arXiv subjects

Joo-Young Kim

Publications and source records attributed to Joo-Young Kim.

At least 37 records · Page 2Linked to original sources

arXiv2025

V-Rex: Real-Time Streaming Video LLM Acceleration via Dynamic KV Cache Retrieval

Streaming video large language models (LLMs) are increasingly used for real-time multimodal tasks such as video captioning, question answering, conversational agents, and augmented reality. However, these models face fundamental memory and computational challenges because their key-value (KV) caches grow substantially with continuous streaming video input. This process requires an iterative prefill stage, which is a unique feature of streaming video LLMs. Due to its iterative prefill stage, it suffers from significant limitations, including extensive computation, substantial data transfer, and degradation in accuracy. Crucially, this issue is exacerbated for edge deployment, which is the primary target for these models. In this work, we propose V-Rex, the first software-hardware co-designed accelerator that comprehensively addresses both algorithmic and hardware bottlenecks in streaming video LLM inference. At its core, V-Rex introduces ReSV, a training-free dynamic KV cache retrieval algorithm. ReSV exploits temporal and spatial similarity-based token clustering to reduce excessive KV cache memory across video frames. To fully realize these algorithmic benefits, V-Rex offers a compact, low-latency hardware accelerator with a dynamic KV cache retrieval engine (DRE), featuring bit-level and early-exit based computing units. V-Rex achieves unprecedented real-time of 3.9-8.3 FPS and energy-efficient streaming video LLM inference on edge deployment with negligible accuracy loss. While DRE only accounts for 2.2% power and 2.0% area, the system delivers 1.9-19.7x speedup and 3.1-18.5x energy efficiency improvements over AGX Orin GPU. This work is the first to comprehensively tackle KV cache retrieval across algorithms and hardware, enabling real-time streaming video LLM inference on resource-constrained edge devices.

eess.IV↗

arXiv2025

ABC-FHE : A Resource-Efficient Accelerator Enabling Bootstrappable Parameters for Client-Side Fully Homomorphic Encryption

As the demand for privacy-preserving computation continues to grow, fully homomorphic encryption (FHE)-which enables continuous computation on encrypted data-has become a critical solution. However, its adoption is hindered by significant computational overhead, requiring 10000-fold more computation compared to plaintext processing. Recent advancements in FHE accelerators have successfully improved server-side performance, but client-side computations remain a bottleneck, particularly under bootstrappable parameter configurations, which involve combinations of encoding, encrypt, decoding, and decrypt for large-sized parameters. To address this challenge, we propose ABC-FHE, an area- and power-efficient FHE accelerator that supports bootstrappable parameters on the client side. ABC-FHE employs a streaming architecture to maximize performance density, minimize area usage, and reduce off-chip memory access. Key innovations include a reconfigurable Fourier engine capable of switching between NTT and FFT modes. Additionally, an on-chip pseudo-random number generator and a unified on-the-fly twiddle factor generator significantly reduce memory demands, while optimized task scheduling enhances the CKKS client-side processing, achieving reduced latency. Overall, ABC-FHE occupies a die area of 28.638 mm2 and consumes 5.654 W of power in 28 nm technology. It delivers significant performance improvements, achieving a 1112x speed-up in encoding and encryption execution time compared to a CPU, and 214x over the state-of-the-art client-side accelerator. For decoding and decryption, it achieves a 963x speed-up over the CPU and 82x over the state-of-the-art accelerator.

cs.AR↗

arXiv2025

AoP-SAM: Automation of Prompts for Efficient Segmentation

The Segment Anything Model (SAM) is a powerful foundation model for image segmentation, showing robust zero-shot generalization through prompt engineering. However, relying on manual prompts is impractical for real-world applications, particularly in scenarios where rapid prompt provision and resource efficiency are crucial. In this paper, we propose the Automation of Prompts for SAM (AoP-SAM), a novel approach that learns to generate essential prompts in optimal locations automatically. AoP-SAM enhances SAM's efficiency and usability by eliminating manual input, making it better suited for real-world tasks. Our approach employs a lightweight yet efficient Prompt Predictor model that detects key entities across images and identifies the optimal regions for placing prompt candidates. This method leverages SAM's image embeddings, preserving its zero-shot generalization capabilities without requiring fine-tuning. Additionally, we introduce a test-time instance-level Adaptive Sampling and Filtering mechanism that generates prompts in a coarse-to-fine manner. This notably enhances both prompt and mask generation efficiency by reducing computational overhead and minimizing redundant mask refinements. Evaluations of three datasets demonstrate that AoP-SAM substantially improves both prompt generation efficiency and mask generation accuracy, making SAM more effective for automated segmentation tasks.

cs.CV↗

arXiv2025

Oaken: Fast and Efficient LLM Serving with Online-Offline Hybrid KV Cache Quantization

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.

cs.AR↗

arXiv2025

LightNobel: Improving Sequence Length Limitation in Protein Structure Prediction Model via Adaptive Activation Quantization

Recent advances in Protein Structure Prediction Models (PPMs), such as AlphaFold2 and ESMFold, have revolutionized computational biology by achieving unprecedented accuracy in predicting three-dimensional protein folding structures. However, these models face significant scalability challenges, particularly when processing proteins with long amino acid sequences (e.g., sequence length > 1,000). The primary bottleneck that arises from the exponential growth in activation sizes is driven by the unique data structure in PPM, which introduces an additional dimension that leads to substantial memory and computational demands. These limitations have hindered the effective scaling of PPM for real-world applications, such as analyzing large proteins or complex multimers with critical biological and pharmaceutical relevance. In this paper, we present LightNobel, the first hardware-software co-designed accelerator developed to overcome scalability limitations on the sequence length in PPM. At the software level, we propose Token-wise Adaptive Activation Quantization (AAQ), which leverages unique token-wise characteristics, such as distogram patterns in PPM activations, to enable fine-grained quantization techniques without compromising accuracy. At the hardware level, LightNobel integrates the multi-precision reconfigurable matrix processing unit (RMPU) and versatile vector processing unit (VVPU) to enable the efficient execution of AAQ. Through these innovations, LightNobel achieves up to 8.44x, 8.41x speedup and 37.29x, 43.35x higher power efficiency over the latest NVIDIA A100 and H100 GPUs, respectively, while maintaining negligible accuracy loss. It also reduces the peak memory requirement up to 120.05x in PPM, enabling scalable processing for proteins with long sequences.

cs.AR↗

arXiv2025

SCRec: A Scalable Computational Storage System with Statistical Sharding and Tensor-train Decomposition for Recommendation Models

Deep Learning Recommendation Models (DLRMs) play a crucial role in delivering personalized content across web applications such as social networking and video streaming. However, with improvements in performance, the parameter size of DLRMs has grown to terabyte (TB) scales, accompanied by memory bandwidth demands exceeding TB/s levels. Furthermore, the workload intensity within the model varies based on the target mechanism, making it difficult to build an optimized recommendation system. In this paper, we propose SCRec, a scalable computational storage recommendation system that can handle TB-scale industrial DLRMs while guaranteeing high bandwidth requirements. SCRec utilizes a software framework that features a mixed-integer programming (MIP)-based cost model, efficiently fetching data based on data access patterns and adaptively configuring memory-centric and compute-centric cores. Additionally, SCRec integrates hardware acceleration cores to enhance DLRM computations, particularly allowing for the high-performance reconstruction of approximated embedding vectors from extremely compressed tensor-train (TT) format. By combining its software framework and hardware accelerators, while eliminating data communication overhead by being implemented on a single server, SCRec achieves substantial improvements in DLRM inference performance. It delivers up to 55.77$\times$ speedup compared to a CPU-DRAM system with no loss in accuracy and up to 13.35$\times$ energy efficiency gains over a multi-GPU system.

cs.AR↗

arXiv2025

ADOR: A Design Exploration Framework for LLM Serving with Enhanced Latency and Throughput

The growing adoption of Large Language Models (LLMs) across various domains has driven the demand for efficient and scalable AI-serving solutions. Deploying LLMs requires optimizations to manage their significant computational and data demands. The prefill stage processes large numbers of input tokens in parallel, increasing computational load, while the decoding stage relies heavily on memory bandwidth due to the auto-regressive nature of LLMs. Current hardware, such as GPUs, often fails to balance these demands, leading to inefficient utilization. While batching improves hardware efficiency, it delays response times, degrading Quality-of-Service (QoS). This disconnect between vendors, who aim to maximize resource efficiency, and users, who prioritize low latency, highlights the need for a better solution. To address this, we propose ADOR, a framework that automatically identifies and recommends hardware architectures tailored to LLM serving. By leveraging predefined architecture templates specialized for heterogeneous dataflows, ADOR optimally balances throughput and latency. It efficiently explores design spaces to suggest architectures that meet the requirements of both vendors and users. ADOR demonstrates substantial performance improvements, achieving 2.51x higher QoS and 4.01x better area efficiency compared to the A100 at high batch sizes, making it a robust solution for scalable and cost-effective LLM serving.

cs.AR↗

arXiv2025

APINT: A Full-Stack Framework for Acceleration of Privacy-Preserving Inference of Transformers based on Garbled Circuits

As the importance of Privacy-Preserving Inference of Transformers (PiT) increases, a hybrid protocol that integrates Garbled Circuits (GC) and Homomorphic Encryption (HE) is emerging for its implementation. While this protocol is preferred for its ability to maintain accuracy, it has a severe drawback of excessive latency. To address this, existing protocols primarily focused on reducing HE latency, thus making GC the new latency bottleneck. Furthermore, previous studies only focused on individual computing layers, such as protocol or hardware accelerator, lacking a comprehensive solution at the system level. This paper presents APINT, a full-stack framework designed to reduce PiT's overall latency by addressing the latency problem of GC through both software and hardware solutions. APINT features a novel protocol that reallocates possible GC workloads to alternative methods (i.e., HE or standard matrix operation), substantially decreasing the GC workload. It also suggests GC-friendly circuit generation that reduces the number of AND gates at the most, which is the expensive operator in GC. Furthermore, APINT proposes an innovative netlist scheduling that combines coarse-grained operation mapping and fine-grained scheduling for maximal data reuse and minimal dependency. Finally, APINT's hardware accelerator, combined with its compiler speculation, effectively resolves the memory stall issue. Putting it all together, APINT achieves a remarkable end-to-end reduction in latency, outperforming the existing protocol on CPU platform by 12.2x online and 2.2x offline. Meanwhile, the APINT accelerator not only reduces its latency by 3.3x but also saves energy consumption by 4.6x while operating PiT compared to the state-of-the-art GC accelerator.

cs.AR↗

arXiv2025

RED: Energy Optimization Framework for eDRAM-based PIM with Reconfigurable Voltage Swing and Retention-aware Scheduling

In the era of artificial intelligence (AI), Transformer demonstrates its performance across various applications. The excessive amount of parameters incurs high latency and energy overhead when processed in the von Neumann architecture. Processing-in-memory (PIM) has shown the potential in accelerating data-intensive applications by reducing data movement. While previous works mainly optimize the computational part of PIM to enhance energy efficiency, the importance of memory design, which consumes the most power in PIM, has been rather neglected. In this work, we present RED, an energy optimization framework for eDRAM-based PIM. We first analyze the PIM operations in eDRAM, obtaining two key observations: 1) memory access energy consumption is predominant in PIM, and 2) read bitline (RBL) voltage swing, sense amplifier power, and retention time are in trade-off relations. Leveraging them, we propose a novel reconfigurable eDRAM and retention-aware scheduling that minimizes the runtime energy consumption of the eDRAM macro. The framework pinpoints the optimal operating point by pre-estimating energy consumption across all possible tiling schemes and memory operations. Then, the reconfigurable eDRAM controls the RBL voltage swing at runtime according to the scheduling, optimizing the memory access power. Moreover, RED employs refresh skipping and sense amplifier power gating to mitigate the energy consumption overhead coming from the trade-off relation. Finally, the RED framework achieves up to 3.05x higher energy efficiency than the prior SRAM-based PIM, reducing the energy consumption of eDRAM macro up to 74.88% with reconfigurable eDRAM and optimization schemes, requiring only 3.5% area and 0.77% energy overhead for scheduling.

cs.AR↗

arXiv2025

Securing DRAM at Scale: ARFM-Driven Row Hammer Defense with Unveiling the Threat of Short tRC Patterns

To address the issue of powerful row hammer (RH) attacks, our study involved an extensive analysis of the prevalent attack patterns in the field. We discovered a strong correlation between the timing and density of the active-to-active command period, ${tRC}$, and the likelihood of RH attacks. In this paper, we introduce MARC, an innovative ARFM-driven RH mitigation IP that significantly reinforces existing RH mitigation IPs. MARC dynamically adjusts the frequency of RFM in response to the severity of the RH attack environment, offering a tailored security solution that not only detects the threats but also adapts to varying threat levels. MARC's detection mechanism has demonstrated remarkable efficiency, identifying over 99\% of attack patterns. Moreover, MARC is designed as a compact hardware module, facilitating tight integration either on the memory controller-side or DRAM-side within the memory system. It only occupies a negligible hardware area of 3363~\textit{$μm^2$}. By activating ARFM based on MARC's detection, the additional energy overhead is also negligible in normal workloads. We conduct experiments to compare the highest row count throughout the patterns, defined as max exposure, between the vanilla RH mitigation IPs and the MARC-enhanced versions of the same IPs, focusing on both DRAM-side and memory controller-side. On the DRAM-side, MARC + probabilistic scheme and MARC + counter-based tracking scheme achieve 8.1$\times$ and 1.5$\times$ improvement in max exposure ratio compared to the vanilla IPs, respectively. On the memory controller-side, the MARC + PARA and MARC + Graphene achieve 50$\times$ and 5.7$\times$ improvement in max exposure ratio compared to the vanilla IPs, respectively. MARC ensures optimal security without sacrificing system performance, making MARC a pioneering solution in the realm of RH attack mitigation.

cs.CR↗

arXiv2025

EXION: Exploiting Inter- and Intra-Iteration Output Sparsity for Diffusion Models

Over the past few years, diffusion models have emerged as novel AI solutions, generating diverse multi-modal outputs from text prompts. Despite their capabilities, they face challenges in computing, such as excessive latency and energy consumption due to their iterative architecture. Although prior works specialized in transformer acceleration can be applied, the iterative nature of diffusion models remains unresolved. In this paper, we present EXION, the first SW-HW co-designed diffusion accelerator that solves the computation challenges by exploiting the unique inter- and intra-iteration output sparsity in diffusion models. To this end, we propose two SW-level optimizations. First, we introduce the FFN-Reuse algorithm that identifies and skips redundant computations in FFN layers across different iterations (inter-iteration sparsity). Second, we use a modified eager prediction method that employs two-step leading-one detection to accurately predict the attention score, skipping unnecessary computations within an iteration (intra-iteration sparsity). We also introduce a novel data compaction mechanism named ConMerge, which can enhance HW utilization by condensing and merging sparse matrices into compact forms. Finally, it has a dedicated HW architecture that supports the above sparsity-inducing algorithms, translating high output sparsity into improved energy efficiency and performance. To verify the feasibility of the EXION, we first demonstrate that it has no impact on accuracy in various types of multi-modal diffusion models. We then instantiate EXION in both server- and edge-level settings and compare its performance against GPUs with similar specifications. Our evaluation shows that EXION achieves dramatic improvements in performance and energy efficiency by 3.2-379.3x and 45.1-3067.6x compared to a server GPU and by 42.6-1090.9x and 196.9-4668.2x compared to an edge GPU.

cs.AR↗

arXiv2024

HURRY: Highly Utilized, Reconfigurable ReRAM-based In-situ Accelerator with Multifunctionality

Resistive random-access memory (ReRAM) crossbar arrays are suitable for efficient inference computations in neural networks due to their analog general matrix-matrix multiplication (GEMM) capabilities. However, traditional ReRAM-based accelerators suffer from spatial and temporal underutilization. We present HURRY, a reconfigurable and multifunctional ReRAM-based in-situ accelerator. HURRY uses a block activation scheme for concurrent activation of dynamically sized ReRAM portions, enhancing spatial utilization. Additionally, it incorporates functional blocks for convolution, ReLU, max pooling, and softmax computations to improve temporal utilization. System-level scheduling and data mapping strategies further optimize performance. Consequently, HURRY achieves up to 3.35x speedup, 5.72x higher energy efficiency, and 7.91x greater area efficiency compared to current ReRAM-based accelerators.

cs.AR↗

arXiv2024

LPU: A Latency-Optimized and Highly Scalable Processor for Large Language Model Inference

The explosive arrival of OpenAI's ChatGPT has fueled the globalization of large language model (LLM), which consists of billions of pretrained parameters that embodies the aspects of syntax and semantics. HyperAccel introduces latency processing unit (LPU), a latency-optimized and highly scalable processor architecture for the acceleration of LLM inference. LPU perfectly balances the memory bandwidth and compute logic with streamlined dataflow to maximize performance and efficiency. LPU is equipped with expandable synchronization link (ESL) that hides data synchronization latency between multiple LPUs. HyperDex complements LPU as an intuitive software framework to run LLM applications. LPU achieves 1.25 ms/token and 20.9 ms/token for 1.3B and 66B model, respectively, which is 2.09x and 1.37x faster than the GPU. LPU, synthesized using Samsung 4nm process, has total area of 0.824 mm2 and power consumption of 284.31 mW. LPU-based servers achieve 1.33x and 1.32x energy efficiency over NVIDIA H100 and L4 servers, respectively.

cs.AR↗

arXiv2024

SAL-PIM: A Subarray-level Processing-in-Memory Architecture with LUT-based Linear Interpolation for Transformer-based Text Generation

Text generation is a compelling sub-field of natural language processing, aiming to generate human-readable text from input words. In particular, the decoder-only generative models, such as generative pre-trained transformer (GPT), are widely used for text generation, with two major computational stages: summarization and generation. Unlike the summarization stage, which can process the input tokens in parallel, the generation stage is difficult to accelerate due to its sequential generation of output tokens through iteration. Moreover, each iteration requires reading a whole model with little data reuse opportunity. Therefore, the workload of transformer-based text generation is severely memory-bound, making the external memory bandwidth system bottleneck. In this paper, we proposed a subarray-level processing-in-memory architecture named SAL-PIM, HBM-based PIM architecture for the end-to-end acceleration of transformer-based text generation. The SAL-PIM architecture includes three architectural features. First, the SAL-PIM architecture utilizes higher internal bandwidth by integrating multiple subarray-level arithmetic logic units with optimized data mapping schemes. Second, the SAL-PIM architecture adopts LUT-based linear interpolation to perform complex non-linear functions in PIM. Third, the SAL-PIM architecture accelerates end-to-end inference on PIM in text generation. Furthermore, to validate the SAL-PIM architecture, we built cycle-accurate simulator and implemented the SAL-PIM's logic units in 28-nm CMOS technology. As a result, when the input size is from 32 to 128 and the output size is from 1 to 256, SAL-PIM achieves a maximum of 4.72 times speedup and an average of 1.83 times speedup for the text generation based on the GPT-2 medium model compared to the server-level GPU.

cs.AR↗

arXiv2023

Darwin: A DRAM-based Multi-level Processing-in-Memory Architecture for Data Analytics

Processing-in-memory (PIM) architecture is an inherent match for data analytics application, but we observe major challenges to address when accelerating it using PIM. In this paper, we propose Darwin, a practical LRDIMM-based multi-level PIM architecture for data analytics, which fully exploits the internal bandwidth of DRAM using the bank-, bank group-, chip-, and rank-level parallelisms. Considering the properties of data analytics operators and DRAM's area constraints, Darwin maximizes the internal data bandwidth by placing the PIM processing units, buffers, and control circuits across the hierarchy of DRAM. More specifically, it introduces the bank processing unit for each bank in which a single instruction multiple data (SIMD) unit handles regular data analytics operators and bank group processing unit for each bank group to handle workload imbalance in the condition-oriented data analytics operators. Furthermore, Darwin supports a novel PIM instruction architecture that concatenates instructions for multiple thread executions on bank group processing entities, addressing the command bottleneck by enabling separate control of up to 512 different in-memory processing units simultaneously. We build a cycle-accurate simulation framework to evaluate Darwin with various DRAM configurations, optimization schemes and workloads. Darwin achieves up to 14.7x speedup over the non-optimized version. Finally, the proposed Darwin architecture achieves 4.0x-43.9x higher throughput and reduces energy consumption by 85.7% than the baseline CPU system (Intel Xeon Gold 6226 + 4 channels of DDR4-2933). Compared to the state-of-the-art PIM, Darwin achieves up to 7.5x and 7.1x in the basic query operators and TPC-H queries, respectively. Darwin is based on the latest GDDR6 and requires only 5.6% area overhead, suggesting a promising PIM solution for the future main memory system.

eess.SY↗

arXiv2023

Strix: An End-to-End Streaming Architecture with Two-Level Ciphertext Batching for Fully Homomorphic Encryption with Programmable Bootstrapping

Homomorphic encryption (HE) enables computations on encrypted data by concealing information under noise for security. However, the process of bootstrapping, which resets the noise level in the ciphertext, is computationally expensive and requires a large bootstrapping key. The TFHE scheme offers a faster and programmable bootstrapping algorithm called PBS, crucial for security-focused applications like machine learning. Nevertheless, the current TFHE scheme lacks support for ciphertext packing, resulting in low throughput. This work thoroughly analyzes TFHE bootstrapping, identifies the bottleneck in GPUs caused by the blind rotation fragmentation problem, and proposes a hardware TFHE accelerator called Strix. Strix introduces a two-level batching approach to enhance the batch size in PBS, utilizes a specialized microarchitecture for efficient streaming data processing, and incorporates a fully-pipelined FFT microarchitecture to improve performance. It achieves significantly higher throughput than state-of-the-art implementations on both CPUs and GPUs, outperforming existing TFHE accelerators by a factor of 7.4.

cs.CR↗

arXiv2022

LearningGroup: A Real-Time Sparse Training on FPGA via Learnable Weight Grouping for Multi-Agent Reinforcement Learning

Multi-agent reinforcement learning (MARL) is a powerful technology to construct interactive artificial intelligent systems in various applications such as multi-robot control and self-driving cars. Unlike supervised model or single-agent reinforcement learning, which actively exploits network pruning, it is obscure that how pruning will work in multi-agent reinforcement learning with its cooperative and interactive characteristics. \par In this paper, we present a real-time sparse training acceleration system named LearningGroup, which adopts network pruning on the training of MARL for the first time with an algorithm/architecture co-design approach. We create sparsity using a weight grouping algorithm and propose on-chip sparse data encoding loop (OSEL) that enables fast encoding with efficient implementation. Based on the OSEL's encoding format, LearningGroup performs efficient weight compression and computation workload allocation to multiple cores, where each core handles multiple sparse rows of the weight matrix simultaneously with vector processing units. As a result, LearningGroup system minimizes the cycle time and memory footprint for sparse data generation up to 5.72x and 6.81x. Its FPGA accelerator shows 257.40-3629.48 GFLOPS throughput and 7.10-100.12 GFLOPS/W energy efficiency for various conditions in MARL, which are 7.13x higher and 12.43x more energy efficient than Nvidia Titan RTX GPU, thanks to the fully on-chip training and highly optimized dataflow/data format provided by FPGA. Most importantly, the accelerator shows speedup up to 12.52x for processing sparse data over the dense case, which is the highest among state-of-the-art sparse training accelerators.

cs.AR↗

arXiv2022

DFX: A Low-latency Multi-FPGA Appliance for Accelerating Transformer-based Text Generation

Transformer is a deep learning language model widely used for natural language processing (NLP) services in datacenters. Among transformer models, Generative Pre-trained Transformer (GPT) has achieved remarkable performance in text generation, or natural language generation (NLG), which needs the processing of a large input context in the summarization stage, followed by the generation stage that produces a single word at a time. The conventional platforms such as GPU are specialized for the parallel processing of large inputs in the summarization stage, but their performance significantly degrades in the generation stage due to its sequential characteristic. Therefore, an efficient hardware platform is required to address the high latency caused by the sequential characteristic of text generation. In this paper, we present DFX, a multi-FPGA acceleration appliance that executes GPT-2 model inference end-to-end with low latency and high throughput in both summarization and generation stages. DFX uses model parallelism and optimized dataflow that is model-and-hardware-aware for fast simultaneous workload execution among devices. Its compute cores operate on custom instructions and provide GPT-2 operations end-to-end. We implement the proposed hardware architecture on four Xilinx Alveo U280 FPGAs and utilize all of the channels of the high bandwidth memory (HBM) and the maximum number of compute resources for high hardware efficiency. DFX achieves 5.58x speedup and 3.99x energy efficiency over four NVIDIA V100 GPUs on the modern GPT-2 model. DFX is also 8.21x more cost-effective than the GPU appliance, suggesting that it is a promising solution for text generation workloads in cloud datacenters.

eess.SY↗