This tutorial paper surveys the past 20 years of near/sub-threshold digital integrated circuit design. Most of the chips have been highly characterized for voltage scaling down to near/sub-threshold, and a significantly custom design effort has been reported to achieve reliable operation. In this tutorial, we address the challenges of process variations and discuss the circuits and methods used over the years to minimize this impact. We discuss the advantages of standard-cell library design and provide a more involved pruning method to improve performance and robustness. Finally, we discuss the developments of the usually ignored power delivery techniques for near/sub-threshold circuits. We motivate the use of voltage stacking as a new power delivery technique for near/sub-threshold. This paper provides the basic enablement approaches for designing a chip operating in the near/sub-threshold region based on our experience.
19 October 2021 in IEEE Transactions on Biomedical Circuits and Systems
We present a systematic evaluation and optimization of a complex bio-medical signal processing application on the BrainWave prototype system, targeted towards ambulatory EEG monitoring within a tiny power budget of $< $ 1 mW. The considered BrainWave processor is completely programmable, while maintaining energy-efficiency by means of a Coarse-Grained Reconfigurable Array (CGRA). This is demonstrated through the mapping and evaluation of a state-of-the-art non-convulsive epileptic seizure detection algorithm, while ensuring real-time operation and seizure detection accuracy. Exploiting the CGRA leads to an energy reduction of 73.1%, compared to a highly tuned software implementation (SW-only). A total of 9 complex kernels were benchmarked on the CGRA, resulting in an average 4.7 × speedup and average 4.4 × energy savings over highly tuned SW-only implementations. The BrainWave processor is implemented in 28-nm FDSOI technology with 80 kB of Foundry-provided SRAM. By exploiting near-threshold computing for the logic and voltage-stacking to minimize on-chip voltage-conversion overhead, additional 15.2% and 19.5% energy savings are obtained, respectively. At the Minimum-Energy-Point (MEP) (223 $\mu \rm {W}$ , 8 MHz) we report a measured state-of-the-art 90.6% system conversion efficiency, while executing the epileptic seizure detection in real-time.
2 April 2003 in IEEE Transactions on Very Large Scale Integration (VLSI) Systems
Functional yield is a term used to describe the percentage of dies on a wafer that are not affected by catastrophic defects. Within the interconnect these defects are usually caused by particle contamination and are divided into bridging defects, which join adjacent wires and cuts, which result in broken wires. Functional yield is therefore determined by the geometry of the routing channels, how these channels are filled with wire and the distribution of defect sizes. Since the wire spacing and width are usually fixed and the distribution of defects within a mature production facility is well known, the problem reduces to estimating individual wire lengths for cuts and to estimating the overlapping distance that two wires share in neighboring sections of the routing grid for bridges. Previous work in this area has analyzed the problem by assuming that all wiring tracks are occupied with wire, leading to overestimates for the probability of failure due to both cuts and bridges. This paper utilizes statistical models of the placement/routing process to provide a more realistic approach for cut and bridge yield estimation. A comparison of the predicted probability of failure within each wiring layer with postlayout data indicate an average error of 20% for cuts and 26% for bridges.
The advent of deep submicron technologies brings new challenges to digital circuit design. A reduced threshold voltage (VT) and power supply (Vdd) in addition to process variabilities have a direct impact on circuit design. In a semiconductor environment, it is conventionally thought that parametric yield is high and stable and that the main yield losses are functional. Although functional yield remains the focus of attention, modern and future circuits may not have the presumed high parametric yield. We present a study that compares the tolerance to process variability of various design families for metrics including timing and power consumption under VT-Vdd scalability using a NAND gate as a test vehicle. The fundamental limitations to the scaling of the supply voltage due to the statistical variation of MOS VT are investigated and defined. The four logic families under study are: static CMOS, Differential Complementary Voltage Swing Logic (DCVSL), Domino and Pass Logic.
Flip-flops and latches are two options to construct pipelines in digital integrated circuits (ICs). In this paper, the implications for converting a flip-flop based design to a latch-based design are investigated by performing timing and power analysis. Design flows are also proposed to convert a flip-flop based design to a latch-based design as well as a latch/flip-flop-mixed design. With a new retiming strategy, the optimum operating condition is identified for both the latch based design and the mixed design, where the maximum time borrowing or performance enhancement can be obtained. Compared to the flip-flop based design, 48% and 45% frequency boosting are achieved by the latch based design and the mixed design, respectively. While maintaining the same performance as the flip-flop based design with the aid of supply voltage scaling, the latch based design and the mixed design reduce the power consumption by 21% and 16%, respectively, in an industrial 28-nm FDSOI CMOS technology.
23 January 2023 in IEEE Transactions on Circuits and Systems II: Express Briefs
Data-retention flip-flops (DR-FF) provide an efficient state retention capability for any processor with frequent active-to-sleep mode transitions. This paper proposes new ultra-low-power balloon-based DR-FFs operating over a wide supply voltage range achieving a sleep mode power consumption as low as 380 fW, an improvement of 185× over the previously reported CMOS DR-FFs. The proposed DR-FFs have significantly lower transition mode energy consumption compared to prior-art non-volatile DR-FFs (> 18×). The proposed DR-FFs consume the lowest total energy (sleep, store and restore modes) for sleep mode times less than 520 ms. Additionally, the designed DR-FFs are comparable with most of the NV-FFs in literature for sleep times of up to 100 s, while eliminating the drawbacks of Non-volatile data retention FFs (higher/multiple supplies, integration/endurance issues).