semi·news
Headlines要闻 / Research研究 / /
Research digest · Friday, August 21, 2026 研究摘要 · 2026年8月21日 星期五

Specialized Hardware Moves Compute Toward the Sensor 专用硬件将计算推向传感端

This thin research window centers on fixed-point biosignal analytics, memristive in-sensor computing, and probabilistic hardware for 3D reconstruction. The efficiency claims are still bounded by synthesis, review-level evidence, or simulation, leaving measured deployment as the central gap. 本期较薄的研究窗口聚焦于定点生物信号分析、忆阻器传感内计算，以及用于3D重建的概率计算硬件。其效率结论仍受限于综合结果、综述性证据或仿真，实测部署仍是最关键的缺口。

Look-back window: 7 days · 3 paper(s) 回溯窗口: 7天 · 3篇

Circuits, Architecture & Reliability 电路、架构与可靠性

Fixed-Point Hardware for Hjorth-Feature Cardiac Fibrillation Diagnosis 用于Hjorth特征心颤诊断的定点硬件

I. Kouretas, A. G. Skrivanos, N. Sagias, et al.

Electronics · 2026-08-17

A fully synthesizable fixed-point pipeline computes Hjorth activity, mobility, and complexity in real time, then feeds those features directly to a quantized neural network. At 10–12 bits, the design reports about 93.7% classification accuracy and an AUC above 0.97 on the SPHD dataset, close to its floating-point baseline. The result is application-specific and synthesis-focused rather than measured silicon, so area, energy, and generalization beyond the selected clinical dataset remain open questions. 该工作实现了一条可完全综合的定点流水线，可实时计算Hjorth活动度、移动度和复杂度，并将特征直接送入量化神经网络。在10–12 bit配置下，设计在SPHD数据集上报告约93.7%的分类准确率和高于0.97的AUC，接近浮点基线。该结果面向特定应用，且主要基于综合而非实测芯片，因此面积、能耗以及对所选临床数据集之外的泛化能力仍待验证。

Devices & Process 器件与工艺

From Memristive Biosensors to In-Sensor Computing 从忆阻式生物传感器到传感内计算

H. Ryu, W.-C. Lee, J. Lee

Biosensors · 2026-08-16

This review traces memristive biosensors from direct and indirect transduction through on-chip integration and in-sensor computing, where sensing, storage, and processing share one physical platform. The architecture could reduce data movement in point-of-care systems, but the paper identifies CMOS compatibility, device variability, and reliable multi-threshold operation as unresolved barriers. It is a field map rather than a new measured device result, making it most useful for understanding integration tradeoffs and open engineering problems. 这篇综述梳理了忆阻式生物传感器从直接、间接换能，到片上集成和传感内计算的发展路径；在后一种架构中，感知、存储与处理共处同一物理平台。该架构有望减少即时检测系统的数据搬运，但论文指出CMOS兼容性、器件波动以及可靠的多阈值工作仍是未解决障碍。这是一幅领域路线图，而非新的实测器件结果，主要价值在于理解集成权衡和待解工程问题。

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

ProbSplat: Probabilistic Hardware for Gaussian Splatting ProbSplat：面向Gaussian Splatting的概率计算硬件

S. Gottumukkula, M. P. Samartha, V. Pahariya, et al.

arXiv:2608.13143 · 2026-08-13

ProbSplat stores independently programmable Gaussian means and variances in floating-gate inverter columns and evaluates log-likelihoods for 3D scene reconstruction. A simulated 180nm, 1.8V design operating at 50MHz reports less than 2.4% mean-variance deviation and 18pJ per 4-bit log-likelihood inference across 500 mixture functions. The approach targets a data-movement-heavy edge workload, but all headline results are simulation-only and need array-level silicon measurements, endurance data, and comparisons against current digital accelerators. ProbSplat在浮栅反相器列中分别编程并存储Gaussian分布的均值与方差，并为3D场景重建计算对数似然。其180nm、1.8V、50MHz仿真设计在500个混合函数上报告低于2.4%的均值—方差偏差，以及每次4 bit对数似然推理18pJ的能耗。该方案瞄准数据搬运密集的边缘工作负载，但核心结果全部来自仿真，仍需阵列级芯片实测、耐久性数据，并与当前数字加速器比较。

Newsletter 邮件订阅

Daily semiconductor briefing. 每日半导体简报。