An Accurate Interconnect Test Structure for Parasitic Validation in On-Chip Machine Learning Accelerators

TL;DR

This paper introduces a compact, easily embeddable test structure for accurately measuring interconnect parasitics in on-chip machine learning accelerators, validated on TSMC 28nm and 7nm processes.

Contribution

It presents a novel, compact test structure for interconnect parasitic measurement that is suitable for real chip implementation and validation.

Findings

Validated on TSMC 28nm process

Modified for TSMC 7nm process to identify process issues

Developed empirical model for RC parasitic estimation

Abstract

For nanotechnology nodes, the feature size is shrunk rapidly, the wire becomes narrow and thin, it leads to high RC parasitic, especially for resistance. The overall system performance are dominated by interconnect rather than device. As such, it is imperative to accurately measure and model interconnect parasitic in order to predict interconnect performance on silicon. Despite many test structures developed in the past to characterize device models and layout effects, only few of them are available for interconnects. Nevertheless, they are either not suitable for real chip implementation or too complicated to be embedded. A compact yet comprehensive test structure to capture all interconnect parasitic in a real chip is needed. To address this problem, this paper describes a set of test structures that can be used to study the timing performance (i.e. propagation delay and crosstalk) of various interconnect configurations. Moreover, an empirical model is developed to estimate the actual RC parasitic. Compared with the state-of-the-art interconnect test structures, the new structure is compact in size and can be easily embedded on die as a parasitic variation monitor. We have validated the proposed structure on a test chip in TSMC 28nm HPM process. Recently, the test structure is further modified to identify the serious interconnect process issues for critical path design using TSMC 7nm FF process.