semi·news
Headlines要闻 / Research研究 / /
Research digest · Friday, November 13, 2026 研究摘要 · 2026年11月13日 星期五

Hardware Efficiency Depends on Circuits and Dataflow 电路与数据流共同决定硬件效率

New device and EDA papers join earlier-November circuit and neuromorphic work from the merged feed. The selection examines reconfigurable logic, chiplet communication, parasitic-aware computation, and low-power quantum control. 本期将新近的器件与EDA论文,同合并候选池中11月较早发表的电路及类脑计算研究一并选读。关注方向包括可重构逻辑、chiplet通信、考虑寄生效应的计算,以及低功耗量子控制。

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

Devices & Process 器件与工艺

Reconfigurable Neuromorphic Logic Circuits Driven by Ion-Mediated Organic Electrochemical Diodes 离子调控有机电化学二极管驱动的可重构类脑逻辑电路

Dongmi Heo, Min-Hu Huang, Youngseok Kim, et al.

Advanced Functional Materials · 2026-11-10

Organic electrochemical diodes implement multiple logic functions in a single diode-resistor configuration, with input-voltage polarity and frequency reprogramming channel conductance. Cascading two multistate circuits produces a proof-of-concept half-adder. The result connects device-level state control to reconfigurable logic, with the demonstrated scope centered on small circuits. 有机电化学二极管通过输入电压的极性与频率调节沟道电导,在单个二极管—电阻结构中实现多种逻辑功能。研究将两个多状态逻辑电路级联,演示了概念验证性质的半加器。这一结果把器件状态调控与可重构逻辑连接起来,当前演示集中于小规模电路。

Circuits & Readout Systems 电路与读出系统

A Multi-Functional Streaming Readout ASIC With Channel-Embedded DSP and Neural Inference 通道内集成DSP与神经网络推理的多功能流式读出ASIC

S. Mandal, G. Deptuch, Jack Fried, et al.

IEEE Transactions on Circuits and Systems Part 1: Regular Papers · 2026-11-01

A 65 nm CMOS radiation-detector readout ASIC integrates 23 independent channels, each combining an analog front end, a 12-bit ADC, programmable DSP, and a fixed-point multilayer perceptron. The neural engine uses 8-bit weights and 12-bit activations for on-chip regression and classification. Supporting feature streaming alongside full waveforms moves data reduction into the acquisition chip and targets output-bandwidth demand. 一款采用65 nm CMOS的辐射探测器读出ASIC集成23个独立通道,每通道包含模拟前端、12 bit ADC、可编程DSP及定点多层感知机。神经网络引擎使用8 bit权重和12 bit激活值,在片上执行回归与分类。芯片同时支持完整波形和特征流输出,将数据压缩处理前移到采集芯片,以降低输出带宽需求。

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

SpikeLet Combines Hierarchical Interconnects with Spatial-Temporal Parallelism for Neuromorphic Chiplets SpikeLet结合分层互连与时空并行,扩展类脑chiplet系统

Jia-Jun Lu, Xiao-Yang Zeng, Chi-Xiao Chen, et al.

IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems · 2026-11-01

SpikeLet combines hierarchical spike aggregation, cross-layer and cross-timestep parallelism, and communication-aware mapping for multichiplet spiking neural networks. On a 4-by-4-chiplet prototype, the authors report a 74.2% reduction in transmission latency in their evaluation. The design addresses the communication and scheduling costs that become important when a spiking network grows beyond one chip's capacity. SpikeLet面向多chiplet脉冲神经网络,结合分层脉冲聚合、跨层与跨时间步并行,以及通信感知映射。在4×4个chiplet的原型上,作者报告其评估中的传输延迟降低74.2%。该设计针对脉冲网络规模超过单芯片容量后逐渐突出的通信与调度开销。

HyVox Splits Low-Power Speaker Identification Across Neuromorphic and von Neumann Processors HyVox在类脑与冯·诺依曼处理器间划分低功耗说话人识别任务

S. Addy, Eduardo Calle-Ortiz, Prashant Shenoy, et al.

arXiv:2611.03735 · 2026-11-01

HyVox assigns speech embeddings to a neuromorphic processor and speaker matching, unknown-speaker rejection, and enrollment to a microcontroller. Its backbone compresses 6.5-fold to 29,000 synapses, while a 32-neuron embedding extractor passes only spike counts between the processors. This partition lets new speakers enter through gallery-prototype updates and concentrates sparse event-driven computation on the neuromorphic chip. HyVox将语音嵌入计算交给类脑处理器,将说话人匹配、未知说话人拒识及注册交给微控制器。其骨干网络压缩6.5倍至29000个突触,32神经元嵌入提取器仅通过脉冲计数与另一处理器交换信息。这种划分允许通过更新说话人库中的原型注册新用户,并将稀疏事件驱动计算集中到类脑芯片上。

Hardware-Aware AI Research 硬件感知AI研究

An Educated Guess Derives Activation-Aligned Gradients for Zeroth-Order Optimization An Educated Guess利用激活几何构造零阶优化的梯度估计

Justus Huebotter, Serge Thill, M. van Gerven, et al.

arXiv:2611.08712 · 2026-11-05

Spectral Zeroth-Order methods use activation geometry available during forward passes to derive gradient-search directions under explicit stochastic assumptions. The approach targets the poor alignment between random perturbations and true gradients in high-dimensional models. The preprint connects inference-only gradient estimation to learning on neuromorphic, asynchronous, and other hardware with demanding synchronization or differentiability constraints. Spectral Zeroth-Order方法利用前向计算中可获取的激活几何,在明确的随机假设下推导梯度搜索方向。该方法针对高维模型中随机扰动与真实梯度对齐程度较差的问题。这篇预印本将仅依赖前向计算的梯度估计,与类脑、异步及其他在同步或可微性方面受限的硬件学习联系起来。

Quantum Hardware & Control 量子硬件与控制

A 16-Channel Cryo-CMOS Flux Control Pulse Generator ASIC 16通道低温CMOS磁通控制脉冲发生器ASIC

K. Tien, D. Frank, J. Bulzacchelli, et al.

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

A 16-channel cryo-CMOS flux-control ASIC achieves a qubit-limited two-qubit gate error rate of 0.084% with a superconducting quantum processor while consuming less than 7 mW per flux-control channel. Three 10-bit sub-DACs operating at rates up to 1 GS/s combine to deliver 14-bit level resolution, with on-chip digital precompensation for off-chip distortion. The measured combination of gate fidelity and channel power provides a concrete integration point for scaling cryogenic qubit control. 一款16通道低温CMOS磁通控制ASIC与超导量子处理器联合演示,实现了受量子比特本身限制的0.084%双量子比特门错误率,每个磁通控制通道功耗低于7 mW。3个以最高1 GS/s速率工作的10 bit子DAC组合出14 bit电平分辨率,并由片上数字预补偿修正片外失真。门保真度与通道功耗的联合实测,为扩展低温量子比特控制提供了具体的集成参考。

Electronic Design Automation 电子设计自动化

Agentic Design of Approximate DFT IP in TSMC 28 nm CMOS with RF-SoC Verification 基于智能体设计台积电28 nm近似DFT IP,并开展RF-SoC验证

M. S. de Silva, R. J. Cintra, A. Madanayake

Electronics · 2026-11-07

A Claude-based agent designs approximate-DFT digital IP around sparse butterfly factorizations, targeting TSMC 28 nm CMOS with RF-SoC verification. The approximate transforms are described as requiring O(N) multiplication operations, making arithmetic structure a central design constraint. The work provides a domain-specific case for examining AI-assisted HDL generation and hardware verification. 研究使用基于Claude的智能体,围绕稀疏蝶形分解设计近似离散傅里叶变换数字IP,面向台积电28 nm CMOS并开展RF-SoC验证。这类近似变换所需的乘法运算量被描述为O(N),因此算术结构成为设计中的核心约束。该工作为考察AI辅助HDL生成与硬件验证提供了一个专门领域的案例。

Signal-Integrity-Aware Training and Readout for Memristor Crossbar Arrays 面向忆阻器交叉阵列的信号完整性感知训练与读出

Han-Zhi Ma, Jia-Rui Qiu, Ye Shi, et al.

IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems · 2026-11-01

A parasitic-aware training and inference scheme combines fast partial-element equivalent-circuit modeling with a readout rule based on quasi-steady-state arrival time. Matrix decomposition and sparse techniques accelerate array modeling, while dynamic unstructured pruning and parasitic correction address data-dependent signal distortion. The method links interconnect behavior to model deployment, targeting delay, crosstalk, and voltage-drop errors in memristor crossbars. 一套考虑寄生效应的训练与推理方案,将快速部分元等效电路建模与基于准稳态到达时间的读出规则相结合。矩阵分解和稀疏技术用于加速阵列建模,动态非结构化剪枝与寄生效应校正则处理数据相关的信号失真。该方法将互连行为纳入模型部署,针对忆阻器交叉阵列中的延迟、串扰及压降误差。

Newsletter 邮件订阅

Daily semiconductor briefing. 每日半导体简报。