Introduction
As transistors shrink, they become increasingly susceptible to "aging" mechanisms like Bias Temperature Instability (BTI) and Hot-Carrier Degradation (HCD), which slow down circuits over time. To prevent chip failure, engineers typically add "timing guardbands"—extra slack in the clock cycle—based on worst-case simulations. However, these simulations often rely on data from simple inverters or single-gate transistors. Real-world processors are built from complex "standard cells" (like And-Or-Inverters) and long, heterogeneous logic paths. This paper argues that relying on simple inverter models creates a "pessimism gap": designers either over-design chips, making them slower than necessary, or under-design them, risking premature failure.
Methodology
The researchers designed and manufactured a custom test chip, dubbed "KIPT," using a 1.2V 65nm CMOS process. Their approach moved beyond standard testing in two major ways:
- Complex Cell-Based Oscillators: Instead of just inverters, they built Ring Oscillators (ROs) out of complex combinational cells like AOI (And-Or-Inverter) and OAI (Or-And-Inverter), as well as full-adders.
- Path-Based Oscillators: They extracted actual critical logic paths from two open-source processors (PULP and RISC-V) and converted these paths into oscillators. This allowed them to measure how a "representative" slice of a CPU ages.
To accelerate the aging process, the chip was subjected to high-voltage stress (up to 2.5V) at 80°C. The study distinguished between DC stress (measuring BTI alone) and AC stress (measuring the combined impact of BTI and HCD).
Results and Discussion
The experimental results confirm that inverter-based oscillators—the industry standard—are unreliable proxies for complex logic. They tend to underestimate degradation under static (DC) stress but overestimate it under switching (AC) stress.
Key findings from the characterization include:
- HCD Vulnerability Varies Wildly: NAND and OAI cells are particularly sensitive to switching stress, with degradation nearly doubling when moving from DC to AC stress. In contrast, AOI and NOR cells only see a ~40% increase.
- The "Averaging Effect" of Logic Paths: While individual cell types show high variation in aging (ranging from 7% to 14% degradation), actual processor logic paths show much lower variation (hovering around 11%). The different cells in a path essentially "average out" each other's aging characteristics.
- Frequency Loss: After 10,000 seconds of stress, the PULP processor path RO showed a frequency drop from ~103.7 MHz to 91.6 MHz.
Conclusion
This study provides a "ground truth" that can be used to calibrate SPICE models, potentially allowing designers to reduce timing guardbands and reclaim lost performance. While the study was performed on 65nm CMOS, the principles of cell-specific aging are universal. Future work will need to explore modern FinFET nodes (7nm, 5nm) and the recovery effects of BTI when stress is removed.
Summary updated