VS) fundamentally improves power delivery efficiency (PDE) by series-stacking multiple voltage domains to eliminate explicit step-down voltage conversion and reduce energy loss along the power delivery path. However, it suffers from aggravated supply noise, preventing its adoption in mainstream computing systems. In this paper, we investigate a practical approach to enabling efficient and reliable power delivery in voltage-stacked manycore systems that can ensure worst-case supply noise reliability without excessive costly over-design. We start by developing an analytical model to capture the essential noise behaviors in VS. It allows us to identify dominant noise contributor and derive the worst-case conditions. With this in-depth understanding, we propose a hybrid voltage regulation solution to effectively mitigate noise with worst-case guarantees. When evaluated with real-world benchmarks, our solution can achieve 93.8% power delivery efficiency, an improvement of 13.9% over the conventional baseline.

Today, many data analytic workloads such as graph processing and neural network desire efficient memory read operation. The need for preprocessing various raw data also demands enhanced memory read bandwidth. Unfortunately, due to the necessity of dynamic refresh, modern DRAM system has to stall memory access during each refresh cycle. As DRAM device density continues to grow, the refresh time also needs to extend to cover more memory rows. Consequently, DRAM refresh operation can be a crucial throughput bottleneck for memory read intensive (MRI) data processing tasks. To fully unleash the performance of these applications, we revisit conventional DRAM architecture and refresh mechanism. We propose DR DRAM, an application-specific memory design approach that makes a novel tradeoff between read and write performance. Simply put, DR has two layers of meaning: device refresh and data recovery. It aims at eliminating stall by enabling read and refresh operations to be done simultaneously. Unlike traditional schemes, DR explores device refresh that only refreshes a specific device at a time. Meanwhile, DR increases read efficiency by recovering the inaccessible data that resides on a device under refreshing. Our design can be implemented on existing redundant data storage area on DRAM. In this paper we detail DR's architecture and protocol design. We evaluate it on a cycle accurate simulator. Our results show that DR can nearly eliminate refresh overhead for memory read operation and brings up to 12% extra maximum read bandwidth and 50~60% latency improvement on present DRR4 device.

More than 20% of the available energy is lost in "the last centimeter" from the PCB board to the microprocessor chip due to inherent inefficiencies of power delivery subsystems (PDSs) in today's computing systems. By series-stacking multiple voltage domains to eliminate explicit voltage conversion and reduce loss along the power delivery path, voltage stacking (VS) is a novel configuration that can improve power delivery efficiency (PDE). However, VS suffers from aggravated levels of supply noise caused by current imbalance between the stacking layers, preventing its practical adoption in mainstream computing systems. Throughput-centric manycore architectures such as GPUs intrinsically exhibit more balanced workloads, yet suffer from lower PDE, making them ideal platforms to implement voltage stacking. In this paper, we present a cross-layer approach to practical voltage stacking implementation in GPUs. It combines circuit-level voltage regulation using distributed charge-recycling integrated voltage regulators (CR-IVRs) with architecture-level voltage smoothing guided by control theory. Our proposed voltage-stacked GPUs can eliminate 61.5% of total PDS energy loss and achieve 92.3% system-level power delivery efficiency, a 12.3% improvement over the conventional single-layer based PDS. Compared to the circuit-only solution, the cross-layer approach significantly reduces the implementation cost of voltage stacking (88% reduction in area overhead) without compromising supply reliability under worst-case scenarios and across a wide range of real-world benchmarks. In addition, we demonstrate that the cross-layer solution not only complements on-chip CR-IVRs to transparently manage current imbalance and restore stable layer voltages, but also serves as a seamless interface to accommodate higher-level power optimization techniques, traditionally thought to be incompatible with a VS configuration.

Despite being employed in burgeoning efforts to improve power delivery efficiency, integrated voltage regulators (IVRs) have yet to be evaluated in a rigorous, systematic, or quantitative manner. To fulfill this need, we present Ivory, a high-level design space exploration tool capable of providing accurate conversion efficiency, static performance characteristics, and dynamic transient responses of an IVR-enabled power delivery subsystem (PDS), enabling rapid trade-off exploration at early design stage, approximately 1000x faster than SPICE simulation. We demonstrate and validate Ivory with a wide spectrum of IVR topologies. In addition, we present a case study using Ivory to reveal the optimal PDS configurations, with underlying power break-downs and area overheads for the GPU manycore architecture, which has yet to embrace IVRs.

Energy efficiency is undoubtedly important for GPU architectures. Besides the traditionally explored energy-efficiency optimization techniques, exploiting the supply voltage guard-band remains a promising yet unexplored opportunity. Our hardware measurements show that up to 23% of the nominal supply voltage can be eliminated to improve CPU energy efficiency by as much as 25%. The key obstacle for exploiting this opportunity lies in understanding the characteristics and root causes of large voltage droops in GPU architectures and subsequently smoothing them away without severe performance penalties. The CPU's manycore nature complicates the voltage noise phenomenon, and its distinctive architecture features from the CPU necessitate a CPU-specific voltage noise analysis. In this paper, we make the following contributions. First, we provide a voltage noise categorization framework to identify, characterize, and understand voltage noise in the manycore CPU architecture. Second, we perform a microarchitecture-level voltage-droop root-cause analysis for the two major droop types we identify, namely the local first-order droop and the global second-order droop. Third, on the basis of our categorization and characterization, we propose a hierarchical voltage smoothing mechanism that mitigates each type of voltage droop. Our evaluation shows it can reduce up to 31% worst-case droop, which translates to 11.8% core-level and 7.8% processor-level energy reduction.

Energy efficiency of GPU architectures has emerged as an important aspect of computer system design. In this paper, we explore the energy benefits of reducing the GPU chip's voltage to the safe limit, i.e. Vmin point. We perform such a study on several commercial off-the-shelf GPU cards. We find that there exists about 20% voltage guardband on those GPUs spanning two architectural generations, which, if "eliminated" completely, can result in up to 25% energy savings on one of the studied GPU cards. The exact improvement magnitude depends on the program's available guardband, because our measurement results unveil a program dependent Vmin behavior across the studied programs. We make fundamental observations about the program-dependent Vmin behavior. We experimentally determine that the voltage noise has a larger impact on Vmin compared to the process and temperature variation, and the activities during the kernel execution cause large voltage droops. From these findings, we show how to use a kernel's microarchitectural performance counters to predict its Vmin value accurately. The average and maximum prediction errors are 0.5% and 3%, respectively. The accurate Vmin prediction opens up new possibilities of a cross-layer dynamic guardbanding scheme for GPUs, in which software predicts and manages the voltage guardband, while the functional correctness is ensured by a hardware safety net mechanism.

The traditional guardbanding approach to ensure processor reliability is becoming obsolete because it always over-provisions voltage and wastes a lot of energy. As a next-generation alternative, adaptive guardbanding dynamically adjusts chip clock frequency and voltage based on timing margin measured at runtime. With adaptive guardbanding, voltage guardband is only provided when needed, thereby promising significant energy efficiency improvement.

Voltage noise is a major obstacle in improving processor energy efficiency because it necessitates large operating voltage guardbands that increase overall power consumption and limit peak performance. Identifying the leading root causes of voltage noise is essential to minimize the unnecessary guardband and maximize the overall energy efficiency. We provide the first-ever modeling and characterization of voltage noise in GPUs based on a new simulation infrastructure called GPUVolt. Using it, we identify the key intracore microarchitectural components (e.g., the register file and special functional units) that significantly impact the GPU's voltage noise. We also demonstrate that intercore-aligned microarchitectural activity detrimentally impacts the chip- wide worst-case voltage droops. On the basis of these findings, we propose a combined register-file and execution-unit throttling mechanism that smooths GPU voltage noise and reduces the guardband requirement by as much as 29%.

General-purpose GPUs (GPGPUs) are becoming prevalent in mainstream computing, and performance per watt has emerged as a more crucial evaluation metric than peak performance. As such, GPU architects require robust tools that will enable them to quickly explore new ways to optimize GPGPUs for energy efficiency. We propose a new GPGPU power model that is configurable, capable of cycle-level calculations, and carefully validated against real hardware measurements. To achieve configurability, we use a bottom-up methodology and abstract parameters from the microarchitectural components as the model's inputs. We developed a rigorous suite of 80 microbenchmarks that we use to bound any modeling uncertainties and inaccuracies. The power model is comprehensively validated against measurements of two commercially available GPUs, and the measured error is within 9.9% and 13.4% for the two target GPUs (GTX 480 and Quadro FX5600). The model also accurately tracks the power consumption trend over time. We integrated the power model with the cycle-level simulator GPGPU-Sim and demonstrate the energy savings by utilizing dynamic voltage and frequency scaling (DVFS) and clock gating. Traditional DVFS reduces GPU energy consumption by 14.4% by leveraging within-kernel runtime variations. More finer-grained SM cluster-level DVFS improves the energy savings from 6.6% to 13.6% for those benchmarks that show clustered execution behavior. We also show that clock gating inactive lanes during divergence reduces dynamic power by 11.2%.

As GPU's compute capabilities grow, their memory hierarchy increasingly becomes a bottleneck. Current GPU memory hierarchies use coarse-grained memory accesses to exploit spatial locality, maximize peak bandwidth, simplify control, and reduce cache meta-data storage. These coarse-grained memory accesses, however, are a poor match for emerging GPU applications with irregular control flow and memory access patterns. Meanwhile, the massive multi-threading of GPUs and the simplicity of their cache hierarchies make CPU-specific memory system enhancements ineffective for improving the performance of irregular GPU applications. We design and evaluate a locality-aware memory hierarchy for throughput processors, such as GPUs. Our proposed design retains the advantages of coarse-grained accesses for spatially and temporally local programs while permitting selective fine-grained access to memory. By adaptively adjusting the access granularity, memory bandwidth and energy are reduced for data with low spatial/temporal locality without wasting control overheads or prefetching potential for data with high spatial locality. As such, our locality-aware memory hierarchy improves GPU performance, energy-efficiency, and memory throughput for a large range of applications.