arXiv subjects

Shinya Takamaeda-Yamazaki

Publications and source records attributed to Shinya Takamaeda-Yamazaki.

At least 19 recordsLinked to original sources

arXiv2026

Accelerating LLM Decoding with In-DRAM Parallelism

Autoregressive decoding of large language models (LLMs) on edge devices is dominated by memory-bound general matrix--vector multiplication (GeMV) operations. Processing-using-DRAM (PUD) is a promising computing paradigm that can alleviate this bottleneck by reducing processor--memory data transfers and enabling massively parallel computation within DRAM (e.g., across 65,536 columns). However, weight matrices in individual LLM layers (e.g., a QKV matrix in Llama3 8B with 6,144 rows) are typically too small to fully utilize parallelism in PUD. To address this challenge, we propose an algorithm--system co-design for PUD-based LLM inference. At the algorithm level, we enable a subsequent projection to be packed with the current projection by reformulating it as a computation on the current residual state plus a correction for the residual update. Layer redundancy allows this correction to be approximated using fewer bit-serial steps. At the system level, we tailor data mapping and execution to improve utilization of either columns within DRAM subarrays (SIMD) or independently controlled banks (MIMD). Overall, we achieves up to $1.41\times$ end-to-end decoding speedup over the PUD baseline, corresponding to a $2.45\times$ speedup over processor-based execution.

cs.AR↗

arXiv2026

RAM-H1200: A Unified Evaluation and Dataset on Hand Radiographs for Rheumatoid Arthritis

Rheumatoid arthritis (RA) assessment from hand radiographs requires multi-level analysis and modeling of anatomical structures and fine-grained local pathological changes. However, existing public resources do not support such unified multi-level analysis, often lacking full-hand coverage, fine-grained annotations, and consistent integration with clinical scoring systems. In particular, annotations that enable quantitative analysis of bone erosion (BE) remain scarce. RAM-H1200 contains 1,200 hand radiographs collected from six medical centers, with multi-level annotations including (i) whole-hand bone structure instance segmentation, (ii) pixel-level BE masks, (iii) SvdH-defined joint regions of interest, and (iv) joint-level SvdH scores for both BE and joint space narrowing (JSN). It is designed to evaluate whether models can jointly capture anatomical structure, localized erosive pathology, and clinically standardized RA severity from hand radiographs. The proposed BE masks enable, for the first time, quantitative BE analysis beyond coarse categorical grading by providing explicit spatial supervision for lesion extent and morphology. To our knowledge, RAM-H1200 is the first public large-scale benchmark that jointly supports whole-hand bone structure instance segmentation, pixel-level BE delineation, and clinically grounded joint-level SvdH scoring for both BE and JSN. Results across benchmark tasks show that anatomical modeling is substantially more mature than quantitative BE analysis: whole-hand bone segmentation achieves strong performance, whereas BE segmentation remains a major open challenge. By unifying anatomical structure modeling, quantitative lesion analysis, and clinically grounded SvdH scoring, RAM-H1200 provides a single benchmark for comprehensive RA analysis on hand radiographs.

cs.CV↗

arXiv2026

TRIP-Shapley: Propagation-Aware Contribution Attribution Framework for Peer-to-Peer Federated Learning

Contribution attribution is a fundamental problem in federated learning (FL), yet it remains challenging in peer-to-peer FL (P2PFL), where participant influence propagates through repeated model exchanges. Recent work, DICE, addresses this issue by directly tracking multi-round propagation paths. However, its practical deployment is hindered by two challenges: (i) communication explosion caused by multi-hop propagation path tracking, and (ii) vulnerability to misreporting attacks. To overcome these challenges, we consider contribution attribution in P2PFL with a model-free coordinator used solely for auditing. Based on this setting, we propose TRIP-Shapley, a propagation-aware Shapley-inspired surrogate for contribution attribution. Instead of explicitly tracking multi-round paths, TRIP-Shapley computes round-wise contribution scores independently and propagates only scalar contribution values. This enables attribution of influence accumulated throughout the entire training process while maintaining practical communication costs. In addition, we introduce lightweight mechanisms to detect and mitigate misreporting attacks. Rather than theoretically approximating Shapley values, TRIP-Shapley is empirically evaluated against Shapley values as a reference. Experiments demonstrate communication savings of up to $45.8\times$ relative to the compared DICE configurations, with favorable cost--reference-alignment trade-offs overall, although alignment is weaker in some FMNIST and 25-client settings. The proposed defenses reduce contribution inflation under the two evaluated attacks.

cs.LG↗

arXiv2026

TALLOP: Joint Optimization of Local Training and Communication Compression for Time-Constrained Federated Learning

Minimizing the wall-clock time required to reach a target accuracy is crucial in federated learning (FL), where many applications require frequent model updates as data distributions evolve over time. Nevertheless, in heterogeneous environments, slow clients with limited computational and communication capabilities can delay global synchronization, forcing faster clients to wait and increasing the overall training time. To mitigate this straggler effect, recent studies have introduced per-round time constraints and adapted client workloads within the available time. However, these approaches primarily focus on local computation, whereas communication can become a major bottleneck in communication-limited environments, motivating the joint optimization of training workload and communication compression. Although previous studies have explored such joint optimization, determining an effective training--communication balance under a fixed per-round time budget remains challenging, as the resulting model performance depends on both training progress and compression-induced degradation. We therefore propose TALLOP, which jointly optimizes the local training workload and communication compression under per-round time constraints. TALLOP predicts the training loss achievable under different computation budgets from the early-stage mini-batch loss trajectory (LENS), and estimates the degradation caused by communication compression using a compression-error model (CLEN). TALLOP combines these predictions to select the training-communication allocation that minimizes the post-compression loss under the given round time constraint. Experiments in heterogeneous FL environments show that TALLOP generally achieves competitive or improved time-to-accuracy compared with approaches that optimize only local training or neither, although the benefit over training-only adaptation depends on the setting.

cs.LG↗

arXiv2026

Annotation-Driven Migration of CUDA Programs to Tenstorrent Blackhole

Tenstorrent Blackhole combines distributed local memories, explicit inter-core communication, and Tensix cores decoupling data movement from computation. CUDA offers a substantial HPC software base but leaves physical data placement and scheduling largely implicit. Migrating CUDA HPC kernels to Blackhole requires spatial mapping across cores and per-core coordination of compute and data-movement kernels. We present an MLIR-based compiler deriving data placement, inter-core communication, and tile computation from a statically shaped affine CUDA subset. Declarative annotations express choices not fixed by the source, including compute/DM operation placement and streaming granularity. We evaluate it on Gaussian elimination, five-point stencil, and symmetric rank-k update. Relative to our compiler's default realizations, the best measured configurations achieve a 4.2x speedup for BF16 Gaussian and 2.1-2.2x for FP32 Gaussian and Jacobi. A choice's performance impact can reverse with surrounding policies, precision, and loop schedule, motivating comparison of alternative per-core realizations rather than independent policy selection.

cs.DC↗

arXiv2026

CARVY-FL: Client Anticlustering for Robust Voting in Provably Secure Federated Learning

Federated learning (FL) enables collaborative training without directly sharing raw data, but remains vulnerable to malicious clients. Voting-based FL improves robustness by partitioning clients into groups, training one model per group, and aggregating predictions by plurality voting. However, under class-disjoint non-IID data, distribution-oblivious grouping can yield highly variable certified accuracy (CA). We propose CARVY-FL, which estimates client distribution types from one-epoch model updates and uses anticlustering to increase within-group distributional diversity. Under a fixed grouping, CARVY-FL retains the voting-based CA guarantee while increasing vote margins. Experiments on MNIST and Fashion-MNIST show higher CA than FLCert. Under BadNets with model replacement, CARVY-FL improves the AUC of 100-ASR by 11.1% and 14.9%, respectively.

cs.CR↗

arXiv2026

Clutch: High Performance Vector-Scalar Comparison using DRAM via Chunked Temporal Coding

Vector-scalar comparison is a fundamental computation primitive that compares each element in a vector against a single scalar value. It is widely used in various data-intensive workloads from databases to machine learning. Due to its low computational intensity, its execution tends to be memory-bound, limiting the utilization of compute resources. Processing-using-DRAM (PuD) is an emerging computing paradigm that performs massively parallel bitwise operations directly inside DRAM arrays, alleviating off-chip data movement. Existing PuD-based approaches require many DRAM commands because the comparison's algorithmic complexity grows with operand bit-width in the bit-serial execution model. This command overhead becomes the dominant bottleneck, limiting application-level speedup. We propose Clutch, a data representation and comparison algorithm that accelerates vector-scalar comparisons in PuD systems with high efficiency and scalability. Clutch first uses temporal coding, encoding each vector value as a sequence of leading ones, which enables lookup-based comparison against a scalar by accessing the corresponding DRAM row. To avoid the prohibitive memory footprint of lookup tables at high precision, Clutch partitions operands into multiple multi-bit chunks, compares chunks independently using compact lookup tables, and merges the per-chunk results with a PuD-efficient procedure. By adjusting the number of chunks, Clutch provides a flexible tradeoff between throughput and memory usage. Across predicate evaluation and decision tree inference, Clutch improves end-to-end application throughput and energy efficiency by an average of 12x and 69x over highly optimized CPU and GPU execution, and by 2.9x and 3.0x over the state-of-the-art bit-serial PuD implementation. We also present the first mapping of decision tree inference to PuD execution, extending PuD to a new application domain.

cs.AR↗

arXiv2026

PuDGhost: Experimental Analysis of Computation Result Corruption in Processing-using-DRAM Operations on Real DRAM Chips and Implications for Future Systems

Processing-using-DRAM (PuD) is a promising computation paradigm that alleviates frequent data movement between main memory and processing units by using each DRAM column as a computation engine via simultaneous multiple-row activation (SiMRA). Unfortunately, DRAM density scaling may hinder PuD's benefits: denser cell arrays bring rows and columns closer, making regular DRAM operations susceptible to noise and interference from neighboring cells. Yet no prior work investigates whether interference from rows or columns not intended to participate in computation can compromise PuD robustness. In this work, we reveal PuDGhost, an interference phenomenon where a PuD operation in a given column produces erroneous results due to interference from 1) data in non-activated DRAM rows and 2) data in other columns that compute concurrently under the same SiMRA operation. PuDGhost violates the ideal picture that each column's computation depends solely on its own operand data, threatening future PuD systems. We present the first extensive characterization of PuDGhost using 96 real DDR4 DRAM chips from 12 modules, quantifying these two interference sources under various conditions. Among our 15 new empirical observations, we highlight two major results: 1) data in adjacent non-activated rows affects SiMRA outputs by up to 10% for random inputs, and 2) data in concurrently computing columns affects SiMRA outputs by up to 48% for random inputs. Guided by these findings, we propose countermeasures across multiple layers of the PuD computing stack. Specifically, we evaluate on real DDR4 DRAM chips: 1) robust column screening that reduces the risk of using unreliable columns in the presence of PuDGhost, and 2) a compute row layout that mitigates PuDGhost via dedicated rows between compute rows. Our solutions greatly improve PuD computation accuracy and provide a foundation for robust future PuD systems.

cs.AR↗

arXiv2026

Relational Hoare Logic for High-Level Synthesis of Hardware Accelerators

High-level synthesis (HLS) is a powerful tool for developing efficient hardware accelerators that rely on specialized memory systems to achieve sufficient on-chip data reuse and off-chip bandwidth utilization. However, even with HLS, designing such systems still requires careful manual tuning, as automatic optimizations provided by existing tools are highly sensitive to programming style and often lack transparency. To address these issues, we present a formal translation framework based on relational Hoare logic, which enables robust and transparent transformations. Our method recognizes complex memory access patterns in naïve HLS programs and automatically transforms them by inserting on-chip buffers to enforce linear access to off-chip memory, and by replacing non-sequential processing with stream processing, while preserving program semantics. Experiments using our prototype translator, combined with an off-the-shelf HLS compiler and a real FPGA board, have demonstrated significant performance improvements.

cs.PL↗

arXiv2025

MVDRAM: Enabling GeMV Execution in Unmodified DRAM for Low-Bit LLM Acceleration

General matrix-vector multiplication (GeMV) remains a critical latency bottleneck in large language model (LLM) inference, even with quantized low-bit models. Processing-Using-DRAM (PUD), an analog in-DRAM computing technique, has the potential to repurpose on-device DRAM as a GeMV engine, offering additional high-throughput processing capabilities to widespread consumer devices without DRAM modifications. However, applying PUD to GeMV operations in the LLM inference pipeline incurs significant overheads $\textit{before}$ and $\textit{after}$ in-DRAM computation, diminishing the benefits of its high-throughput processing capabilities. This paper presents MVDRAM, the first practical system to accelerate GeMV operations for low-bit LLM inference using unmodified DRAM. By leveraging the data sharing patterns and mathematical linearity in GeMV operations, MVDRAM orchestrates the processor and DRAM to eliminate the costs associated with pre-arranging inputs and bit-transposition of outputs required in conventional PUD approaches. Our experimental evaluation with four DDR4 DRAM modules shows that MVDRAM achieves comparable or even better inference speed than the processor-based implementation for GeMV operations in low-bit (under 4-bit) LLM. In particular, MVDRAM achieves up to 7.29$\times$ speedup and 30.5$\times$ energy efficiency for low-bit GeMV operations. For end-to-end LLM inference, MVDRAM achieves 2.18$\times$ and 1.31$\times$ throughput improvements, along with 3.04$\times$ and 2.35$\times$ energy efficiency, for 2-bit and 4-bit quantized low-bit models, respectively. MVDRAM has the potential to redefine the AI hardware landscape by demonstrating the feasibility of standard DRAM as an LLM accelerator.

cs.AR↗

arXiv2025

PUDTune: Multi-Level Charging for High-Precision Calibration in Processing-Using-DRAM

Recently, practical analog in-memory computing has been realized using unmodified commercial DRAM modules. The underlying Processing-Using-DRAM (PUD) techniques enable high-throughput bitwise operations directly within DRAM arrays. However, the presence of inherent error-prone columns hinders PUD's practical adoption. While selectively using only error-free columns would ensure reliability, this approach significantly reduces PUD's computational throughput. This paper presents PUDTune, a novel high-precision calibration technique for increasing the number of error-free columns in PUD. PUDTune compensates for errors by applying pre-identified column-specific offsets to PUD operations. By leveraging multi-level charge states of DRAM cells, PUDTune generates fine-grained and wide-range offset variations despite the limited available rows. Our experiments with DDR4 DRAM demonstrate that PUDTune increases the number of error-free columns by 1.81$\times$ compared to conventional implementations, improving addition and multiplication throughput by 1.88$\times$ and 1.89$\times$ respectively.

cs.AR↗

arXiv2025

PRIOT: Pruning-Based Integer-Only Transfer Learning for Embedded Systems

On-device transfer learning is crucial for adapting a common backbone model to the unique environment of each edge device. Tiny microcontrollers, such as the Raspberry Pi Pico, are key targets for on-device learning but often lack floating-point units, necessitating integer-only training. Dynamic computation of quantization scale factors, which is adopted in former studies, incurs high computational costs. Therefore, this study focuses on integer-only training with static scale factors, which is challenging with existing training methods. We propose a new training method named PRIOT, which optimizes the network by pruning selected edges rather than updating weights, allowing effective training with static scale factors. The pruning pattern is determined by the edge-popup algorithm, which trains a parameter named score assigned to each edge instead of the original parameters and prunes the edges with low scores before inference. Additionally, we introduce a memory-efficient variant, PRIOT-S, which only assigns scores to a small fraction of edges. We implement PRIOT and PRIOT-S on the Raspberry Pi Pico and evaluate their accuracy and computational costs using a tiny CNN model on the rotated MNIST dataset and the VGG11 model on the rotated CIFAR-10 dataset. Our results demonstrate that PRIOT improves accuracy by 8.08 to 33.75 percentage points over existing methods, while PRIOT-S reduces memory footprint with minimal accuracy loss.

cs.LG↗

arXiv2025

Accelerating Elliptic Curve Point Additions on Versal AI Engine for Multi-scalar Multiplication

Multi-scalar multiplication (MSM) is crucial in cryptographic applications and computationally intensive in zero-knowledge proofs. MSM involves accumulating the products of scalars and points on an elliptic curve over a 377-bit modulus, and the Pippenger algorithm converts MSM into a series of elliptic curve point additions (PADDs) with high parallelism. This study investigates accelerating MSM on the Versal ACAP platform, an emerging hardware that employs a spatial architecture integrating 400 AI Engines (AIEs) with programmable logic and a processing system. AIEs are SIMD-based VLIW processors capable of performing vector multiply-accumulate operations, making them well-suited for multiplication-heavy workloads in PADD. Unlike simpler multiplication tasks in previous studies, cryptographic computations also require complex operations such as carry propagation. These operations necessitate architecture-aware optimizations, including intra-core dedicated coding style to fully exploit VLIW capabilities and inter-core strategy for spatial task mapping. We propose various optimizations to accelerate PADDs, including (1) algorithmic optimizations for carry propagation employing a carry-save-like technique to exploit VLIW and SIMD capabilities and (2) a comparison of four distinct spatial mappings to enhance intra- and inter-task parallelism. Our approach achieves a computational efficiency that utilizes 50.2% of the theoretical memory bandwidth and provides 568 speedup over the integrated CPU on the AIE evaluation board.

cs.AR↗

arXiv2025

Exploring the Versal AI Engine for 3D Gaussian Splatting

Dataflow-oriented spatial architectures are the emerging paradigm for higher computation performance and efficiency. AMD Versal AI Engine is a commercial spatial architecture consisting of tiles of VLIW processors supporting SIMD operations arranged in a two-dimensional mesh. The architecture requires the explicit design of task assignments and dataflow configurations for each tile to maximize performance, demanding advanced techniques and meticulous design. However, a few works revealed the performance characteristics of the Versal AI Engine through practical workloads. In this work, we provide the comprehensive performance evaluation of the Versal AI Engine using Gaussian feature computation in 3D Gaussian splatting as a practical workload, and we then propose a novel dedicated algorithm to fully exploit the hardware architecture. The computations of 3D Gaussian splatting include matrix multiplications and color computations utilizing high-dimensional spherical harmonic coefficients. These tasks are processed efficiently by leveraging the SIMD capabilities and their instruction-level parallelism. Additionally, pipelined processing is achieved by assigning different tasks to individual cores, thereby fully exploiting the spatial parallelism of AI Engines. The proposed method demonstrated a 226-fold throughput increase in simulation-based evaluation, outperforming a naive approach. These findings provide valuable insights for application development that effectively harnesses the spatial and architectural advantages of AI Engines.

cs.AR↗

arXiv2025

RustSFQ: A Domain-Specific Language for SFQ Circuit Design

Cell-based design of a single-flux-quantum (SFQ) digital circuit requires input-output consistency; every output signal must be consumed only once by the input of the following component, which is a unique constraint, unlike the traditional CMOS digital circuit design. While there are some cell libraries and simulation tools for SFQ circuit development, they do not verify the input-output consistency, and designers have significant responsibilities to ensure it manually. Additionally, designers have to carefully manage net names without unintended duplication and correct connectivity among nets in a netlist for simulations. We propose RustSFQ, a domain-specific language (DSL) embedded in Rust that automatically ensures the input-output consistency in the SFQ circuit by leveraging the ownership system of Rust. Each SFQ circuit element is represented as a function while wires are represented as instances, and the Rust compiler verifies that multiple elements do not share a single wire through the ownership system. Circuit descriptions in the RustSFQ are successfully compiled into low-level netlists for both analog and digital circuit simulations, and the DSL provides higher productivity than the conventional design flow. Using the RustSFQ, we developed an SFQ-based Reed-Solomon encoder with a 4-bit width for the first time as a case study. We confirmed that the circuit operated correctly at 10 GHz through circuit simulations.

cs.PL↗

arXiv2024

Federated Learning with Relative Fairness

This paper proposes a federated learning framework designed to achieve \textit{relative fairness} for clients. Traditional federated learning frameworks typically ensure absolute fairness by guaranteeing minimum performance across all client subgroups. However, this approach overlooks disparities in model performance between subgroups. The proposed framework uses a minimax problem approach to minimize relative unfairness, extending previous methods in distributionally robust optimization (DRO). A novel fairness index, based on the ratio between large and small losses among clients, is introduced, allowing the framework to assess and improve the relative fairness of trained models. Theoretical guarantees demonstrate that the framework consistently reduces unfairness. We also develop an algorithm, named \textsc{Scaff-PD-IA}, which balances communication and computational efficiency while maintaining minimax-optimal convergence rates. Empirical evaluations on real-world datasets confirm its effectiveness in maintaining model performance while reducing disparity.

stat.ML↗

arXiv2024

PACiM: A Sparsity-Centric Hybrid Compute-in-Memory Architecture via Probabilistic Approximation

Approximate computing emerges as a promising approach to enhance the efficiency of compute-in-memory (CiM) systems in deep neural network processing. However, traditional approximate techniques often significantly trade off accuracy for power efficiency, and fail to reduce data transfer between main memory and CiM banks, which dominates power consumption. This paper introduces a novel probabilistic approximate computation (PAC) method that leverages statistical techniques to approximate multiply-and-accumulation (MAC) operations, reducing approximation error by 4X compared to existing approaches. PAC enables efficient sparsity-based computation in CiM systems by simplifying complex MAC vector computations into scalar calculations. Moreover, PAC enables sparsity encoding and eliminates the LSB activations transmission, significantly reducing data reads and writes. This sets PAC apart from traditional approximate computing techniques, minimizing not only computation power but also memory accesses by 50%, thereby boosting system-level efficiency. We developed PACiM, a sparsity-centric architecture that fully exploits sparsity to reduce bit-serial cycles by 81% and achieves a peak 8b/8b efficiency of 14.63 TOPS/W in 65 nm CMOS while maintaining high accuracy of 93.85/72.36/66.02% on CIFAR-10/CIFAR-100/ImageNet benchmarks using a ResNet-18 model, demonstrating the effectiveness of our PAC methodology.

cs.AR↗

arXiv2023

HALO-CAT: A Hidden Network Processor with Activation-Localized CIM Architecture and Layer-Penetrative Tiling

To address the 'memory wall' problem in NN hardware acceleration, we introduce HALO-CAT, a software-hardware co-design optimized for Hidden Neural Network (HNN) processing. HALO-CAT integrates Layer-Penetrative Tiling (LPT) for algorithmic efficiency, reducing intermediate result sizes. Furthermore, the architecture employs an activation-localized computing-in-memory approach to minimize data movement. This design significantly enhances energy efficiency, achieving a 14.2x reduction in activation memory capacity and a 17.8x decrease in energy consumption, with only a 1.5% loss in accuracy, compared to traditional HNN processors.

cs.AR↗