Search

Leveraging attention sparsity to accelerate long-context large language models (LLMs) has been a hot research topic. However, current algorithms such as sparse attention or key-value (KV) cache compression tend to use a fixed budget, which presents a significant challenge during deployment because it fails to account for the dynamic nature of real-world scenarios, where the optimal balance between accuracy and efficiency can vary greatly. In this paper, we find that borrowing top-$p$ sampling (nucleus sampling) to sparse attention can surprisingly achieve adaptive budgeting. Based on this, we propose Twilight, a framework to bring adaptive sparsity to any existing sparse attention algorithm without sacrificing their accuracy. Empirical results show that Twilight can adaptively prune at most 98% of redundant tokens, leading to $15.4\times$ acceleration in self-attention operations and $3.9\times$ acceleration in end-to-end per token latency in long context LLM decoding.

Modern Deep Neural Network (DNN) accelerators are equipped with increasingly larger on-chip buffers to provide more opportunities to alleviate the increasingly severe DRAM bandwidth pressure. However, most existing research on buffer utilization still primarily focuses on single-layer dataflow scheduling optimization. As buffers grow large enough to accommodate most single-layer weights in most networks, the impact of single-layer dataflow optimization on DRAM communication diminishes significantly. Therefore, developing new paradigms that fuse multiple layers to fully leverage the increasingly abundant on-chip buffer resources to reduce DRAM accesses has become particularly important, yet remains an open challenge. To address this challenge, we first identify the optimization opportunities in DRAM communication scheduling by analyzing the drawbacks of existing works on the layer fusion paradigm and recognizing the vast optimization potential in scheduling the timing of data prefetching from and storing to DRAM. To fully exploit these optimization opportunities, we develop a Tensor-centric Notation and its corresponding parsing method to represent different DRAM communication scheduling schemes and depict the overall space of DRAM communication scheduling. Then, to thoroughly and efficiently explore the space of DRAM communication scheduling for diverse accelerators and workloads, we develop an end-to-end scheduling framework, SoMa, which has already been developed into a compiler for our commercial accelerator product. Compared with the state-of-the-art (SOTA) Cocco framework, SoMa achieves, on average, a 2.11x performance improvement and a 37.3% reduction in energy cost simultaneously. Then, we leverage SoMa to study optimizations for LLM, perform design space exploration (DSE), and analyze the DRAM communication scheduling space through a practical example, yielding some..(more), Comment: Accepted by 2025 IEEE International Symposium on High-Performance Computer Architecture (HPCA)

Serving numerous users and requests concurrently requires good fairness in Large Language Models (LLMs) serving system. This ensures that, at the same cost, the system can meet the Service Level Objectives (SLOs) of more users , such as time to first token (TTFT) and time between tokens (TBT), rather than allowing a few users to experience performance far exceeding the SLOs. To achieve better fairness, the preemption-based scheduling policy dynamically adjusts the priority of each request to maintain balance during runtime. However, existing systems tend to overly prioritize throughput, overlooking the overhead caused by preemption-induced context switching, which is crucial for maintaining fairness through priority adjustments. In this work, we identify three main challenges that result in this overhead. 1) Inadequate I/O utilization. 2) GPU idleness. 3) Unnecessary I/O transmission during multi-turn conversations. Our key insight is that the block-based KV cache memory policy in existing systems, while achieving near-zero memory waste, leads to discontinuity and insufficient granularity in the KV cache memory. To respond, we introduce FastSwitch, a fairness-aware serving system that not only aligns with existing KV cache memory allocation policy but also mitigates context switching overhead. Our evaluation shows that FastSwitch outperforms the state-of-the-art LLM serving system vLLM with speedups of 1.4-11.2x across different tail TTFT and TBT.

The desires for better prediction accuracy and higher execution performance in neural networks never end. Neural architecture search (NAS) and tensor compilers are two popular techniques to optimize these two goals, but they are both limited to composing or optimizing existing manually designed operators rather than coming up with completely new designs. In this work, we explore the less studied direction of neural operator synthesis, which aims to automatically and efficiently discover novel neural operators with better accuracy and/or speed. We develop an end-to-end framework Syno, to realize practical neural operator synthesis. Syno makes use of a novel set of fine-grained primitives defined on tensor dimensions, which ensure various desired properties to ease model training, and also enable expression canonicalization techniques to avoid redundant candidates during search. Syno further adopts a novel guided synthesis flow to obtain valid operators matched with the specified input/output dimension sizes, and leverages efficient stochastic tree search algorithms to quickly explore the design space. We demonstrate that Syno discovers better operators with an average of $2.06\times$ speedup and less than $1\%$ accuracy loss, even on NAS-optimized models.

Cloud servers use accelerators for common tasks (e.g., encryption, compression, hashing) to improve CPU/GPU efficiency and overall performance. However, users' Service-level Objectives (SLOs) can be violated due to accelerator-related contention. The root cause is that existing solutions for accelerators only focus on isolation or fair allocation of compute and memory resources; they overlook the contention for communication-related resources. Specifically, three communication-induced challenges drive us to re-think the problem: (1) Accelerator traffic patterns are diverse, hard to predict, and mixed across users, (2) communication-related components lack effective low-level isolation mechanism to configure, and (3) computational heterogeneity of accelerators lead to unique relationships between the traffic mixture and the corresponding accelerator performance. The focus of this work is meeting SLOs in accelerator-rich systems. We present \design{}, treating accelerator SLO management as traffic management with proactive traffic shaping. We develop an SLO-aware protocol coupled with an offloaded interface on an architecture that supports precise and scalable traffic shaping. We guarantee accelerator SLO for various circumstances, with up to 45% tail latency reduction and less than 1% throughput variance.

I/O devices in public clouds have integrated increasing numbers of hardware accelerators, e.g., AWS Nitro, Azure FPGA and Nvidia BlueField. However, such specialized compute (1) is not explicitly accessible to cloud users with performance guarantee, (2) cannot be leveraged simultaneously by both providers and users, unlike general-purpose compute (e.g., CPUs). Through ten observations, we present that the fundamental difficulty of democratizing accelerators is insufficient performance isolation support. The key obstacles to enforcing accelerator isolation are (1) too many unknown traffic patterns in public clouds and (2) too many possible contention sources in the datapath. In this work, instead of scheduling such complex traffic on-the-fly and augmenting isolation support on each system component, we propose to model traffic as network flows and proactively re-shape the traffic to avoid unpredictable contention. We discuss the implications of our findings on the design of future I/O management stacks and device interfaces.

Printing custom DNA sequences is essential to scientific and biomedical research, but the technology can be used to manufacture plagues as well as cures. Just as ink printers recognize and reject attempts to counterfeit money, DNA synthesizers and assemblers should deny unauthorized requests to make viral DNA that could be used to ignite a pandemic. There are three complications. First, we don't need to quickly update printers to deal with newly discovered currencies, whereas we regularly learn of new viruses and other biological threats. Second, anti-counterfeiting specifications on a local printer can't be extracted and misused by malicious actors, unlike information on biological threats. Finally, any screening must keep the inspected DNA sequences private, as they may constitute valuable trade secrets. Here we describe SecureDNA, a free, privacy-preserving, and fully automated system capable of verifiably screening all DNA synthesis orders of 30+ base pairs against an up-to-date database of hazards, and its operational performance and specificity when applied to 67 million base pairs of DNA synthesized by providers in the United States, Europe, and China., Comment: Main text 10 pages, 4 figures. 5 supplementary figures. Total 21 pages. Direct correspondence to: Ivan B. Damgard (ivan@cs.au.dk), Andrew C. Yao (andrewcyao@mail.tsinghua.edu.cn), Kevin M. Esvelt (esvelt@mit.edu)

Hybrid main memory systems combine both performance and capacity advantages from heterogeneous memory technologies. With larger capacities, higher associativities, and finer granularities, hybrid memory systems currently exhibit significant metadata storage and lookup overheads for flexibly remapping data blocks between the two memory tiers. To alleviate the inefficiencies of existing designs, we propose Trimma, the combination of a multi-level metadata structure and an efficient metadata cache design. Trimma uses a multi-level metadata table to only track truly necessary address remap entries. The saved memory space is effectively utilized as extra DRAM cache capacity to improve performance. Trimma also uses separate formats to store the entries with non-identity and identity address mappings. This improves the overall remap cache hit rate, further boosting the performance. Trimma is transparent to software and compatible with various types of hybrid memory systems. When evaluated on a representative hybrid memory system with HBM3 and DDR5, Trimma achieves up to 1.68$\times$ and on average 1.33$\times$ speedup benefits, compared to state-of-the-art hybrid memory designs. These results show that Trimma effectively addresses metadata management overheads, especially for future scalable large-scale hybrid memory architectures., Comment: Accepted by PACT 2024

The multi-level design of Log-Structured Merge-trees (LSM-trees) naturally fits the tiered storage architecture: the upper levels (recently inserted/updated records) are kept in fast storage to guarantee performance while the lower levels (the majority of records) are placed in slower but cheaper storage to reduce cost. However, frequently accessed records may have been compacted and reside in slow storage. Existing algorithms are inefficient in promoting these ``hot'' records to fast storage, leading to compromised read performance. We present HotRAP, a key-value store based on RocksDB that can timely promote hot records individually from slow to fast storage and keep them in fast storage while they are hot. HotRAP uses an on-disk data structure (a specially-made LSM-tree) to track the hotness of keys and includes three pathways to ensure that hot records reach fast storage with short delays. Our experiments show that HotRAP outperforms state-of-the-art LSM-trees on tiered storage by up to 5.4$\times$ compared to the second best under read-only and read-write-balanced YCSB workloads with common access skew patterns, and by up to 1.9$\times$ compared to the second best under Twitter production workloads.

Chiplet technology enables the integration of an increasing number of transistors on a single accelerator with higher yield in the post-Moore era, addressing the immense computational demands arising from rapid AI advancements. However, it also introduces more expensive packaging costs and costly Die-to-Die (D2D) interfaces, which require more area, consume higher power, and offer lower bandwidth than on-chip interconnects. Maximizing the benefits and minimizing the drawbacks of chiplet technology is crucial for developing large-scale DNN chiplet accelerators, which poses challenges to both architecture and mapping. Despite its importance in the post-Moore era, methods to address these challenges remain scarce., Comment: Accepted by 2024 IEEE International Symposium on High-Performance Computer Architecture (HPCA)

Dataflow scheduling decisions are of vital importance to neural network (NN) accelerators. Recent scalable NN accelerators support a rich set of advanced dataflow techniques. The problems of comprehensively representing and quickly finding optimized dataflow schemes thus become significantly more complicated and challenging. In this work, we first propose comprehensive and pragmatic dataflow representations for temporal and spatial scheduling on scalable multi-node NN architectures. An informal hierarchical taxonomy highlights the tight coupling across different levels of the dataflow space as the major difficulty for fast design exploration. A set of formal tensor-centric directives accurately express various inter-layer and intra-layer schemes, and allow for quickly determining their validity and efficiency. We then build a generic, optimized, and fast dataflow solver, KAPLA, which makes use of the pragmatic directives to explore the design space with effective validity check and efficiency estimation. KAPLA decouples the upper inter-layer level for fast pruning, and solves the lower intra-layer schemes with a novel bottom-up cost descending method. KAPLA achieves within only 2.2% and 7.7% energy overheads on the result dataflow for training and inference, respectively, compared to the exhaustively searched optimal schemes. It also outperforms random and machine-learning-based approaches, with more optimized results and orders of magnitude faster search speedup.

Two-stage point-to-box network acts as a critical role in the recent popular 3D Siamese tracking paradigm, which first generates proposals and then predicts corresponding proposal-wise scores. However, such a network suffers from tedious hyper-parameter tuning and task misalignment, limiting the tracking performance. Towards these concerns, we propose a simple yet effective one-stage point-to-box network for point cloud-based 3D single object tracking. It synchronizes 3D proposal generation and center-ness score prediction by a parallel predictor without tedious hyper-parameters. To guide a task-aligned score ranking of proposals, a center-aware focal loss is proposed to supervise the training of the center-ness branch, which enhances the network's discriminative ability to distinguish proposals of different quality. Besides, we design a binary target classifier to identify target-relevant points. By integrating the derived classification scores with the center-ness scores, the resulting network can effectively suppress interference proposals and further mitigate task misalignment. Finally, we present a novel one-stage Siamese tracker OSP2B equipped with the designed network. Extensive experiments on challenging benchmarks including KITTI and Waymo SOT Dataset show that our OSP2B achieves leading performance with a considerable real-time speed.Code will be available at https://github.com/haooozi/OSP2B., Comment: Accepted to IJCAI'23. Code will be available at https://github.com/haooozi/OSP2B

The demands for higher performance and accuracy in neural networks (NNs) never end. Existing tensor compilation and Neural Architecture Search (NAS) techniques orthogonally optimize the two goals but actually share many similarities in their concrete strategies. We exploit such opportunities by combining the two into one and make a case for Kernel Architecture Search (KAS). KAS reviews NAS from a system perspective and zooms into a more fine-grained level to generate neural kernels with both high performance and good accuracy. To demonstrate the potential of KAS, we build an end-to-end framework, Canvas, to find high-quality kernels as convolution replacements. Canvas samples from a rich set of fine-grained primitives to stochastically and iteratively construct new kernels and evaluate them according to user-specified constraints. Canvas supports freely adjustable tensor dimension sizes inside the kernel and uses two levels of solvers to satisfy structural legality and fully utilize model budgets. The evaluation shows that by replacing standard convolutions with generated new kernels in common NNs, Canvas achieves average 1.5x speedups compared to the previous state-of-the-art with acceptable accuracy loss and search efficiency. Canvas verifies the practicability of KAS by rediscovering many manually designed kernels in the past and producing new structures that may inspire future machine learning innovations. For source code and implementation, we open-sourced Canvas at https://github.com/tsinghua-ideal/Canvas.

Siamese trackers based on 3D region proposal network (RPN) have shown remarkable success with deep Hough voting. However, using a single seed point feature as the cue for voting fails to produce high-quality 3D proposals. Additionally, the equal treatment of seed points in the voting process, regardless of their significance, exacerbates this limitation. To address these challenges, we propose a novel transformer-based voting scheme to generate better proposals. Specifically, a global-local transformer (GLT) module is devised to integrate object- and patch-aware geometric priors into seed point features, resulting in robust and accurate cues for offset learning of seed points. To train the GLT module, we introduce an importance prediction branch that learns the potential importance weights of seed points as a training constraint. Incorporating this transformer-based voting scheme into 3D RPN, a novel Siamese method dubbed GLT-T is developed for 3D single object tracking on point clouds. Moreover, we identify that the highest-scored proposal in the Siamese paradigm may not be the most accurate proposal, which limits tracking performance. Towards this concern, we approach the binary score prediction task as a ranking problem, and design a target-aware ranking loss and a localization-aware ranking loss to produce accurate ranking of proposals. With the ranking losses, we further present GLT-T++, an enhanced version of GLT-T. Extensive experiments on multiple benchmarks demonstrate that our GLT-T and GLT-T++ outperform state-of-the-art methods in terms of tracking accuracy while maintaining a real-time inference speed. The source code will be made available at https://github.com/haooozi/GLT-T., Comment: Need further revision

Current 3D single object tracking methods are typically based on VoteNet, a 3D region proposal network. Despite the success, using a single seed point feature as the cue for offset learning in VoteNet prevents high-quality 3D proposals from being generated. Moreover, seed points with different importance are treated equally in the voting process, aggravating this defect. To address these issues, we propose a novel global-local transformer voting scheme to provide more informative cues and guide the model pay more attention on potential seed points, promoting the generation of high-quality 3D proposals. Technically, a global-local transformer (GLT) module is employed to integrate object- and patch-aware prior into seed point features to effectively form strong feature representation for geometric positions of the seed points, thus providing more robust and accurate cues for offset learning. Subsequently, a simple yet effective training strategy is designed to train the GLT module. We develop an importance prediction branch to learn the potential importance of the seed points and treat the output weights vector as a training constraint term. By incorporating the above components together, we exhibit a superior tracking method GLT-T. Extensive experiments on challenging KITTI and NuScenes benchmarks demonstrate that GLT-T achieves state-of-the-art performance in the 3D single object tracking task. Besides, further ablation studies show the advantages of the proposed global-local transformer voting scheme over the original VoteNet. Code and models will be available at https://github.com/haooozi/GLT-T., Comment: Accepted to AAAI 2023. The source code and models will be available at https://github.com/haooozi/GLT-T

The second-order training methods can converge much faster than first-order optimizers in DNN training. This is because the second-order training utilizes the inversion of the second-order information (SOI) matrix to find a more accurate descent direction and step size. However, the huge SOI matrices bring significant computational and memory overheads in the traditional architectures like GPU and CPU. On the other side, the ReRAM-based process-in-memory (PIM) technology is suitable for the second-order training because of the following three reasons: First, PIM's computation happens in memory, which reduces data movement overheads; Second, ReRAM crossbars can compute SOI's inversion in $O\left(1\right)$ time; Third, if architected properly, ReRAM crossbars can perform matrix inversion and vector-matrix multiplications which are important to the second-order training algorithms. Nevertheless, current ReRAM-based PIM techniques still face a key challenge for accelerating the second-order training. The existing ReRAM-based matrix inversion circuitry can only support 8-bit accuracy matrix inversion and the computational precision is not sufficient for the second-order training that needs at least 16-bit accurate matrix inversion. In this work, we propose a method to achieve high-precision matrix inversion based on a proven 8-bit matrix inversion (INV) circuitry and vector-matrix multiplication (VMM) circuitry. We design \archname{}, a ReRAM-based PIM accelerator architecture for the second-order training. Moreover, we propose a software mapping scheme for \archname{} to further optimize the performance by fusing VMM and INV crossbar. Experiment shows that \archname{} can achieve an average of 115.8$\times$/11.4$\times$ speedup and 41.9$\times$/12.8$\times$energy saving compared to a GPU counterpart and PipeLayer on large-scale DNNs., Comment: 13pages, 13 figures

Siamese tracking paradigm has achieved great success, providing effective appearance discrimination and size estimation by the classification and regression. While such a paradigm typically optimizes the classification and regression independently, leading to task misalignment (accurate prediction boxes have no high target confidence scores). In this paper, to alleviate this misalignment, we propose a novel tracking paradigm, called SiamLA. Within this paradigm, a series of simple, yet effective localization-aware components are introduced, to generate localization-aware target confidence scores. Specifically, with the proposed localization-aware dynamic label (LADL) loss and localization-aware label smoothing (LALS) strategy, collaborative optimization between the classification and regression is achieved, enabling classification scores to be aware of location state, not just appearance similarity. Besides, we propose a separate localization branch, centered on a localization-aware feature aggregation (LAFA) module, to produce location quality scores to further modify the classification scores. Consequently, the resulting target confidence scores, are more discriminative for the location state, allowing accurate prediction boxes tend to be predicted as high scores. Extensive experiments are conducted on six challenging benchmarks, including GOT-10k, TrackingNet, LaSOT, TNL2K, OTB100 and VOT2018. Our SiamLA achieves state-of-the-art performance in terms of both accuracy and efficiency. Furthermore, a stability analysis reveals that our tracking paradigm is relatively stable, implying the paradigm is potential to real-world applications.

Traffic sign detection is a challenging task for the unmanned driving system, especially for the detection of multi-scale targets and the real-time problem of detection. In the traffic sign detection process, the scale of the targets changes greatly, which will have a certain impact on the detection accuracy. Feature pyramid is widely used to solve this problem but it might break the feature consistency across different scales of traffic signs. Moreover, in practical application, it is difficult for common methods to improve the detection accuracy of multi-scale traffic signs while ensuring real-time detection. In this paper, we propose an improved feature pyramid model, named AF-FPN, which utilizes the adaptive attention module (AAM) and feature enhancement module (FEM) to reduce the information loss in the process of feature map generation and enhance the representation ability of the feature pyramid. We replaced the original feature pyramid network in YOLOv5 with AF-FPN, which improves the detection performance for multi-scale targets of the YOLOv5 network under the premise of ensuring real-time detection. Furthermore, a new automatic learning data augmentation method is proposed to enrich the dataset and improve the robustness of the model to make it more suitable for practical scenarios. Extensive experimental results on the Tsinghua-Tencent 100K (TT100K) dataset demonstrate the effectiveness and superiority of the proposed method when compared with several state-of-the-art methods.

FPGAs are now used in public clouds to accelerate a wide range of applications, including many that operate on sensitive data such as financial and medical records. We present ShEF, a trusted execution environment (TEE) for cloud-based reconfigurable accelerators. ShEF is independent from CPU-based TEEs and allows secure execution under a threat model where the adversary can control all software running on the CPU connected to the FPGA, has physical access to the FPGA, and can compromise the FPGA interface logic of the cloud provider. ShEF provides a secure boot and remote attestation process that relies solely on existing FPGA mechanisms for root of trust. It also includes a Shield component that provides secure access to data while the accelerator is in use. The Shield is highly customizable and extensible, allowing users to craft a bespoke security solution that fits their accelerator's memory access patterns, bandwidth, and security requirements at minimum performance and area overheads. We describe a prototype implementation of ShEF for existing cloud FPGAs, map ShEF to a performant and secure storage application, and measure the performance benefits of customizable security using five additional accelerators.