semi·news
Headlines要闻 / Research研究 /
Research digest · Sunday, May 24, 2026 研究摘要 · 2026年5月24日 星期日

Data Movement Is the New Compute 数据搬移就是新的算力

Five of this week's hardware papers attack data movement directly — pushing crypto into DRAM, multicasting collectives instead of duplicating sends, overlapping operators inside a single device, cutting video-DiT tokens, and shifting LLM inference accounting from prefill-bound to capacity-bound. The shared assumption: compute density is no longer the constraint the field thinks it is. 本周硬件论文中,有五篇直接针对数据搬移做文章——把密码运算下沉到DRAM、用多播替代重复发送实现集合通信、在单设备内进行算子重叠、削减video-DiT的token数、并把LLM推理记账从'prefill约束'转向'容量约束'。共识在于:计算密度已不是业界以为的那个核心瓶颈。

Look-back window: 7 days · 10 paper(s) 回溯窗口: 7天 · 10篇

Devices & Process 器件与工艺

An AI-driven robotic system for automated 2D van-der-Waals hetero-assembly 用于自动化二维范德瓦尔斯异质组装的AI驱动机器人系统 device-fab2D-materials

et al. (see paper)

arXiv:2605.20420 · 2026-05-19

2D van-der-Waals stacks (twisted bilayer graphene, hBN/MoS2, etc.) are a key materials platform for quantum simulation and exotic electronic phases, but assembly is essentially handcraft — operator-led, slow, hit-or-miss. The group automates the stacking via an AI-driven robotic system that handles flake identification, alignment, and transfer. Materials-layer infrastructure rather than a new device, but the kind of tooling that scales an entire research-method bottleneck once it works. 二维范德瓦尔斯堆叠结构(扭转双层石墨烯、hBN/MoS2等)是量子模拟与新电子相态的关键材料平台,但组装本质仍是手工——人工操作、慢、命中率低。作者通过AI驱动的机器人系统自动完成薄片识别、对位与转印。这并非新器件而是材料层基础设施,一旦走通,可释放整条研究方法瓶颈的产能。

Circuits & Architecture 电路与体系结构

Aging-aware arithmetic multipliers under NBTI and process variation 考虑NBTI老化与工艺偏差的算术乘法器设计 circuit-designreliability

et al. (see paper)

arXiv:2605.18444 · 2026-05-18

Multipliers are the workhorse of AI accelerators — systolic arrays, tensor cores, FPGAs all multiply constantly. NBTI aging shifts threshold voltages on the most-stressed transistors over years; in a multiplier that runs at near-100% utilization for the chip's life, the cumulative drift is real and pushes timing margins out of the window. The paper proposes a designable margin-and-redundancy combination targeted at AI duty cycles specifically. Worth reading because most aging-aware design work has assumed CPU-style bursty utilization, not LLM-serving sustained 90%+. 乘法器是AI加速器的主力单元——脉动阵列、Tensor Core、FPGA本质都在持续做乘法。NBTI老化在受应力最强的晶体管上长期推高阈值电压;在芯片寿命内接近满负载的乘法器中,累积漂移真实且会把时序裕度推出窗口。本文提出针对AI工作负载占空比的可设计裕度与冗余组合。值得一读:既有的老化感知设计大多默认CPU式的突发占用,而非LLM服务式的持续90%+占用。

iHAC: integrated high-availability cluster architecture with hybrid resource pooling iHAC:具备混合资源池化的集成高可用集群架构 architectureavailability

et al. (see paper)

arXiv:2605.18361 · 2026-05-18

Traditional HA clusters keep a hot standby idle for failover, which both wastes resources and creates single-point-of-failure exposure when the standby drifts out of sync. iHAC integrates compute pooling with availability so the 'standby' capacity does productive work and failover is a routing change, not a resource flip. Concrete architecture proposal with simulation; the interest for AI-serving systems is direct — keeping a hot GPU idle 'just in case' costs more than the workload that displaced it. 传统高可用集群将热备节点闲置以备故障切换,既浪费资源,又因备机长期未同步而构成新的单点风险。iHAC将算力池化与可用性整合,使'备份'容量参与生产负载,故障切换变为路由切换而非资源重置。具体架构方案配仿真;对AI服务系统直接相关——为'以防万一'保留闲置GPU的成本,常常高于被它挤掉的负载本身。

AI Accelerators & Compute-in-Memory AI加速器与存算一体

Near-memory cryptography: AES-128 and SHA-256 inside the DRAM, not on the CPU 近存储密码运算:在DRAM内而非CPU上完成AES-128与SHA-256 near-memorycryptodata-movement

et al. (see paper)

arXiv:2605.20047 · 2026-05-19

AES-128 and SHA-256 are fast on a CPU per-op but their *bandwidth* profile is terrible — they sit at the wrong end of a memory bus. The group implements both primitives inside the DRAM via near-memory processing, returning ciphertext/hashes directly without round-tripping plaintext over the CPU bus. The compute saved isn't the headline; the bandwidth saved is. For full-disk encryption and storage-tier hashing — which scale linearly with data volume — this is the right place to put the primitive. AES-128与SHA-256每次操作在CPU上很快,但其'带宽特征'糟糕——位于内存总线的不利端。作者通过DRAM内近存储处理实现两种原语,直接返回密文/哈希,避免明文在CPU总线上往返。节省的算力不是重点,节省的带宽才是。对于全盘加密与存储层哈希(其规模随数据量线性增长),把原语放在这里才是正确的位置。

ORBIS: output-guided token reduction for video Diffusion-Transformer acceleration ORBIS:面向视频DiT加速的输出引导token裁剪 video-diffusiontoken-reduction

et al. (see paper)

arXiv:2605.22015 · 2026-05-21

Video Diffusion Transformers blow up compute because 3D spatio-temporal attention scales with token-count × frames. Most token-reduction methods key on visual-similarity heuristics, which are noisy and over-prune. ORBIS uses the model's own output distribution to decide which tokens carry information for *this* generation and which are safe to drop. Practical: works on existing video-DiT checkpoints without retraining. The conceptual move — let the output decide what the input should be — is the architecturally interesting bit. 视频Diffusion Transformer因三维时空注意力随token数×帧数膨胀而算力暴涨。多数token裁剪方法依赖视觉相似度启发式,噪声大且常误删。ORBIS利用模型自身输出分布判断哪些token对当前生成实际重要、哪些可安全丢弃。实践层面:可直接用于现有video-DiT权重,无需重训。'让输出反过来决定输入应该是什么'的思路才是架构上的关键贡献。

μ-ORCA: microsecond-scale DNN inference on AMD ACAP for high-energy-physics jet tagging μ-ORCA:面向高能物理喷注识别的AMD ACAP微秒级DNN推理 acceleratorlow-latencyFPGA

et al. (see paper)

arXiv:2605.17683 · 2026-05-17

Jet-tagging at colliders needs DNN inference inside the trigger pipeline at microsecond latency on tiny inputs — exactly the case where most accelerators waste 80% of die area on instruction fetch and dataflow they don't need. The paper targets AMD's ACAP (Versal) heterogeneous fabric: tensor cores for the multiply, programmable logic for the control, NoC for the data movement. Niche application, but the systematic engineering of 'small model + extreme latency' is the relevant pattern for inference workloads pushing into agentic decode and edge. 对撞机喷注识别需在触发流水线内对小输入做微秒级DNN推理——恰是多数加速器为指令取指与不必要的数据流浪费80%晶元面积的场景。本文以AMD ACAP(Versal)异构架构为目标:Tensor Core执行乘法、可编程逻辑做控制、片上网络处理数据搬移。应用面较窄,但'小模型+极致延迟'的系统化工程,对正朝向Agentic解码与边缘扩展的推理负载,有可移植的范式意义。

AI Research (hardware-relevant) AI研究(与硬件相关)

Exploiting multicast for collective-communication acceleration in training and inference 训练与推理中利用多播加速集合通信 collective-commnetworkingtraining

et al. (see paper)

arXiv:2605.22428 · 2026-05-21

State-of-the-art AllGather and AlltoAll implementations rely on unicast writes and transmit the same data multiple times across the fabric. The paper observes that the underlying interconnects (NVLink, InfiniBand, modern Ethernet) all support hardware multicast — it's just never been used by the AI-collective implementations. Replacing unicast-by-default with multicast-where-applicable removes the duplicate transmissions at the wire level. Concrete numbers on real clusters; the punchline is that the optimization existed in the hardware for years, and the software left it on the floor. 当前最先进的AllGather/AlltoAll实现依赖单播写入,在网络结构上多次传输相同数据。作者观察到:底层互连(NVLink、InfiniBand、现代以太网)早就支持硬件多播——只是AI集合通信实现从未使用。将默认单播替换为'适用处多播',即可在线缆层面消除重复传输。论文在实际集群上给出具体数字,要点是:该优化在硬件中存在多年,软件却一直未使用。

Inference scaling for reasoning LLMs: from prefill-bound to capacity-bound 推理型LLM的推理扩展:从prefill约束转向容量约束 LLM-servingreasoningcapacity-planning

et al. (see paper)

arXiv:2605.19775 · 2026-05-19

Classic LLM-serving analysis treats prefill as compute-bound and decode as memory-bound. With reasoning-centric architectures generating chains of thought thousands of tokens long, that mental model breaks: the dominant phase is decode, the dominant cost is *capacity* (per-token KV-cache footprint × concurrent requests), and prefill is comparatively cheap. The paper formalizes the regime, gives the right bottleneck taxonomy, and notes that most existing serving optimizations were built for the wrong workload. Required reading for anyone planning capacity around reasoning-model deployment. 经典LLM服务分析将prefill视为算力受限、decode视为内存带宽受限。在生成长达数千token思维链的推理型架构下,该心智模型失效:主导阶段变为decode,主导成本是'容量'(每token的KV缓存占用×并发请求数),prefill反而相对廉价。论文形式化了这一新范式,给出正确的瓶颈分类,并指出多数现有服务优化原本针对的是另一种工作负载。所有围绕推理型模型规划产能的人都应阅读。

DynaFlow: programmable operator scheduling for transparent intra-device parallelism DynaFlow:面向透明设备内并行的可编程算子调度 schedulingintra-device-parallelism

et al. (see paper)

arXiv:2605.21603 · 2026-05-20

GPUs underutilize their hardware on most ML workloads — an attention layer leaves the tensor cores idle during softmax, a MoE layer leaves them idle during routing, etc. Overlapping operators with different resource profiles closes that gap, but every existing framework demands invasive, model-specific code changes to do it. DynaFlow exposes a programmable scheduler so the overlap is configured rather than hand-coded. Implementation is in PyTorch; the abstraction is the contribution. 在大多数ML工作负载下,GPU硬件并未充分利用——注意力层的softmax阶段Tensor Core空闲、MoE层的路由阶段Tensor Core空闲。重叠资源占用不同的算子可弥合这一差距,但现有框架需对每个模型做侵入式定制改造。DynaFlow暴露一个可编程调度器,使重叠由配置完成而非手工实现。实现位于PyTorch层;真正的贡献是该抽象本身。

EDA, Verification & Design Automation EDA与设计自动化

CPPL: an LLM-targeted intermediate language between natural language and CIRCT CPPL:位于自然语言与CIRCT之间、面向大模型的中间语言 EDALLM-for-hardware

et al. (see paper)

arXiv:2605.17892 · 2026-05-18

LLM-to-RTL has been the obvious EDA application of LLMs since GPT-4, but production usage stays low because raw LLM-emitted RTL is hard to validate and integrate with compiler-based flows (CIRCT, etc.). CPPL is a typed prompt-language at the right level of abstraction for an LLM to emit reliably *and* for CIRCT to consume directly. The architectural contribution: don't try to make the LLM emit RTL, give it a target that's correct-by-construction and that downstream tools already trust. A cleaner separation of concerns than most prior LLM-EDA work. 自GPT-4以来,LLM到RTL一直被视为LLM在EDA的明显应用方向,但实际投产很少——LLM直出的RTL难以校验、也难以与基于编译器的设计流(如CIRCT)整合。CPPL是位于合适抽象层级的类型化提示语言,既便于LLM稳定输出,也可被CIRCT直接消费。架构贡献在于:不要让LLM直出RTL,而是给它一个'构造即正确'且下游工具早已信任的目标。该关注点分离比此前多数LLM-EDA工作更干净。