semi·news
Headlines要闻 / Research研究 /
Research digest · Thursday, May 21, 2026 研究摘要 · 2026年5月21日 星期四

Compute-in-Memory Comes of Age While LLM Serving Treats Hardware as a First-Class Variable 存算一体走向系统,LLM服务把硬件当作一等变量

This week the hardware stack and the software stack converge on the same problem from opposite directions. From the silicon side: a fabricated TiO2/HfZrO2 ferroelectric FET, a 16Kb 3T1R ReRAM CIM macro spanning CNN and SNN inference, and an SNN inference architecture that finally exploits temporal sparsity. From the systems side: three new LLM-serving works treat GPU power caps, MoE expert parallelism, and per-phase quantization not as constraints but as control knobs. 本周硬件栈与软件栈从两端逼近同一个问题。硬件侧:TiO2/HfZrO2铁电FET的工艺、同时支持CNN与SNN推理的16Kb 3T1R ReRAM CIM macro、以及真正利用时间稀疏性的SNN推理架构。系统侧:三项新的LLM服务工作分别将GPU功耗上限、MoE专家并行、按阶段量化精度从硬约束转为可调旋钮。

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

Devices & Process 器件与工艺

2D GaSe single-pixel spectrometer via electro-optical barrier co-modulation 基于电-光势垒协同调制的2D GaSe单像素光谱仪 device-fab2D-materials

et al. (see paper)

arXiv:2605.20375 · 2026-05-21

A few-layer GaSe device is shown to act as a computational spectrometer by jointly modulating an electrostatic gate and an optical bias, replacing the dispersive optics of a conventional spectrometer with a single pixel. Reconstruction is done in software. The interest for chip-scale sensing is real, though spectral resolution and noise floor haven't been benchmarked against silicon-photonic or filter-array competitors. 作者证明一只少层GaSe器件可同时受电压和光学偏置调制,从而以单像素替代传统光谱仪的色散光学,后续由软件重建光谱。对芯片级光谱传感意义明确,但本文尚未与硅光及滤波器阵列方案在光谱分辨率与噪声底上做对比。

Circuits & Architecture 电路与体系结构

Dynamic control-flow execution for runtime-reconfigurable processors 面向运行时可重构处理器的动态控制流执行支持 architecturesimulation-only

et al. (see paper)

arXiv:2605.21203 · 2026-05-21

The paper tackles a long-standing limitation of CGRA/reconfigurable fabrics: irregular control flow forces fallback to a host CPU and wipes out the energy advantage. The proposed extensions support data-dependent branches natively on the reconfigurable substrate. Evaluation is simulation-only, but the design is concrete enough to be implementable, and the energy numbers are competitive with prior CGRA work. 本文针对CGRA等可重构架构长期存在的短板:不规则控制流必须回落到主机CPU,能效优势随之消失。作者提出的扩展可在可重构基底原生支持数据相关分支。评估为仿真级,但设计具体到可落地的程度,能效与前期CGRA工作相当。

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

E-ReCON: a 16Kb 3T1R ReRAM CIM macro for both CNN and SNN edge inference E-ReCON:面向CNN与SNN边缘推理的16Kb 3T1R ReRAM存内计算宏 tape-outcompute-in-memory

et al. (see paper)

arXiv:2605.20717 · 2026-05-21

The macro is real silicon: a 0.85µm² 3T1R ReRAM bitcell driving an AND-based in-memory multiplier that supports both conventional CNN and spiking workloads, with a novel interleaved 10T/28T adder tree to amortize accumulation overhead. The dual-mode CNN/SNN capability is the differentiating claim — most prior CIM macros pick one. Edge-inference targets, not data-center. 该宏为实际流片:0.85µm² 3T1R ReRAM位单元搭配基于AND的存内乘法,同时支持传统CNN与脉冲网络负载,并引入新颖的10T/28T交错加法树以分摊累加开销。'CNN/SNN双模'是其核心差异点——此前的CIM宏多只支持其一。面向边缘推理,而非数据中心。

ELSA: an elastic SNN inference architecture that exploits temporal sparsity ELSA:真正利用时间稀疏性的弹性脉冲神经网络推理架构 architectureneuromorphic

et al. (see paper)

arXiv:2605.20802 · 2026-05-21

SNN accelerators have advertised event-driven efficiency for years, but most are still layer-by-layer pipelines that wait for the full input window before emitting an output. ELSA implements true elastic inference: confident outputs are emitted as soon as enough spikes accumulate, and downstream layers are gated accordingly. The architectural lesson is that SNN efficiency is fundamentally a scheduling problem, not just a circuit one. SNN加速器多年来一直宣传事件驱动的能效,但绝大多数仍是逐层流水、等满输入窗口才输出。ELSA实现了真正的弹性推理:在脉冲累积到置信阈值时立即输出,下游层据此被门控。其架构启示是:SNN的能效首先是调度问题,而非单纯的电路问题。

High-performance TiO2 ferroelectric FETs with HfZrO2 dielectric for neuromorphic computing 面向类脑计算的高性能TiO2铁电FET(HfZrO2介质) device-fabneuromorphic

et al. (see paper)

arXiv:2605.20406 · 2026-05-21

Bottom-gate FeFETs using TiO2 channels on HfZrO2 ferroelectric dielectric, reported with high on/off ratios across two gate-stack thicknesses. The novelty is moving the CMOS-back-end-compatible HZO stack into a configuration that supports gradual conductance modulation — the prerequisite for analog weight storage. Single-device characterization only; no array-level retention or endurance data yet. 作者在HfZrO2铁电介质上以TiO2作为沟道,制备底栅FeFET并报告两种栅叠厚度下的高开关比。新意在于将与CMOS后段兼容的HZO体系配置为支持电导渐变调制——这是模拟权重存储的前提。仅为单器件表征,尚无阵列级保持或耐久数据。

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

Mix-Quant: phase-aware quantization — aggressive in prefill, precise in decode Mix-Quant:阶段感知量化——预填激进、解码精确 quantizationLLM-serving

et al. (see paper)

arXiv:2605.20315 · 2026-05-21

Most LLM quantization schemes pick a single precision for the whole forward pass; Mix-Quant observes that the prefill phase (compute-bound, error-tolerant due to attention averaging) and the decode phase (memory-bound, error-sensitive) want different precisions. The framework runs prefill at FP4 / INT4 and decode at higher precision, with measured throughput gains on agentic, long-context workloads. The framing — phase as a first-class quantization dimension — is the more durable contribution than any specific schedule. 多数LLM量化方案对整次前向使用统一精度;Mix-Quant注意到prefill阶段(计算密集、对误差宽容)与decode阶段(内存密集、对误差敏感)需要不同精度。框架在prefill使用FP4/INT4,decode使用更高精度,在Agentic长上下文负载上获得实测吞吐量提升。比起具体调度,'阶段作为一等量化维度'这一框架本身更具持久价值。

PALS: power-aware LLM serving — GPU power caps as a runtime variable PALS:功耗感知的LLM服务——把GPU功耗上限当成运行时变量 LLM-servingMoEpower

et al. (see paper)

arXiv:2605.21427 · 2026-05-21

Existing LLM-serving systems treat GPU power as a static datacenter constraint and optimize only batching, scheduling, and parallelism. PALS treats power caps as a first-class control surface, dynamically retuning per-GPU power limits across expert-parallel MoE replicas to match the actual computation cost per token. The reported gains in tokens-per-watt are modest but the architectural shift is the right one: in a memory-bound inference regime, the binding constraint is increasingly thermal, not FLOPS. 现有LLM服务系统将GPU功耗视为数据中心层面的静态约束,只在批处理、调度、并行上做优化。PALS将功耗上限当作一等控制面,在专家并行MoE副本之间动态调整每卡功耗上限,以匹配每token的实际计算开销。论文给出的能效提升不算激进,但架构方向是对的:在内存受限的推理形态下,真正的硬约束越来越是热,而非FLOPS。

NanoCP: request-level dynamic context parallelism for data-expert parallel MoE decoding NanoCP:用于数据-专家并行MoE解码的请求级动态上下文并行 LLM-servingMoEKV-cache

et al. (see paper)

arXiv:2605.21100 · 2026-05-21

Modern MoE serving binds each request's attention, MoE communication, and KV cache to a single instance — fine for short prompts, ruinous for long context because attention cost scales with KV-cache size while MoE all-to-all latency scales with batch. NanoCP dynamically splits context parallelism per request, decoupling these two scaling laws. The contribution is small in code but large in implication: MoE serving has to stop pretending requests are uniform. 现代MoE服务系统将每个请求的注意力、MoE通信与KV缓存绑定到单一实例——短提示词没问题,长上下文则灾难性:注意力开销随KV缓存大小增长,MoE的all-to-all延迟随批大小增长,两者的扩展规律相互矛盾。NanoCP按请求动态分配上下文并行度,解耦了这两个维度。代码层改动不大,意义却大:MoE服务必须停止假装请求是同质的。

Quantum Devices 量子器件

Towards transistor-based quantum computing: telesistors and topologically protected Clifford gates 面向晶体管化的量子计算:Telesistor与拓扑保护Clifford门 theorypreprint

et al. (see paper)

arXiv:2605.21045 · 2026-05-21

The authors propose a universal QC architecture in which the elementary unit is a teleportation-based 'telesistor' built from symmetry-protected topological ground states, which suppress certain noise channels and yield Clifford gates without active error correction. The hope is to make the architectural picture of a quantum processor more transistor-like — a tractable, composable element. Theoretical proposal; nothing fabricated. 作者提出一种通用量子计算架构,其基元为基于对称性保护拓扑序基态的'telesistor':通过隐形传态实现门操作,从而被动抑制特定噪声、无需主动纠错即可获得Clifford门。其雄心在于让量子处理器的架构图景更像经典晶体管:可分解、可组合。目前为理论提案,尚无器件实现。

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

LLMs for secure hardware design: opportunities and severe vulnerabilities 大模型用于硬件安全设计:机会与严重漏洞 surveyEDA

et al. (see paper)

arXiv:2605.10807 · 2026-05-21

A comprehensive review of LLMs across the EDA stack — RTL generation, testbench synthesis, spec-to-silicon — that breaks from the usual cheerleading and catalogs concrete vulnerabilities: prompt-injection-induced Trojans, leakage of pre-silicon IP through chat logs, and verification gaps the models reliably miss. The clearest practitioner-facing summary published this year. 对大模型在EDA各环节(RTL生成、测试平台合成、规格到芯片的桥接)应用的全面综述,但跳出业内常见的乐观叙述,系统记录了具体漏洞:提示注入诱导的硬件木马、聊天日志泄漏前期IP、以及模型反复遗漏的验证盲区。是今年面向实务工程师最清晰的一份梳理。