Documents

DOI

To overcome the increasing sensitivity to variability in nanoscale integrated circuits, operation parameters (e.g., supply voltage) are adapted in a customized way exclusively to each chip. AVS is a standard industrial technique which has been adopted widely to compensate for process, voltage, and temperature variations as well as power optimization of integrated circuits. For cost and complexity reasons, AVS techniques are usually implemented by means of on-chip performance monitors (so-called PMBs) allowing fast performance evaluation during production or run time. Such on-chip monitoring approaches estimate operation parameters either based on responses from performance monitors with no interaction with the circuit or by monitoring the actual critical paths of the circuit. In this thesis, we focus on AVS techniques, which estimate operation parameters using responses from on-chip performance monitors with no interaction with the circuit during production. We discuss the challenges that these monitoring methodologies face with decreasing node sizes, in terms of accuracy and effectiveness. We show that the accuracy of these approaches is design dependent, and requires up to 15% added design margin. In addition, we show using silicon measurements of a nanometric FD-SOI device that the required design margin is above 10% of the clock cycle, which leads to significant waste of power. In this thesis, we introduce the new method of using delay test patterns including TF, SDD, and PDLY test patterns for application of AVS during IC production. The proposed method is able to eliminate the need for PMBs, while improving the accuracy of performance estimation. The basic requirement of using delay-based AVS is that there should be a reasonable correlation between the frequency the chip can attain while passing all delay test patterns and the actual frequency of the chip. Based on simulation results of ISCAS’99 benchmarks with a 28 nm FD-SOI library, using delay test patterns result in an error of 5.33% for TF testing, an error of 3.96% for SDD testing, and an error as low as 1.85% using PDLY testing. Accordingly, PDLY patterns have the capacity to achieve the lowest error in performance estimation, followed by SDD patterns and finally TF patterns. We performed the same analysis using a 65 nm technology node, which showed the same results. We also did two different silicon measurements on a 28 nm FD-SOI CPU to investigate the effectiveness of the TF-based approach. The results of the first case study on real silicon comparing the performance estimation using functional test patterns and the TF-based approach show a very close correlation between the two, which proves the effectiveness of the TF approach. The second case study compares the accuracy of voltage estimation using PMBs and the TF-based approach. The results show that the PMB approach can only account for 85% of the uncertainty in voltage measurements, which results in considerable power waste. In comparison, the TF-based approach can account for 99% of that uncertainty, thereby providing the ability to reducing that wasted power.
Original languageEnglish
Awarding Institution
Supervisors/Advisors
Award date21 Nov 2018
DOIs
Publication statusPublished - 2018

    Research areas

  • Adaptive voltage scaling, process variations, performance estimation, process monitoring boxes, delay testing, transition fault testing, path delay testing

ID: 47220158