Won Woo Ro

Zhongkai Yu, Haotian Ye, Chenyang Zhou et al.

All current LLM serving systems place the GPU at the center, from production-level attention-FFN disaggregation to NVIDIA's Rubin GPU-LPU heterogeneous platform. Even academic PIM/PNM proposals still treat the GPU as the central hub for cross-device communication. Yet the GPU's compute-rich architecture is fundamentally mismatched with the memory-bound nature of decode-phase attention, inflating serving latency while wasting power and die area on idle compute units. The problem is compounded as reasoning and agentic workloads push context lengths toward one million tokens, making attention latency the primary user-facing bottleneck. To address these inefficiencies, we present AMMA, a multi-chiplet, memory-centric architecture for low-latency long-context attention. AMMA replaces GPU compute dies with HBM-PNM cubes, roughly doubling the available memory bandwidth to better serve memory-bound attention workloads. To translate this bandwidth into proportional performance gains, we introduce (i) a logic-die microarchitecture that fully exploits per-cube internal bandwidth for decode attention under a minimal power and area budget, (ii) a two-level hybrid parallelism scheme, and (iii) a reordered collective flow that reduces intra-chip die-to-die communication overhead. We further conduct a design-space exploration over per-cube compute power and intra-chip D2D link bandwidth, providing actionable guidance for hardware designers. Evaluations show that AMMA achieves 15.5X lower attention latency and 6.9X lower energy consumption compared with the NVIDIA H100.

Sungbin Kim, Hyunwuk Lee, Wonho Cho et al.

Diffusion models achieve superior performance in image generation tasks. However, it incurs significant computation overheads due to its iterative structure. To address these overheads, we analyze this iterative structure and observe that adjacent time steps in diffusion models exhibit high value similarity, leading to narrower differences between consecutive time steps. We adapt these characteristics to a quantized diffusion model and reveal that the majority of these differences can be represented with reduced bit-width, and even zero. Based on our observations, we propose the Ditto algorithm, a difference processing algorithm that leverages temporal similarity with quantization to enhance the efficiency of diffusion models. By exploiting the narrower differences and the distributive property of layer operations, it performs full bit-width operations for the initial time step and processes subsequent steps with temporal differences. In addition, Ditto execution flow optimization is designed to mitigate the memory overhead of temporal difference processing, further boosting the efficiency of the Ditto algorithm. We also design the Ditto hardware, a specialized hardware accelerator, fully exploiting the dynamic characteristics of the proposed algorithm. As a result, the Ditto hardware achieves up to 1.5x speedup and 17.74% energy saving compared to other accelerators.

Mincheol Park, Dongjin Kim, Cheonjun Park et al.

Channel pruning is widely accepted to accelerate modern convolutional neural networks (CNNs). The resulting pruned model benefits from its immediate deployment on general-purpose software and hardware resources. However, its large pruning granularity, specifically at the unit of a convolution filter, often leads to undesirable accuracy drops due to the inflexibility of deciding how and where to introduce sparsity to the CNNs. In this paper, we propose REPrune, a novel channel pruning technique that emulates kernel pruning, fully exploiting the finer but structured granularity. REPrune identifies similar kernels within each channel using agglomerative clustering. Then, it selects filters that maximize the incorporation of kernel representatives while optimizing the maximum cluster coverage problem. By integrating with a simultaneous training-pruning paradigm, REPrune promotes efficient, progressive pruning throughout training CNNs, avoiding the conventional train-prune-finetune sequence. Experimental results highlight that REPrune performs better in computer vision tasks than existing methods, effectively achieving a balance between acceleration ratio and performance retention.

Chaoyi Jiang, Sungwoo Kim, Lei Gao et al.

Masked autoregressive (MAR) models unify the strengths of masked and autoregressive generation by predicting tokens in a fixed order using bidirectional attention for image generation. While effective, MAR models suffer from significant computational overhead, as they recompute attention and feed-forward representations for all tokens at every decoding step, despite most tokens remaining semantically stable across steps. We propose a training-free generation framework MARché to address this inefficiency through two key components: cache-aware attention and selective KV refresh. Cache-aware attention partitions tokens into active and cached sets, enabling separate computation paths that allow efficient reuse of previously computed key/value projections without compromising full-context modeling. But a cached token cannot be used indefinitely without recomputation due to the changing contextual information over multiple steps. MARché recognizes this challenge and applies a technique called selective KV refresh. Selective KV refresh identifies contextually relevant tokens based on attention scores from newly generated tokens and updates only those tokens that require recomputation, while preserving image generation quality. MARché significantly reduces redundant computation in MAR without modifying the underlying architecture. Empirically, MARché achieves up to 1.7x speedup with negligible impact on image quality, offering a scalable and broadly applicable solution for efficient masked transformer generation.