Jongse Park

Donghwan Kim, Prakhar Singh, Younghoon Min et al.

Agentic AI completes tasks through iterative planning, tool use, and reasoning based on observed outcomes. Despite its popularity, its system-level behavior remains poorly understood, particularly for complex datasets and agent architectures-owing to highly non-deterministic execution, prohibitive evaluation costs, and limited visibility into proprietary models. This paper presents GAIATrace, the first token-level trace dataset of two state-of-the-art agentic systems (MiroThinker and OWL) running GAIA, a benchmark composed of a heterogeneous mix of general-purpose tasks. Unlike prior trace datasets, GAIATrace captures full reasoning tokens, task-level structures, and activities of every major participating LLMs, enabling in-depth systems research. Complementing the dataset, we present Vidur-Agent, a trace-driven simulator that can replay GAIATrace to perform reproducible, low-cost system evaluation across diverse simulated environments. Using both artifacts, we characterize how modern agentic systems handle general tasks and how various system design choices shape their behavior, yielding several unique findings.

Seock-Hwan Noh, Jahyun Koo, Seunghyun Lee et al.

Training deep neural networks (DNNs) is a computationally expensive job, which can take weeks or months even with high performance GPUs. As a remedy for this challenge, community has started exploring the use of more efficient data representations in the training process, e.g., block floating point (BFP). However, prior work on BFP-based DNN accelerators rely on a specific BFP representation making them less versatile. This paper builds upon an algorithmic observation that we can accelerate the training by leveraging multiple BFP precisions without compromising the finally achieved accuracy. Backed up by this algorithmic opportunity, we develop a flexible DNN training accelerator, dubbed FlexBlock, which supports three different BFP precision modes, possibly different among activation, weight, and gradient tensors. While several prior works proposed such multi-precision support for DNN accelerators, not only do they focus only on the inference, but also their core utilization is suboptimal at a fixed precision and specific layer types when the training is considered. Instead, FlexBlock is designed in such a way that high core utilization is achievable for i) various layer types, and ii) three BFP precisions by mapping data in a hierarchical manner to its compute units. We evaluate the effectiveness of FlexBlock architecture using well-known DNNs on CIFAR, ImageNet and WMT14 datasets. As a result, training in FlexBlock significantly improves the training speed by 1.5~5.3x and the energy efficiency by 2.4~7.0x on average compared to other training accelerators and incurs marginal accuracy loss compared to full-precision training.

Jinwoo Hwang, Minsu Kim, Daeun Kim et al.

Modern retrospective analytics systems leverage cascade architecture to mitigate bottleneck for computing deep neural networks (DNNs). However, the existing cascades suffer two limitations: (1) decoding bottleneck is either neglected or circumvented, paying significant compute and storage cost for pre-processing; and (2) the systems are specialized for temporal queries and lack spatial query support. This paper presents CoVA, a novel cascade architecture that splits the cascade computation between compressed domain and pixel domain to address the decoding bottleneck, supporting both temporal and spatial queries. CoVA cascades analysis into three major stages where the first two stages are performed in compressed domain while the last one in pixel domain. First, CoVA detects occurrences of moving objects (called blobs) over a set of compressed frames (called tracks). Then, using the track results, CoVA prudently selects a minimal set of frames to obtain the label information and only decode them to compute the full DNNs, alleviating the decoding bottleneck. Lastly, CoVA associates tracks with labels to produce the final analysis results on which users can process both temporal and spatial queries. Our experiments demonstrate that CoVA offers 4.8x throughput improvement over modern cascade systems, while imposing modest accuracy loss.

Jaehong Cho, Minsu Kim, Hyunmin Choi et al.

Recently, there has been an extensive research effort in building efficient large language model (LLM) inference serving systems. These efforts not only include innovations in the algorithm and software domains but also constitute developments of various hardware acceleration techniques. Nevertheless, there is a lack of simulation infrastructure capable of accurately modeling versatile hardware-software behaviors in LLM serving systems without extensively extending the simulation time. This paper aims to develop an effective simulation tool, called LLMServingSim, to support future research in LLM serving systems. In designing LLMServingSim, we focus on two limitations of existing simulators: (1) they lack consideration of the dynamic workload variations of LLM inference serving due to its autoregressive nature, and (2) they incur repetitive simulations without leveraging algorithmic redundancies in LLMs. To address these limitations, LLMServingSim simulates the LLM serving in the granularity of iterations, leveraging the computation redundancies across decoder blocks and reusing the simulation results from previous iterations. Additionally, LLMServingSim provides a flexible framework that allows users to plug in any accelerator compiler-and-simulation stacks for exploring various system designs with heterogeneous processors. Our experiments demonstrate that LLMServingSim produces simulation results closely following the performance behaviors of real GPU-based LLM serving system with less than 14.7% error rate, while offering 91.5x faster simulation speed compared to existing accelerator simulators.

Jaehong Cho, Hyunmin Choi, Guseul Heo et al.

Large language model (LLM) serving infrastructures are undergoing a shift toward heterogeneity and disaggregation. Modern deployments increasingly integrate diverse accelerators and near-memory processing technologies, introducing significant hardware heterogeneity, while system software increasingly separates computation, memory, and model components across distributed resources to improve scalability and efficiency. As a result, LLM serving performance is no longer determined by hardware or software choices in isolation, but by their runtime interaction through scheduling, data movement, and interconnect behavior. However, understanding these interactions remains challenging, as existing simulators lack the ability to jointly model heterogeneous hardware and disaggregated serving techniques within a unified, runtime-driven framework. This paper presents LLMServingSim 2.0, a unified system-level simulator designed to make runtime-driven hardware-software interactions in heterogeneous and disaggregated LLM serving infrastructures explicit and analyzable. LLMServingSim 2.0 embeds serving decisions and hardware behavior into a single runtime loop, enabling interaction-aware modeling of batching, routing, offloading, memory, and power. The simulator supports extensible integration of emerging accelerators and memory systems through profile-based modeling, while capturing dynamic serving behavior and system-level effects. We validate LLMServingSim 2.0 against real deployments, showing that it reproduces key performance, memory, and power metrics with an average error of 0.97%, while maintaining simulation times of around 10 minutes even for complex configurations. These results demonstrate that LLMServingSim 2.0 provides a practical bridge between hardware innovation and serving-system design, enabling systematic exploration and co-design for next-generation LLM serving infrastructures.

Jaehong Cho, Hyunmin Choi, Jongse Park

This paper introduces LLMServingSim2.0, a system simulator designed for exploring heterogeneous hardware in large-scale LLM serving systems. LLMServingSim2.0 addresses two key limitations of its predecessor: (1) integrating hardware models into system-level simulators is non-trivial due to the lack of a clear abstraction, and (2) existing simulators support only a narrow subset of serving techniques, leaving no infrastructure that captures the breadth of approaches in modern LLM serving. To overcome these issues, LLMServingSim2.0 adopts trace-driven performance modeling, accompanied by an operator-level latency profiler, enabling the integration of new accelerators with a single command. It further embeds up-to-date serving techniques while exposing flexible interfaces for request routing, cache management, and scheduling policies. In a TPU case study, our profiler requires 18.5x fewer LoC and outperforms the predecessor's hardware-simulator integration, demonstrating LLMServingSim2.0's low-effort hardware extensibility. Our experiments further show that LLMServingSim2.0 reproduces GPU-based LLM serving with 1.9% error, while maintaining practical simulation time, making it a comprehensive platform for both hardware developers and LLM service providers.

Yoonsung Kim, Changhun Oh, Jinwoo Hwang et al.

Deep neural network (DNN) video analytics is crucial for autonomous systems such as self-driving vehicles, unmanned aerial vehicles (UAVs), and security robots. However, real-world deployment faces challenges due to their limited computational resources and battery power. To tackle these challenges, continuous learning exploits a lightweight "student" model at deployment (inference), leverages a larger "teacher" model for labeling sampled data (labeling), and continuously retrains the student model to adapt to changing scenarios (retraining). This paper highlights the limitations in state-of-the-art continuous learning systems: (1) they focus on computations for retraining, while overlooking the compute needs for inference and labeling, (2) they rely on power-hungry GPUs, unsuitable for battery-operated autonomous systems, and (3) they are located on a remote centralized server, intended for multi-tenant scenarios, again unsuitable for autonomous systems due to privacy, network availability, and latency concerns. We propose a hardware-algorithm co-designed solution for continuous learning, DaCapo, that enables autonomous systems to perform concurrent executions of inference, labeling, and training in a performant and energy-efficient manner. DaCapo comprises (1) a spatially-partitionable and precision-flexible accelerator enabling parallel execution of kernels on sub-accelerators at their respective precisions, and (2) a spatiotemporal resource allocation algorithm that strategically navigates the resource-accuracy tradeoff space, facilitating optimal decisions for resource allocation to achieve maximal accuracy. Our evaluation shows that DaCapo achieves 6.5% and 5.5% higher accuracy than a state-of-the-art GPU-based continuous learning systems, Ekya and EOMU, respectively, while consuming 254x less power.

Minsu Kim, Jinwoo Hwang, Guseul Heo et al.

Learned indexes use machine learning models to learn the mappings between keys and their corresponding positions in key-value indexes. These indexes use the mapping information as training data. Learned indexes require frequent retrainings of their models to incorporate the changes introduced by update queries. To efficiently retrain the models, existing learned index systems often harness a linear algebraic QR factorization technique that performs matrix decomposition. This factorization approach processes all key-position pairs during each retraining, resulting in compute operations that grow linearly with the total number of keys and their lengths. Consequently, the retrainings create a severe performance bottleneck, especially for variable-length string keys, while the retrainings are crucial for maintaining high prediction accuracy and in turn, ensuring low query service latency. To address this performance problem, we develop an algorithm-hardware co-designed string-key learned index system, dubbed SIA. In designing SIA, we leverage a unique algorithmic property of the matrix decomposition-based training method. Exploiting the property, we develop a memoization-based incremental training scheme, which only requires computation over updated keys, while decomposition results of non-updated keys from previous computations can be reused. We further enhance SIA to offload a portion of this training process to an FPGA accelerator to not only relieve CPU resources for serving index queries (i.e., inference), but also accelerate the training itself. Our evaluation shows that compared to ALEX, LIPP, and SIndex, a state-of-the-art learned index systems, SIA-accelerated learned indexes offer 2.6x and 3.4x higher throughput on the two real-world benchmark suites, YCSB and Twitter cache trace, respectively.

Minsu Kim, Seongmin Hong, RyeoWook Ko et al.

Modern Large Language Model serving system batches multiple requests to achieve high throughput, while batching attention operations is challenging, rendering memory bandwidth a critical bottleneck. The community relies on high-end GPUs with multiple high-bandwidth memory channels. Unfortunately, HBM's high bandwidth often comes at the expense of limited memory capacity, which reduces core utilization and increases costs. Recent advancements enabling longer contexts for LLMs have substantially increased the key-value cache size, further intensifying the pressures on memory capacity. The literature has explored KV cache quantization techniques, which commonly use low bitwidth for most values, selectively using higher bitwidth for outlier values. While this approach helps achieve high accuracy and low bitwidth simultaneously, it comes with the limitation that cost for online outlier detection is excessively high, negating the advantages. We propose Oaken, an acceleration solution that achieves high accuracy and high performance simultaneously through co-designing algorithm and hardware. To effectively find a sweet spot in the accuracy-performance trade-off space of KV cache quantization, Oaken employs an online-offline hybrid approach, setting outlier thresholds offline, which are then used to determine the quantization scale online. To translate the proposed algorithmic technique into tangible performance gains, Oaken also comes with custom quantization engines and memory management units that can be integrated with any LLM accelerators. We built an Oaken accelerator on top of an LLM accelerator, LPU, and conducted a comprehensive evaluation. Our experiments show that for a batch size of 256, Oaken achieves up to 1.58x throughput improvement over NVIDIA A100 GPU, incurring a minimal accuracy loss of only 0.54\% on average, compared to state-of-the-art KV cache quantization techniques.

Wonung Kim, Yubin Lee, Yoonsung Kim et al.

Transformers are the driving force behind today's Large Language Models (LLMs), serving as the foundation for their performance and versatility. Yet, their compute and memory costs grow with sequence length, posing scalability challenges for long-context inferencing. In response, the algorithm community is exploring alternative architectures, such as state space models (SSMs), linear attention, and recurrent neural networks (RNNs), which we refer to as post-transformers. This shift presents a key challenge: building a serving system that efficiently supports both transformer and post-transformer LLMs within a unified framework. To address this challenge, we analyze the performance characteristics of transformer and post-transformer LLMs. Despite their algorithmic differences, both are fundamentally limited by memory bandwidth under batched inference due to attention in transformers and state updates in post-transformers. Further analyses suggest two additional insights: (1) state update operations, unlike attention, incur high hardware cost, making per-bank PIM acceleration inefficient, and (2) different low-precision arithmetic methods offer varying accuracy-area tradeoffs, while we identify Microsoft's MX as the Pareto-optimal choice. Building on these insights, we design Pimba as an array of State-update Processing Units (SPUs), each shared between two banks to enable interleaved access to PIM. Each SPU includes a State-update Processing Engine (SPE) that comprises element-wise multipliers and adders using MX-based quantized arithmetic, enabling efficient execution of state update and attention operations. Our evaluation shows that, compared to LLM-optimized GPU and GPU+PIM systems, Pimba achieves up to 4.1x and 2.1x higher token generation throughput, respectively.

Daeun Kim, Jinwoo Hwang, Changhun Oh et al.

Diffusion Transformer (DiT) has driven significant progress in image generation tasks. However, DiT inferencing is notoriously compute-intensive and incurs long latency even on datacenter-scale GPUs, primarily due to its iterative nature and heavy reliance on GEMM operations inherent to its encoder-based structure. To address the challenge, prior work has explored quantization, but achieving low-precision quantization for DiT inferencing with both high accuracy and substantial speedup remains an open problem. To this end, this paper proposes MixDiT, an algorithm-hardware co-designed acceleration solution that exploits mixed Microscaling (MX) formats to quantize DiT activation values. MixDiT quantizes the DiT activation tensors by selectively applying higher precision to magnitude-based outliers, which produce mixed-precision GEMM operations. To achieve tangible speedup from the mixed-precision arithmetic, we design a MixDiT accelerator that enables precision-flexible multiplications and efficient MX precision conversions. Our experimental results show that MixDiT delivers a speedup of 2.10-5.32 times over RTX 3090, with no loss in FID.

Changhun Oh, Seongryong Oh, Jinwoo Hwang et al.

3D Gaussian Splatting (3DGS) rendering in real-time on resource-constrained devices is essential for delivering immersive augmented and virtual reality (AR/VR) experiences. However, existing solutions struggle to achieve high frame rates, especially for high-resolution rendering. Our analysis identifies the sorting stage in the 3DGS rendering pipeline as the major bottleneck due to its high memory bandwidth demand. This paper presents Neo, which introduces a reuse-and-update sorting algorithm that exploits temporal redundancy in Gaussian ordering across consecutive frames, and devises a hardware accelerator optimized for this algorithm. By efficiently tracking and updating Gaussian depth ordering instead of re-sorting from scratch, Neo significantly reduces redundant computations and memory bandwidth pressure. Experimental results show that Neo achieves up to 10.0x and 5.6x higher throughput than state-of-the-art edge GPU and ASIC solution, respectively, while reducing DRAM traffic by 94.5% and 81.3%. These improvements make high-quality and low-latency on-device 3D rendering more practical.

Donghwan Kim, Xin Gu, Jinho Baek et al.

Machine learning (ML) models memorize and leak training data, causing serious privacy issues to data owners. Training algorithms with differential privacy (DP), such as DP-SGD, have been gaining attention as a solution. However, DP-SGD adds a noise at each training iteration, which degrades the accuracy of the trained model. To improve accuracy, a new family of approaches adds carefully designed correlated noises, so that noises cancel out each other across iterations. We performed an extensive characterization study of these new mechanisms, for the first time to the best of our knowledge, and show they incur non-negligible overheads when the model is large or uses large embedding tables. Motivated by the analysis, we propose Cocoon, a hardware-software co-designed framework for efficient training with correlated noises. Cocoon accelerates models with embedding tables through pre-computing and storing correlated noises in a coalesced format (Cocoon-Emb), and supports large models through a custom near-memory processing device (Cocoon-NMP). On a real system with an FPGA-based NMP device prototype, Cocoon improves the performance by 2.33-10.82x(Cocoon-Emb) and 1.55-3.06x (Cocoon-NMP).

Jinwoo Hwang, Daeun Kim, Sangyeop Lee et al.

Recently, Video-Language Models (VideoLMs) have demonstrated remarkable capabilities, offering significant potential for flexible and powerful video query systems. These models typically rely on Vision Transformers (ViTs), which process video frames individually to extract visual embeddings. However, generating embeddings for large-scale videos requires ViT inferencing across numerous frames, posing a major hurdle to real-world deployment and necessitating solutions for integration into scalable video data management systems. This paper introduces Déjà Vu, a video-language query engine that accelerates ViT-based VideoLMs by reusing computations across consecutive frames. At its core is ReuseViT, a modified ViT model specifically designed for VideoLM tasks, which learns to detect inter-frame reuse opportunities, striking an effective balance between accuracy and reuse. Although ReuseViT significantly reduces computation, these savings do not directly translate into performance gains on GPUs. To overcome this, Déjà Vu integrates memory-compute joint compaction techniques that convert the FLOP savings into tangible performance gains. Evaluations on three VideoLM tasks show that Déjà Vu accelerates embedding generation by up to a 2.64x within a 2% error bound, dramatically enhancing the practicality of VideoLMs for large-scale video analytics.

Seungbeom Choi, Sunho Lee, Yeonjae Kim et al.

As machine learning techniques are applied to a widening range of applications, high throughput machine learning (ML) inference servers have become critical for online service applications. Such ML inference servers pose two challenges: first, they must provide a bounded latency for each request to support consistent service-level objective (SLO), and second, they can serve multiple heterogeneous ML models in a system as certain tasks involve invocation of multiple models and consolidating multiple models can improve system utilization. To address the two requirements of ML inference servers, this paper proposes a new ML inference scheduling framework for multi-model ML inference servers. The paper first shows that with SLO constraints, current GPUs are not fully utilized for ML inference tasks. To maximize the resource efficiency of inference servers, a key mechanism proposed in this paper is to exploit hardware support for spatial partitioning of GPU resources. With the partitioning mechanism, a new abstraction layer of GPU resources is created with configurable GPU resources. The scheduler assigns requests to virtual GPUs, called gpu-lets, with the most effective amount of resources. The paper also investigates a remedy for potential interference effects when two ML tasks are running concurrently in a GPU. Our prototype implementation proves that spatial partitioning enhances throughput by 102.6% on average while satisfying SLOs.

Joongun Park, Seunghyo Kang, Sanghyeon Lee et al.

The widening availability of hardware-based trusted execution environments (TEEs) has been accelerating the adaptation of new applications using TEEs. Recent studies showed that a cloud application consists of multiple distributed software modules provided by mutually distrustful parties. The applications use multiple TEEs (enclaves) communicating through software-encrypted memory channels. Such execution model requires bi-directional protection: protecting the rest of the system from the enclave module with sandboxing and protecting the enclave module from a third-part module and operating systems. However, the current TEE model, such as Intel SGX, cannot efficiently represent such distributed sandbox applications. To overcome the lack of hardware supports for sandboxed TEEs, this paper proposes an extended enclave model called Stockade, which supports distributed sandboxes hardened by hardware. Stockade proposes new three key techniques. First, it extends the hardware-based memory isolation in SGX to confine a user software module only within its enclave. Second, it proposes a trusted monitor enclave that filters and validates systems calls from enclaves. Finally, it allows hardware-protected memory sharing between a pair of enclaves for efficient protected communication without software-based encryption. Using an emulated SGX platform with the proposed extensions, this paper shows that distributed sandbox applications can be effectively supported with small changes of SGX hardware.

Hardik Sharma, Jongse Park, Naveen Suda et al.

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