semi·news
Headlines要闻 / Research研究 / /
Research digest · Thursday, November 5, 2026 研究摘要 · 2026年11月5日 星期四

Memory Placement Shapes Inference Efficiency Across Scales 存储布局影响各尺度的推理效率

Ferroelectric memory arrays, ternary ROM computing, and optical fabrics tackle data movement at different layers of the hardware stack. Device measurements and commercial-board benchmarks sit alongside architecture emulation and software experiments, with each result retaining its evaluation scope. 铁电存储阵列、三值ROM计算与光互连网络从硬件栈的不同层次处理数据搬运问题。本期涵盖器件实测、商用板卡基准、架构仿真和软件实验,解读各项结果时需结合相应的评估范围。

Look-back window: 7 days · 8 paper(s) 回溯窗口: 7天 · 8篇

Devices & Process 器件与工艺

NORmal: Bringing 3D Ferroelectric NOR Flash Onto an LPDDR6 Memory Interface NORmal:将3D铁电NOR闪存接入LPDDR6内存接口

Zhuoran Li, Zhuohang Bian, Yibo Zhao, et al.

arXiv:2610.40000 · 2026-10-29T13:19:45Z

NORmal reports fabricated BEOL 3D ferroelectric NOR arrays combining dense 3-bit MLC storage with fast, writable SLC and sub-7.5 ns sensing. It proposes a DRAM-emulation bridge that presents these arrays as a byte-addressable rank on an unmodified LPDDR6 interface. The approach aims to expand local AI memory capacity while accommodating mutable runtime data. The sensing measurement leaves controller, interface, and complete-system timing as separate evaluation questions. NORmal报告了已制备的后道互连层(BEOL)3D铁电NOR阵列,将高密度3-bit MLC存储与快速、可写的SLC集成,感测时间低于7.5 ns。研究提出通过DRAM仿真桥,将这些阵列作为可按字节寻址的rank接入未经修改的LPDDR6接口。该方案旨在扩大本地AI的内存容量,同时容纳运行时更新的数据。感测时间这一实测指标之外,控制器、接口与完整系统的时序仍需分别评估。

Circuits & Reliability 电路与可靠性

A 65nm CiROM Macro Stores Two Ternary Weights per Cell for 1.58-bit LLM Inference 面向1.58-bit LLM推理的65nm CiROM宏单元:每个单元存储两个三值权重

Wen-Lun Zhang, Xin-Yu Li, Shimpei Ando, et al.

IEEE Transactions on Circuits and Systems I: Regular Papers · 2026-11-01

A 65nm compute-in-ROM macro uses configurable source and bit lines to store two ternary weights in each 1T ROM cell for BitNet-style inference. A local accumulator gates zero-weight operations, while a digital adapter co-designed with LoRA provides downstream task adaptation after fabrication. The design links storage density and sparsity to circuit-level switching reductions, with fixed ROM weights constraining how broadly a deployed chip can be retargeted. 这项65nm ROM内计算宏单元设计通过可配置源极线与位线,在每个1T ROM单元中存储两个三值权重,面向BitNet类推理。本地累加器对零权重运算实施门控,与LoRA协同设计的数字适配器则支持流片后的下游任务适配。设计将存储密度和稀疏性利用落实到电路开关活动的削减,但固定的ROM权重也会约束芯片部署后可迁移的任务范围。

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

Workload-Static Optical Circuit Switching for Scalable LLM Decode 以工作负载静态光路交换扩展LLM解码互连

Maximilian Thürmer, Nicolas Scotto di Perto, Eloise Stein, et al.

arXiv:2611.05104 · 2026-11-03T16:00:07Z

This optical circuit-switching design configures links once per model deployment and co-designs collective schedules with model parallelism. For a modeled 576-accelerator domain serving 5T-parameter MoE models with contexts up to 5M tokens, it reports up to 65% better interactivity and 119% better energy efficiency than a 4:1 oversubscribed packet-switched fabric. The gains come from GXNET workload emulation and analytical and flit-level network models grounded in hardware measurements, making oversubscription and communication schedules key evaluation assumptions. 这项光路交换设计在每次模型部署时配置一次链路,并将集合通信调度与模型并行方式协同优化。在模拟的576颗加速器域中,运行5万亿参数MoE模型、上下文最长500万token时,相较4:1超额订阅的分组交换网络,交互性能最高提升65%,能效最高提升119%。这些结果来自GXNET工作负载仿真及以硬件测量为依据的分析模型、flit级网络模型,超额订阅比例与通信调度是理解收益的关键评估条件。

TinyML Benchmarks Ten Sub-Watt Devices and Exposes Memory Constraints TinyML对比十款亚瓦级器件,检视存储约束

Pietro Bonazzi, Luigi Capogrosso, Ahmet Celik, et al.

arXiv:2611.03392 · 2026-11-02T16:28:53Z

A benchmark evaluates four CNNs and two small transformers across ten commercial edge devices, including sensor-integrated processing and MCUs with and without NPUs. For the reported energy-above-idle measurements, Sony's IMX500 uses 0.13–0.35 mJ versus 0.31–1.05 mJ for STM32N6; different models also favor GAP9 or Alif E7. The workload-dependent rankings show why platform selection needs measured latency and energy alongside operation counts, with the quoted energy figures excluding idle consumption. 这项基准在十款商用边缘器件上评估了四种CNN与两种小型Transformer,覆盖传感器内处理器以及有、无NPU的MCU。按高于空闲基线的增量能耗计算,Sony IMX500为0.13–0.35 mJ,STM32N6为0.31–1.05 mJ;部分模型则以GAP9或Alif E7表现更优。排名随工作负载变化,说明平台选型需要结合实测时延与能耗评估运算量,上述能耗数值未计入空闲消耗。

Streaming NTT Across Three FPGA Dies for Homomorphic Encryption 跨三个FPGA芯粒的流式NTT加速同态加密

Mohamed Allam, Thomas Spendlhofer, Hugo Sanz-González, Antonio J. Peña

arXiv:2611.03325 · 2026-11-02T15:49:14Z

An NTT/INTT kernel mapped across the AMD Alveo V80's three dies reports processing a 65,536-coefficient polynomial every 5.12 microseconds at 400 MHz. Its fully unrolled delay-commutator streaming structure addresses the routing overhead that can grow with parallelism on multi-die FPGAs. This targets a major cost in homomorphic-encryption inference, with application speedup also depending on transfers and the remaining FHE pipeline. 这项NTT/INTT内核映射到AMD Alveo V80的三个芯粒,在400 MHz下每5.12微秒可处理一个含65,536个系数的多项式。完全展开的延迟交换器流式结构针对多芯粒FPGA中随并行度上升的布线开销。该设计瞄准同态加密推理的一项主要计算成本,应用整体加速比还取决于数据传输及FHE流水线的其他部分。

AI Research & Inference Systems AI研究与推理系统

UMLoad Accelerates On-Demand Model Loading on Unified Memory UMLoad加速统一内存平台的按需模型加载

Zimeng Wang, Jinrong Yang, Rong Chen

arXiv:2611.04253 · 2026-11-03T03:44:38Z

UMLoad reduces average on-demand model loading time by 36.6–71.9% against unmodified loaders and 19.9–62.7% against pipelined loaders in tests across Intel, AMD, and NVIDIA unified-memory platforms. It separates logical allocations from physical page backing, eliminating weight copies through on-demand reconstruction and accelerating allocation with GPU-assisted page zeroing. Mean time to first token in LLM-routing tests falls 51.3–59.3%, making loader design relevant to interactive latency; the evaluation covers six models and the tested software stacks. 在Intel、AMD与NVIDIA统一内存平台的测试中,UMLoad相较原始加载器将平均按需模型加载时间缩短36.6–71.9%,相较流水化加载器缩短19.9–62.7%。它将逻辑分配与物理页后备解耦,通过按需重建消除权重复制,并借助GPU辅助页清零加速内存分配。LLM路由测试的平均首token时延降低51.3–59.3%,显示加载器设计会显著影响交互响应;评估范围为六个模型及所测试的软件栈。

PagedMoE Reuses Expert Cache Data Across Precision Levels PagedMoE跨精度复用专家缓存数据

Yutong Lin, Yingdi Shan, Shaoyuan Chen, et al.

arXiv:2611.03173 · 2026-11-02T14:09:55Z

PagedMoE stores each SSD-resident MoE expert in a prefix-compatible representation, allowing cached low-precision data to be extended by fetching only missing bitplanes. It selects precision before I/O using routing weights and expert-specific error profiles, aiming to avoid repeated transfers when requested bit widths change. The design targets DRAM-limited local inference, with practical throughput and quality depending on cache capacity, SSD behavior, and the precision-selection policy. PagedMoE将驻留SSD的每个MoE专家存成前缀兼容的表示,使已缓存的低精度数据能够通过补取缺失位平面扩展到更高精度。系统利用路由权重与各专家的误差特征,在I/O发生前选择精度,力求避免所需位宽变化时重复搬运相同数据。该设计面向DRAM容量受限的本地推理,实际吞吐量与模型质量取决于缓存容量、SSD特性及精度选择策略。

EDA & Design Automation EDA与设计自动化

APPlace Co-Optimizes Analog Device Placement and Pin Access Points APPlace联合优化模拟器件布局与引脚接入点

Bingyang Liu, Xiaohan Gao, Souradip Poddar, David Z. Pan

arXiv:2610.39515 · 2026-10-29T07:14:08Z

APPlace jointly optimizes device coordinates and routing access-point selections during analog global placement, then preserves that structure through legalization. Across four real-world analog benchmarks, it reports better overall post-layout performance than representative baselines at comparable layout area. Modeling the actual connection choices gives placement a more useful view of downstream routing, with the reported evidence scoped to those benchmark circuits. APPlace在模拟电路全局布局阶段联合优化器件坐标与布线接入点选择,并在合法化阶段保留已建立的结构。在四个实际模拟电路基准上,研究报告其在布局面积相近时,整体版图后性能优于代表性基线方法。将实际连接位置纳入建模,有助于布局阶段更准确地考虑后续布线,现有结果的验证范围为这些基准电路。

Newsletter 邮件订阅

Daily semiconductor briefing. 每日半导体简报。