Circuits & Architecture
电路与架构
D. Zhang, K. Ma, J. Zhang, et al.
arXiv:2608.07963 · 2026-08-08
A three-stage SiGe:C-BJT and LTC6431 preamplifier board delivers 115.27 mV ns/fC charge gain across a 0.5-25 fC input range and 34.0-594.3 MHz bandwidth. Measurements with LGAD, PIN, and 3D silicon detectors reach timing resolutions of 36.41 ps, 76.10 ps, and 39.40 ps, respectively. The board is experimentally validated, but it is a discrete readout implementation rather than a monolithic front end.
这块三级前置放大板采用SiGe:C BJT与LTC6431，在0.5-25 fC输入范围内实现115.27 mV ns/fC电荷增益，带宽为34.0-594.3 MHz。搭配LGAD、PIN与3D硅探测器时，实测时间分辨率分别达到36.41 ps、76.10 ps与39.40 ps。该方案经过实验验证，但仍是分立式读出实现，并非单片集成前端。
H. Zheng, X. Liu
IEEE Journal of Solid-State Circuits · 2026-08-01
This chopper-stabilized ECG analog front end combines time-division multiplexing with new impedance-boosting and DC-servo loops to reach 142.8 GOhm input impedance. The ASIC can trade channel count against input-referred noise by reconfiguring among four-, two-, and one-channel modes. The design targets electrode-offset and leakage problems that often dominate wearable biopotential interfaces, with measured silicon reported in JSSC.
这款斩波稳定ECG模拟前端结合时分复用、新型阻抗提升环路与DC伺服环路，实现142.8 GOhm输入阻抗。ASIC可在4通道、2通道与1通道模式间重构，以通道数换取更低的输入参考噪声。设计针对可穿戴生物电接口中常见的电极失调与漏电问题，JSSC论文给出了实测芯片结果。
AI Accelerators & Compute-in-Memory
AI加速器与存算一体
N. Jimenez Olalla, A. Long Ching Ip, M. Baumann, et al.
Neuromorphic Computing and Engineering · 2026-08-06
A physical memristor-and-op-amp neuron implements a biologically inspired local learning rule and learns XOR, AND, and OR in a small network. Moving adaptation into the circuit addresses a core limitation of neuromorphic systems that still depend on offline software training. The hardware proof is limited to elementary logic; the compressed-MNIST scaling result comes from SPICE rather than a larger fabricated network.
一个由忆阻器与运算放大器构成的物理神经元实现了受生物机制启发的局部学习规则，并在小型网络中学会XOR、AND与OR。把适应过程放入电路，可直接触及当前神经形态系统仍依赖离线软件训练的核心限制。硬件验证仅覆盖基础逻辑，压缩MNIST的扩展结果仍来自SPICE仿真，而非更大规模的已制造网络。
P. Sun, M. Ikeda
IEEE Transactions on VLSI Systems · 2026-08-01
This accelerator targets the less-studied client side of CKKS homomorphic encryption, unifying NTT and FFT multiplication while generating twiddle factors dynamically instead of storing large tables. Evaluations on a TSMC 28 nm ASIC and Xilinx UltraScale+ FPGA show more than 40% lower area or gate count and a 1.2x end-to-end speedup against representative hardware baselines. The gains are practical but moderate at system level, making the memory and area reductions as important as raw latency.
该加速器面向研究较少的CKKS同态加密客户端，将NTT与FFT乘法统一，并动态生成旋转因子，从而避免存储大型常量表。在TSMC 28 nm ASIC与Xilinx UltraScale+ FPGA上的评估显示，相较代表性硬件基线，面积或门数降低超过40%，端到端速度提升1.2倍。系统级速度增益较为温和，因此存储与面积节省与延迟改善同样重要。
L. Martis, G. Leone, L. Raffo, et al.
IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems · 2026-08-01
SYNtzulA integrates a RISC-V soft core and sparse SNN accelerator in a 6.8 mm2 SoC built with the open IHP-SG13G2 130 nm PDK and OpenROAD flow. The chip runs at up to 125 MHz, reaches 2 GSOP/s, and reports 36.5 pJ per synaptic operation. Its older node and modest throughput limit direct comparison with commercial neuromorphic processors, but silicon from a fully open flow makes the platform unusually reproducible.
SYNtzulA在6.8 mm2 SoC中集成RISC-V软核与稀疏SNN加速器，采用开放的IHP-SG13G2 130 nm PDK和OpenROAD流程实现。芯片最高运行于125 MHz，达到2 GSOP/s，单次突触操作能耗为36.5 pJ。较老制程与有限吞吐量使其难以直接对标商用神经形态处理器，但完全开放流程产出的实测芯片具备少见的可复现性。
R. Ansari, M. Wielgosz
arXiv:2608.00595 · 2026-08-01
A time-multiplexed spike feeder, localized weight memories, and pipelined argmax readout raise an Artix-7 SNN accelerator's maximum clock from 13.3 MHz to 167 MHz. The measured implementation processes an image in 82 microseconds using a 784-64-10 network. The large frequency gain addresses a real routing bottleneck, but MNIST and the small topology leave efficiency on modern workloads unresolved.
通过时分复用脉冲输入、本地化权重存储与流水线argmax读出，这款Artix-7 SNN加速器的最高频率从13.3 MHz提升至167 MHz。实测实现采用784-64-10网络，单张图像处理延迟为82微秒。频率提升解决了真实的布线关键路径，但MNIST与小型网络拓扑仍不足以说明其在现代工作负载上的效率。