跳到正文
National Science Review· Wang·· 8 天前AI 评分58

后端工艺集成碳纳米管模块压缩硅逻辑与时钟网络延迟

Back-end-of-line-integrated carbon nanotube modules for acceleration of silicon logic and clock networks

AI 导读

研究在 National Science Review 发表,提出 τDrive 方案,在硅 CMOS 后端工艺中只于负载关键的驱动级插入碳纳米管时序单元,以压缩逻辑与时钟路径的主导时间常数。

AI 深度导读基于原文整理的分节解读
背景
先进硅逻辑集成电路中,几何微缩已不再带来成比例的时间收益,工作频率越来越受限于高电容、受扇出制约的逻辑路径与时钟网络。新兴的 τ-scaling 视角把这一限制重新表述为跨越驱动器—互连—负载整个系统的时间常数问题。
做法
作者提出 τDrive,一种单片三维(M3D)碳纳米管(CNT)有源时序覆盖层,选择性地压缩硅逻辑与时钟路径中占主导的 τ。与主要缩短互连的拓扑折叠路线不同,τDrive 保留底层硅 CMOS 结构,只在负载关键的驱动器级插入后道工序(BEOL)集成的 CNT 时序单元。实验上以 180-nm Si/500-nm CNT 平台验证低温 CNT-on-Si 集成;仿真上则在 3-nm 硅 CPU 上使用经过校准的 10-nm CNT CMOS 标准单元。
结论
180-nm Si/500-nm CNT 实验平台将支持 CNT 辅助的硅环形振荡器在扇出(FO)为 5 时实现 294% 的输出摆幅改善和 30% 的有效电流提升。在 3-nm 硅 CPU 仿真中,选择性 CNT 插入使 FO = 39 的局部延迟和跳变分别降低 87.5% 和 89.5%;单个单元替换使整颗 CPU 逻辑路径的最大频率(Fmax)提高 2.3%;选择性时钟树增强把频率上限从 2.85 GHz 提升到 2.93 GHz,而时钟树综合(CTS)功耗仅增加 0.01 mW。
意义
结果把 CNT-on-Si M3D 确立为一种低侵入、可垂直接入的驱动资源,用于几何微缩之后的 τ 压缩,同时也提供了一条新的逻辑折叠路径。
局限
CNT 辅助环形振荡器的 294% 摆幅改善与 30% 有效电流提升基于 180-nm Si/500-nm CNT 实验平台,并表述为「将 enable」;3-nm CPU 的延迟、Fmax 与时钟树结果来自使用校准 10-nm CNT CMOS 标准单元的仿真,而非流片实测。
下一步
实验平台「将 enable」CNT 辅助硅环形振荡器达到上述摆幅与电流指标,即该验证尚待完成;CNT-on-Si M3D 作为后几何 τ 压缩的驱动资源与逻辑折叠路径,仍有待进一步确立。

原文摘要

出版方登记 · Crossref

Geometric transistor scaling no longer delivers proportional timing gains in advanced silicon logic integrated circuits (ICs), where operating frequency is increasingly bounded by high-capacitance, fanout-limited logic paths and clock networks. The emerging τ-scaling perspective reframes this limitation as a time-constant problem across the driver-interconnect-load system. Here we report τDrive, a monolithic three-dimensional (M3D) carbon-nanotube (CNT) active timing overlay that selectively compresses the dominant τ of silicon logic and clock paths. Unlike topology-folding approaches that primarily shorten interconnects, τDrive preserves the underlying silicon complementary metal-oxide-semiconductor (CMOS) fabric and inserts back-end-of-line (BEOL) integrated CNT timing cells only at load-critical driver stages. A 180-nm Si/500-nm CNT experimental platform verifies low-temperature CNT-on-Si integration, and will enable CNT-assisted Si ring oscillators to achieve a 294% output-swing improvement and 30% effective-current enhancement with a fanout (FO) of 5. In simulations of a 3-nm Si central processing unit (CPU) using calibrated 10-nm CNT CMOS standard cells, selective CNT insertion reduces FO = 39 local delay and transition by 87.5% and 89.5%, respectively, and a single-cell replacement improves the full-CPU logic-path maximum frequency (Fmax) by 2.3%, while selective clock-tree enhancement raises the frequency bound from 2.85 to 2.93 GHz with only a 0.01 mW increase in clock-tree synthesis (CTS) power. These results establish CNT-on-Si M3D as a low-intrusion, vertically accessible drive resource for post-geometric τ compression, as well as a novel logic-folding approach.

页面未能读取全文,以上为出版方登记的摘要;方法与图表见原文

本页是 AI 精读,不转载论文原文;完整论文(含方法、图表与参考文献)见原文

来源:National Science Review · doi.org