Near-threshold computing is essential for energy-efficient operation of VLSI systems, but wide performance variation and nonlinearity to process variations block the proliferation. To cope with this, in this article, we propose a holistic hardware performance monitoring methodology for accurate timing prediction in a near-threshold voltage regime. Precisely, 1) we formulate the problem of finding an efficient configuration of monitoring circuits into an instance of optimal experiment design problem and 2) propose a new timing prediction flow, consisting of statistical estimation of FEOL and BEOL process variations and a neural network-based timing inference model. For accurate control of timing margin and overcoming simulation-silicon discrepancies, 3) we introduce uncertainty learning in the prediction model construction and calibrate it through transfer learning. Furthermore, 4) we avoid time-consuming SPICE simulations in our methodology by employing efficient but accurate surrogate models. Through simulations using a 28-nm industry PDK and DK characterized at 0.6-V operation, it is shown that our methodology is highly effective, reducing the average prediction pessimism of maximum delay by 77.9% over conventional signoff results while respecting target prediction yield. Besides, for test chips fabricated using a 10-nm process, we demonstrated that our holistic approach from design to postsilicon phase in conjunction with adaptive voltage scaling reduces dynamic power consumption by 28.2%–28.8% on average, in comparison with typical supply voltage operation.

An elaborate hardware performance monitor (HPM) has become increasingly important for handling huge performance variation of near-threshold computing and recent process technologies. In this paper, we propose a new approach to the problem of predicting critical path delays (CPDs) using HPM. Precisely, for a target circuit or system, we formulate the problem of finding an efficient combination of ring oscillators (ROs) for accurate prediction of CPDs on the circuit as a mixed integer second-order cone programming and propose a method of minimizing the total number of ROs for a given pessimism level of prediction. Then, we propose a prediction flow of CPDs through statistical estimation of process parameters from measurements of the customized HPM and machine learning based delay mapping from the estimation. For a set of benchmark circuits tested using 28nm PDK and 0.6V operation, it is shown that our approach is very effective, reducing the pessimism of CPDs and minimum supply voltages by 6.7~52.9% and 20.6~50.8% over those of conventional approaches, respectively.

In mobile systems, the problems of short battery life and increased temperature are exacerbated by wasted leakage power. Leakage power waste can be reduced by power-gating a core while it is stalled waiting for a resource. In this work, we propose and model memory access power gating (MAPG), a low-overhead technique to enable power gating of an active core when it stalls during a long memory access. We describe a programmable two-stage power gating switch design that can vary a core's wake-up delay while maintaining voltage noise limits and leakage power savings. We also model the processor power distribution network and the effect of memory access power gating on neighboring cores. Last, we apply our power gating technique to actual benchmarks, and examine energy savings and overheads from power gating stalled cores during long memory accesses. Our analyses show the potential for over 38% energy savings given “perfect” power gating on memory accesses; we achieve energy savings exceeding 20% for a practical, counter-based implementation.

Double patterning lithography (DPL) is in current production for memory products, and is widely viewed as inevitable for logic products at the 32 nm node. DPL decomposes and prints the shapes of a critical-layer layout in two exposures. In traditional single-exposure lithography, adjacent identical layout features will have identical mean critical dimension (CD), and spatially correlated CD variations. However, with DPL, adjacent features can have distinct mean CDs, and uncorrelated CD variations. This introduces a new set of “bimodal” challenges for timing analysis and optimization. We assess the potential impact of bimodal CD distribution on timing analysis and guardbanding, and find that the traditional “unimodal” characterization and analysis framework may not be viable for DPL. We propose new bimodal-aware timing analysis and optimization methods to improve timing yield of standard-cell based designs that are manufactured using DPL. Our first contribution is a DPL-aware approach to timing modeling, based on detailed analysis of cell layouts. Our second contribution is an integer linear programming-based maximization of “alternate” mask coloring of instances in timing-critical paths, to minimize harmful covariance and performance variation. Third, we propose a dynamic programming-based detailed placement algorithm that solves mask coloring conflicts and can be used to ensure “double patterning correctness” after placement or even after detailed routing, while minimizing the displacement of timing-critical cells with manageable engineering change order (ECO) impact. With a 45 nm library and open-source design testcases, our timing-aware recoloring and placement optimization together achieve up to 271 ps (respectively, 55.75 ns) reduction in worst (respectively, total) negative slack, and 70% (respectively, 72%) reduction in worst (respectively, total) negative slack variation, respectively.

In sub-100nm CMOS processes, delay and leakage power reduction continue to be among the most critical design concerns. We propose to exploit the recent availability of fine-grain exposure dose control in the step-and-scan tool to achieve both design-time (placement) and manufacturing-time (yield-aware dose mapping) optimizations of timing yield and leakage power. Our placement and dose map co-optimization can improve both timing yield and leakage power of a given design. We formulate the placement-aware dose map optimization as quadratic and quadratic constraint programs which are solved using efficient quadratic program solvers. In this paper, we mainly focus on the placement-aware dose map optimization problem; in the Appendix, we describe a complementary but less impactful dose map-aware placement optimization based on an efficient cell swapping heuristic. Experimental results show noticeable improvements in minimum cycle time without leakage power increase, or in leakage power reduction without degradation of circuit performance.

Double patterning lithography (DPL) is in current production for memory products, and is widely viewed as inevitable for logic products at the 32 nm node. DPL decomposes and prints the shapes of a critical-layer layout in two exposures. In traditional single-exposure lithography, adjacent identical layout features will have identical mean critical dimension (CD), and spatially correlated CD variations. However, with DPL, adjacent features can have distinct mean CDs, and uncorrelated CD variations. This introduces a new set of dasiabimodalpsila challenges for timing analysis and optimization. We assess the potential impact of DPL on timing analysis error and guard banding, and find that the traditional dasiaunimodalpsila characterization and analysis framework may not be viable for DPL. For example, using 45 nm models, we find that different DPL mask layout solutions can cause 50 ps skew in clock distribution that is unseen by traditional analyses. Different mask layouts can also result in 20% or more change in timing path delays. Such results lead to insights into physical design optimizations for clock and data path placement and mask coloring that can help mitigate the error and guardband costs of DPL.

Double patterning lithography (DPL) is a likely resolution enhancement technique for IC production in 32nm and below technology nodes. However, DPL gives rise to two independent, uncorrelated distributions of linewidth on a chip, resulting in a 'bimodal' linewidth distribution and an increase in performance variation. [13] suggested that new physical design mechanisms could reduce harmful covariance terms that contribute to this performance variation.

In sub-100nm CMOS processes, delay and leakage power reduction continue to be among the most critical design concerns. We propose to exploit the recent availability of fine-grain exposure dose control in the stepper to achieve both design-time (placement) and manufacturing-time (yield-aware dose mapping) optimizations of timing yield and leakage power. Our placement and dose map co-optimization can simultaneously improve both timing yield and leakage power of a given design. We formulate the placement-aware dose map optimization as a quadratic program, and solve it using an efficient quadratic programming solver. In this paper, we mainly focus on the placement-aware dose map optimization problem; in the Appendix, we describe the complementary but less impactful dose map-aware placement optimization, and an efficient cell swapping heuristic. Experimental results are promising: with typical 90nm stepper (ASML Dose Mapper) parameters, we achieve more than 8% improvement in minimum cycle time of the circuit without any leakage power degradation.