semi·news
Headlines要闻 / Research研究 / /
Research digest · Saturday, November 7, 2026 研究摘要 · 2026年11月7日 星期六

Efficiency Depends on Devices, Mapping and Scheduling 效率取决于器件、映射与调度协同

Device experiments and measured CMOS prototypes address leakage, memory access and early vision processing, while architecture studies examine mapping and inference partitioning. Quantum scheduling and HLS repair extend the focus to execution efficiency and design productivity. 器件实验与CMOS原型实测聚焦漏电、存储访问和前端视觉处理,架构研究则考察映射与推理任务拆分。量子调度和HLS修复进一步将效率问题延伸到执行与设计环节。

Look-back window: 7 days · 9 paper(s) 回溯窗口: 7天 · 9篇

Devices & Process 器件与工艺

Fluorine Surface Engineering Tunes HZO FeFET Thresholds and Expands the Memory Window 氟表面工程调节HZO FeFET阈值并扩大存储窗口

Kyungsoo Park, Sangmyun Lim, Sangkuk Han, et al.

Nano Convergence · 2026-11-04

Fluorine surface treatment shifts the threshold voltage of HZO-based silicon FeFETs positively by about 0.5 V, expands the memory window by about 0.5 V and reduces leakage at 0 V by 40x. The interface treatment enables normally-off operation of the low-threshold state while largely preserving HZO ferroelectric switching. CIFAR-10 accuracy improves from about 78% to 83% in simulations using measured device characteristics, linking the device changes to a modeled inference benefit. 氟表面处理使HZO基硅FeFET的阈值电压正向移动约0.5 V,存储窗口扩大约0.5 V,并使0 V下的漏电流降至原来的约1/40。该界面处理在基本保持HZO铁电翻转特性的同时,使低阈值态实现常关断。采用实测器件特性的仿真将CIFAR-10分类准确率从约78%提高至83%,表明器件改进可在模型层面带来推理收益。

Composition-Dependent Atomic Layer Etching Preserves IGZO Electrical Performance 组分相关的原子层刻蚀保留IGZO电学性能

Min Chan Kim, Min-seo Kim, Hyeseong Lee, et al.

ACS Applied Materials and Interfaces · 2026-11-04

Hydrogen-based atomic layer etching controls IGZO removal through its elemental composition while largely preserving transistor electrical performance. The near-surface reaction limits channel damage and enables post-deposition trimming inside three-dimensional trenches, improving step coverage to nearly 100%. This provides a process route toward more uniform oxide-semiconductor channels for scaled memory devices. 研究通过氢基原子层刻蚀,利用IGZO各元素反应活性的差异调节材料去除速率,同时基本保留晶体管电学性能。反应主要局限于近表面区域,可减少沟道损伤,并在三维沟槽中进行沉积后修整,使台阶覆盖率接近100%。这一工艺为缩尺度存储器件获得更均匀的氧化物半导体沟道提供了路径。

Circuits & Architecture 电路与架构

A CMOS Vision Sensor Computes Temporal Changes Without Frame Memory 无需帧存储的CMOS视觉传感器实现时域变化计算

Seong-Min Ko, Kishore Kasichainula, Injune Yeo, et al.

IEEE Journal of Solid-State Circuits · 2026-11-01

A 324 × 252 vision-sensor prototype fabricated in 0.18-μm CMOS consumes 4.79 μW at 12.5 frames/s in patch-mode temporal-change processing. A checkerboard pixel architecture and column-shared circuits detect changes without frame memory, while in-pixel time memory supports mixed-signal temporal-derivative computation. The measured design shows how moving initial motion processing into the sensor can reduce memory requirements for low-power vision. 采用0.18 μm CMOS制造的324 × 252视觉传感器原型,在图像块模式下以12.5帧/秒处理时域变化,实测功耗为4.79 μW。棋盘式像素架构与列共享电路无需帧存储即可检测变化,像素内时间存储电路则支持混合信号时域导数计算。该设计展示了将初步运动处理前移到传感器内部,以降低低功耗视觉系统存储需求的方法。

Memory Fit Learns Page Affinity for Heterogeneous Server Memory Memory Fit学习内存页亲和性,优化异构服务器存储放置

Chun Deng, Hui Sub Shim, Evan Laufer, et al.

arXiv:2611.07506 · 2026-11-05T06:10:55Z

Memory Fit supplies the operating system with page-to-memory affinity information by matching access traces against learned models of suitable behavior. The authors implement the abstraction and its soft trace matching algorithm in a BOOMv3 RISC-V processor running Linux 6.6. On extended GUPS, its placements recover 2.2–3.8 times as much of the performance gap to optimal placement as Linux TPP, showing the value of better hardware guidance for heterogeneous-memory management. Memory Fit将访问轨迹与学习得到的适配行为模型进行匹配,向操作系统提供内存页与不同存储介质之间的亲和性信息。作者在运行Linux 6.6的BOOMv3 RISC-V处理器中实现了这一抽象及其软轨迹匹配算法。在扩展GUPS测试中,其页面放置方案弥合的最优放置性能差距为Linux TPP的2.2—3.8倍,体现了更精细硬件信息对异构内存管理的价值。

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

A Valuewise Hybrid-Domain CIM Macro Trades Precision Against Energy 按数值重要性分配计算域的混合CIM宏权衡精度与能耗

Yuan-Zhe Zhao, Yang Wang, Yu-Heng Wang, et al.

IEEE Journal of Solid-State Circuits · 2026-11-01

A 28-nm valuewise hybrid CIM prototype computes a small proportion of the most significant values digitally and processes most values in analog memory arrays. A heterogeneous SRAM/DRAM memory fabric and asynchronous sparsity management address latency and unnecessary activity. Measured energy efficiencies are 470.5 TOPS/W at 4-bit precision, 109.8 TOPS/W at 8-bit and 48.5 TOPS/W at 12-bit, making the precision dependence of the energy benefit explicit. 这款28 nm按数值重要性划分计算域的混合CIM原型,将少量最重要的数值交由数字电路精确计算,其余大部分数值在模拟存储阵列中处理。异构SRAM/DRAM存储结构与异步稀疏管理用于降低时延和无效活动。实测能效在4位、8位和12位精度下分别为470.5、109.8和48.5 TOPS/W,清楚呈现了能效收益对计算精度的依赖。

Isobar Models Topology-Aware Mapping for Scaled Analog CIM Isobar为扩展模拟CIM建立拓扑感知映射模型

Yuxin Ren, Danella Zhao, Huanrui Yang

arXiv:2611.08399 · 2026-11-05T18:11:00Z

Isobar models how weight-mapping topology, heterogeneous execution and resource contention affect analog CIM systems as capacity scales through chiplets. For Gemma-7B prefill, its mapping search reduces modeled communication time by 23.3% and end-to-end latency by 18.6% relative to the strongest baseline. The simulator is validated against silicon measurements and cycle-accurate simulation, providing evidence for mapping decisions while keeping the reported system gains at the simulation level. Isobar对模拟CIM通过chiplet扩展容量时,权重映射拓扑、异构执行与资源争用如何影响系统性能进行建模。在Gemma-7B预填充任务中,其映射搜索相对最强基线将仿真通信时间降低23.3%,端到端时延降低18.6%。仿真器经过芯片实测与周期精确仿真校验,可为映射决策提供依据,但所报告的系统收益仍属于仿真结果。

AI Models & Inference Systems AI模型与推理系统

Splitting Decode Shows No Robust Efficiency Advantage Across Modeled Future Systems 未来系统建模显示:进一步拆分解码未呈现稳健能效优势

Ramyad Hadidi

arXiv:2611.08299 · 2026-11-05T17:02:06Z

SiliconForge compares prefill/decode separation with an additional split between attention and feed-forward processing across four model directions, four accelerator classes and two interconnect fabrics. Under a common power budget and optimization rules, the study finds no robust deployment-level advantage for the extra split in SLO-compliant goodput per kW. The analysis shows how added communication and provisioning can offset specialization benefits, with conclusions dependent on modeled workloads and hardware assumptions. SiliconForge在四类模型发展方向、四类加速器和两种互连网络上,对比预填充与解码分离,以及在此基础上进一步拆分注意力和前馈计算的方案。在相同功率预算与优化规则下,研究未发现进一步拆分能在每千瓦满足SLO的有效吞吐量上获得稳健的部署级优势。分析说明,新增通信与资源池配置开销可能抵消专用化收益,结论仍取决于所建模的工作负载与硬件假设。

Quantum Computing 量子计算

DynaMite Selects Measurement Latencies for Fault-Tolerant Quantum Execution DynaMite动态选择测量时延,优化容错量子执行

Dongwhee Kim, Suhas Vittal, Poulami Das

arXiv:2611.05233 · 2026-11-03T17:18:12Z

DynaMite proposes real-time selection of measurement latency to shorten fault-tolerant quantum programs. The paper identifies qubit measurements as accounting for 43–86% of logical cycle time and argues for scheduling around logical performance rather than relying solely on per-qubit readout calibration. The approach links measurement scheduling to error-correction performance by balancing readout speed, state-discrimination accuracy and decoherence. DynaMite提出实时选择测量时延的策略,以缩短容错量子程序的执行时间。论文指出,量子比特测量占逻辑周期时间的43%—86%,因此调度应围绕逻辑层性能展开,并将单比特读出校准之外的因素纳入考量。该方法通过权衡读出速度、状态判别准确率与退相干,将测量调度和纠错性能联系起来。

EDA & Design Automation EDA与设计自动化

MementoHLS Uses Verified Repair History to Improve LLM-Generated HLS MementoHLS利用经验证的修复历史改进LLM生成HLS

Xianjing Zhu, Wenqi Lou, Yuanbo Wen, et al.

arXiv:2611.07324 · 2026-11-05T02:09:44Z

MementoHLS combines reusable repair rules with task-specific revision history to improve LLM generation of functionally correct, synthesizable HLS-C. On 94 HLS-Eval tasks, the rate of candidates passing both testbench and synthesis checks rises from 41.7% to 58.9% at pass@1 and from 57.4% to 68.1% at pass@5. Binding all checks to the same candidate strengthens the evidence for generation success, while these metrics establish correctness and synthesizability rather than final chip performance. MementoHLS将可复用修复规则与任务内修改历史结合,用于提高LLM生成功能正确、可综合HLS-C的成功率。在94项HLS-Eval任务上,同一候选同时通过测试平台与综合检查的pass@1从41.7%提高至58.9%,pass@5从57.4%提高至68.1%。将所有检查结果绑定到同一候选增强了成功判定的可信度,这些指标衡量的是功能正确性与可综合性,最终芯片性能仍需另行评估。

Newsletter 邮件订阅

Daily semiconductor briefing. 每日半导体简报。