Low-Power Timing Optimization Via Power And Clock Gating In Advanced Nodes: Techniques, Challenges and Future Directions

Low-Power Timing Optimization Via Power And Clock Gating In Advanced Nodes: Techniques, Challenges and Future Directions

Ujjwal Singh

Electrical and Computer Engineering Cornell University, Ithaca, USA

Abstract – The continued scaling of semiconductor technology to advanced nodes at 7 nm, 5 nm, and 3 nm has intensified challenges in power leakage, timing convergence, and process variability for modern System-on-Chip (SoC) designs. Among the most effective low-power design strategies are power gating and clock gating, which reduce both dynamic and static power consumption while preserving performance objectives. However, implementing these techniques at advanced nodes introduces substantial complexity in timing optimization, clock tree synthesis (CTS), wake-up delay management, and sensitivity to process variations. This paper presents a structured review of the underlying principles, architectural implementations, and optimization strategies for these techniques, with particular emphasis on their impact on timing closure, design verification, and physical implementation. The review synthesizes journal and conference literature spanning device, circuit, architecture, and EDA (Electronic Design Automation)-level perspectives, organized thematically to address clock gating methodologies, power gating with retention strategies, timing closure constraints, and AI/ML (Artificial Intelligence/Machine Learning)-assisted optimization approaches. Additionally, the paper examines how EDA toolchain integration and machine learning (ML) enable adaptive low-power optimization across the design space. Finally, emerging technologies including AI/ML-assisted timing prediction, multi-domain power-intent specifications using UPF (Unified Power Format) and CPF (Common Power Format), and low-power architectures for 3D (Three-Dimensional) integrated circuits (ICs) are explored as key enablers for next-generation SoC design. Together, these perspectives provide a foundation for understanding current state-of-the-art approaches and the trajectory of low-power timing optimization in advanced-node semiconductor design.

Keywords – Low-Power Design, Timing Optimization, Clock Gating, Power Gating, Advanced Nodes, EDA, FinFET (Fin Field Effect Transistor), SoC Design

1. INTRODUCTION

   Further reduction of semiconductor technology to smaller process nodes of 7 nm, 5 nm, and 3 nm has been a game- changer to the contemporary concept of an integrated circuit (IC) design in that it has provided more transistor densities, greater performance, and increased functionality in smaller System-on-Chip (SoC) designs. Nonetheless, the same improvements have also strengthened the problems related to the dissipation of power, timing closure, and signal integrity, and have become the performance and energy efficiency bottlenecks that can be attained [1], [2]. The dynamic power

   induced by clock activity and switching capacitance, and the static power, which is caused by the subthreshold leakage and the gate leakage, have become significant in the overall power efficiency [3], [4] due to smaller devices. This has resulted in the balance of power and performance efficiency taking center stage in the design objectives of the advanced node technologies. Clock gating and power gating [5], [6] are amongst the most useful and most commonly used low-power operation techniques, albeit without timing reliability degradation. Clock gating saves dynamic power by shutting down clock signals to things to blocks of idle circuitry, and power gating saves leakage power by shutting down the flow of power to idle logic blocks via a high voltage sleep transistor [7]. They are efficient, but they introduce new challenges in terms of timing optimization, clock skew, and wake-up latency, particularly when a multi-domain SoC with fine- grained power partitioning is taken into consideration [8]. Moreover, the sub-10 nm technology design closure is more complex due to process variation, IR drop, timing variation dependence on temperature, and strong optimization structures are required to combine power and timing analysis at various design levels [9], [10].

   To overcome them, current design flows are becoming based on automation of EDA and machine learning (ML) based optimization to enhance timing predictability and power of hierarchical SoC design [11]. ML-based EDA systems are capable of timing violations prediction, efficient addition of clock gating, and optimization of power domain configurations with changing workloads [12]. The standardized low-power design languages like the Unified Power Format (UPF) and the Common Power Format (CPF) may also be provided to provide the desired interoperability and consistency to the formulation of the power will across the design exercise. All this is an indicator of a paradigm shift in the design philosophy of low-power design in technologies based on nanoscale: intelligent and timing-conscious, and context-dependent. It is in this way that we will be reviewing the low-power timing optimization of the power and clock gating in advanced nodes in this paper. It reviews their theoretical experiences, plans of implementation, and performance trade-offs in sub-7 nm technologies. Furthermore, it is also associated with the current streams of EDA and the ML-based design strategies capable of dealing with the timing and power challenges that the strategies

   present. It attempts to get just one thing to consider the co- optimization of the power efficiency, timing reliability, and scalability of the next generation integrated circuits.

2. RELATED WORK

   The low-power implementation of high-CMOS nodes has been stimulated by literature in circuit design systems, architecture, and EDA automation. Some of the more recent articles focus on the concepts of fine-grained clock gating, power-domain partitioning, and machine-learning-based timing closure as the concepts to assist in ameliorating the dynamic and leakage power to more performance-needy demands. The earlier papers [10]-[12] focused more on optimization of gating efficiency and were obtained by logic- level control and hierarchy, but later papers [13]-[18] applied multi-objective optimization in conjunction with predictive modeling to obtain timing, energy, and reliability optimization. All those additions combined have led to the intent-based, predictive, and adaptive power-timing co- optimization needed to be used with sub-5 nm system-on-a- chip (SoC) designs. Xie et al. (2020) introduce PowerNet, a machine learning framework designed to accelerate dynamic IR-drop analysis, a critical power integrity verification step that directly influences timing closure in modern integrated circuit design. The work addresses two key limitations in existing approaches: the prohibitively high computational cost of commercial dynamic IR-drop analysis tools and the poor generalizability of previous supervised learning methods, which typically require retraining for each new design. PowerNet tackles these challenges through a transferable convolutional neural network architecture that generalizes across different designs while supporting both vectorless and vector-based dynamic IR-drop prediction. The methodology centers on two key innovations: preserving the spatial characteristics of power and voltage distributions, and constructing more comprehensive power features, including time-decomposed power maps, to more accurately capture the worst-case transient behavior of local IR-drop phenomena. Experimental results demonstrate that PowerNet achieves improved accuracy in vectorless IR-drop estimation compared to prior learning-based methods, while delivering significant runtime reductions relative to commercial tools. This performance enables faster design iterations during the critical late-stage closure phase. Beyond prediction, PowerNet has been integrated into an automated mitigation workflow that selectively reinforces the power delivery network in identified hotspot regions. This approach successfully reduces IR-drop violations in industrial designs while minimizing modifications to the power distribution network. The work illustrates how transferable machine learning models can substantially reduce the verification burden of power integrity analysis and enable more efficient, targeted design corrections in advanced-node system-on-chip development flows [10].

   Huang et al. (2021) provide a comprehensive survey of machine learning applications in electronic design automation, observing that the growing complexity of VLSI systems has sparked renewed interest in ML-based approaches to enhance both design productivity and quality of results. The authors

   organize existing research according to the EDA design hierarchy, highlighting that machine learning techniques have now permeated major stages of the design flowincluding logic synthesis, placement and routing, verification, and manufacturingwhere they frequently outperform traditional heuristic methods. The survey identifies four recurring themes in ML-for-EDA research: improving decision-making within design tools, predicting performance metrics, performing black-box optimization for design space exploration, and enabling automated design through deep learning and reinforcement learning techniques. This categorization provides valuable context for our work by illustrating how predictive models and optimization policies can enhance clock and power gating strategies, ultimately accelerating the co- optimization of timing and power in advanced-node system- on-chip designs [11]. Zhao et al. present a machine learning framework for multi-corner timing prediction that accelerates static timing analysis by predicting timing metrics at non- dominant process corners using results from a strategically selected subset of dominant corners. The central technical innovation is an iterative dominant-corner selection algorithm that leverages cross-corner timing correlations through correlation analysis, combined with nonlinear regression modelsincluding Ridge regression, multi-layer perceptrons, and Random Foreststo achieve high prediction accuracy while maintaining low computational overhead. Experimental results on industry benchmarks demonstrate that the framework achieves up to 98.2% prediction accuracy with timing errors below 10 picoseconds, while delivering more than a twofold speedup in timing closure runtime. This performance improvement is achieved by replacing traditional static timing analysis at most corners with fast machine learning inference, significantly reducing the verification burden in multi-corner design flows [12].

   Park (2023) present a clock-gating synthesis framework that improves dynamic-power reduction by combining power- aware selection with learning-driven flip-flop grouping at the synthesis stage. The first contribution is a selective clock- gating methodology that estimates expected savings using enable-signal probability analysis, including joint-probability effects, to avoid overestimating gating benefit under realistic activation behavior. The second contribution introduces a representation-learning approach that compresses long flip- flop activity sequences into compact embedding vectors using an LSTM autoencoder and a stacked denoising autoencoder, enabling scalable similarity-based grouping for clock-gate insertion without requiring prohibitively long simulations. Together, these components position clock-gating decisions as data-driven and power-quantified, improving the practicality and effectiveness of gating-centric low-power optimization in timing-feasible implementation flows [13]. Neto et al. propose FlowTune, an end-to-end logic-optimization exploration framework that reduces manual closure effort by automating the selection of optimization sequences across the implementation flow. The central technical idea is to cast flow tuning as sequential decision making and solve it using a domain-specific, multistage multi-armed bandit formulation that efficiently explores candidate transformations while remaining lightweight compared to high-overhead

   reinforcement learning approaches. FlowTune is designed to operate across stages and to incorporate downstream evaluation (including post-place-and-route assessment), enabling more reliable quality-of-result comparisons than synthesis-only tuning. Experimental results reported in the paper show that FlowTune improves QoR relative to baseline and handcrafted flows while maintaining practical runtime, making it a relevant enabling technique for adaptive EDA- driven timingpower co-optimization in advanced-node SoC design [14].

   Won et al. present a machine-learning-driven clock-gating synthesis framework that targets dynamic-power reduction while explicitly preserving timing feasibility and physical implementability. The central technical innovation is an embedding-based flip-flop grouping pipeline: a convolutional autoencoder compresses long flip-flop activity sequences into compact latent vectors, and CNN-based similarity ranking is used to form gating groups that maximize switching- correlation benefits. To ensure closure-quality insertion, the method enforces two key constraints(i) timing safety to prevent violations from added gating-logic delay and (ii) physical proximity to avoid excessive routing overhead. Experimental results reported in the paper show additional dynamic-power reduction over prior grouping approaches, while maintaining no timing violation attributable to gated- logic delay under the applied constraints, making it directly relevant to timing-conscious clock-gating in advanced-node flows [15].

   Yeh et al. propose a process-variation-aware power-gating control method that addresses the variability sensitivity and wake-up management challenges introduced by power gating. The key idea is to adapt the power-switch control timing based on detected process corner behavior, enabling robust sequencing of wake-up events while keeping surge current within a prescribed constraint. The technique introduces a corner-aware controller that selects appropriate buffer combinations to adjust the effective control delay, compensating for process-induced delay shifts that would otherwise increase wake-up variability and power integrity risk. Results presented by the authors demonstrate that the adaptive control can satisfy a surge-current constraint while maintaining appropriate power-switch timing across process variation, illustrating a practical reliability-oriented approach to power gating that complements timingpower co- optimization discussions in low-power SoC design [16].

   Lai et al. present BTI-Gater, an aging-resilient clock-gating methodology that addresses timing reliability degradation caused by workload-dependent NBTI/PBTI aging in clock distribution networks. The central insight is that conventional clock gating can bias the clocks idle duty ratio, leading to asymmetric threshold-voltage shifts and additional clock skew over lifetime. To mitigate this, the authors propose two integrated clock-gating (ICG) cell circuits that alternate the clock idle state between logic-high and logic-low, targeting a balanced ~50% duty ratio that reduces BTI-induced skew. They further provide a selection methodology for choosing

   appropriate ICG configurations at the architecture/microarchitecture level and discuss software sleep scheduling to avoid worst-case aging patterns. The results show meaningful reduction of BTI-induced clock skew and enable tighter timing margins without excessive guard- banding, aligning directly with timing-conscious clock-gating in advanced-node designs [17]. Kwon and Shin propose a recurrent U-Net framewok for fast prediction of dynamic IR- drop in power distribution networks, targeting the high runtime cost of signoff-grade power integrity analysis. The key technical contribution is a recurrent chaining of U-Net blocks that propagates intermediate feature information across time steps, improving modelling fidelity when decoupling capacitors are present and IR-drop behaviour is temporally dependent. The model uses layout-derived input maps (e.g., PDN resistance/capacitance proxies and distance features) to infer transient IR-drop distributions, enabling rapid hotspot identification. Experimental results show substantial speedup relative to commercial tools while maintaining bounded prediction error, and the approach improves accuracy compared with a non-recurrent U-Net under capacitor- influenced scenarios. In the context of low-power timing optimization, this work is relevant as an ML-enabled signoff accelerator that reduces power-integrity iteration overhead and indirectly supports faster timingpower convergence in

   advanced-node flows [18].

3. POWER AND TIMING ADVANCED NODES

   The extended term developments in the semiconductor technology to achieve sub-7 nm node to deliver tremendous performance and integration advantages, but at the expense of power consumption and timing management challenges. The speed/leakage/energy efficiency ratio has been among the most notable features in the present IC design because of the ever-present network of transistors. Contemporary integrated circuit design faces escalating challenges from both dynamic and leakage power consumption, increasingly stringent timing closure requirements, and amplified process variation effects. These factors collectively demand advanced optimization strategies capable of navigating the complex power- performance trade-offs inherent at various abstraction levels of the design process.

   1. Evolution of Power Challanges

      The relationship between power and timing in modern integrated circuits is illustrated in Figure 1. Technology and voltage scaling create inherent tradeoffs between power consumption and performance. Dynamic power scales with switching activity and capacitive load, while static power arises from various leakage mechanisms that have become dominant in advanced nodes. These power components, coupled through interconnect parasitics, jointly determine timing delay. The final timing slack must account for additional degradation from IR drop, temperature effects, aging, and PVT variations, highlighting the multi-dimensional nature of contemporary design optimization challenges. The aggressive scaling of threshold voltage (Vth) and gate oxide thickness in nanoscale CMOS technologies exacerbates static

      power consumption through enhanced leakage currents. Specifically, subthreshold leakage exhibits exponential dependence on Vth reduction, while gate oxide tunnelling current increases dramatically with thinner oxide layers, compounded by junction leakage contributions at elevated temperatures. Technology scaling below 10nm significantly exacerbates leakage mechanisms, with subthreshold leakage remaining the dominant component due to its exponential dependence on threshold voltage (Vth) reduction. Additionally, gate-induced drain leakage (GIDL) increases substantially due to high electric field concentrations at the drain junction, while gate tunnelling currents intensify with ultra-thin gate dielectrics. [19].

      Despite the superior electrostatic integrity afforded by FinFET and GAA transistor structures, which substantially alleviate short-channel effects compared to planar devices, static power dissipation persists as a critical design constraint under high- performance operating conditions and elevated thermal environments where subthreshold and junction leakage currents remain substantial. [20]. Given that clock distribution networks must serve millions of flip-flops with minimal skew and meet critical timing constraints, they represent one of the largest contributors to dynamic power in modern SoCs, often accounting for a big chunk of the total dynamic power dissipation. The high switching activity of clock signals, combined with the substantial capacitive load of extensive clock tree structures and buffers required to maintain signal integrity across the chip, results in significant energy expenditure. Furthermore, clock networks operate continuously at the system’s maximum frequency, unlike data paths that may experience lower average switching activity. This makes clock power optimization a high-impact strategy for reducing overall system energy consumption, with even modest improvements in clock tree efficiency translating to substantial energy savings at the chip level. [21].

      Figure 1: Overview of Power and Timing Interactions in Advanced Nodes

   2. Timing Closure Complexity

      At a higher level of process node, the timing closure problem has been a complicated one as the level of interconnect parasitics, IR drop, and process voltage temperature (PVT) variations are on the rise [22]. RC delay of interconnects is increasing disproportionately to metal pitches and aspect ratios as well, and wire delay has been a significant source of overall timing degradation. At the same time, the localized performance loss and timing non-conformance at power domains could be predetermined by IR drop, the loss of voltage in a resistive network of power delivery. These effects are also increased by electromigration (EM), as well as thermal gradients and time margins are distorted at nanometer scales. The sub-5 nm designers make an ill-hearted trade-off between timing slack, power consumption, and performance [23]. The higher supply voltage (VDD) could be much more effective in providing timing slack, but at the cost of dynamically consuming more power and having lower power at lower timing uncertainty and slower transitions. The aggressiveness of clock and power gating, likewise, will translate into longer latency of enabling/disabling the domain, which should be well scanned by traversing some corners of operation analysis using static timing analysis (STA). They are also to be co-optimized to perform correctly under worst-case conditions of PVT, in particular, high-speed PCs, GPUs, and AI accelerators.

   3. Overview of Power Optimization Techniques

      Initial low-power design techniques were device-based and circuit-based techniques of voltage scaling, MTCMOS, and DVFS in order to trade off energy and delay [24]. Nevertheless, such techniques are more susceptible in sub-7 nm nodes because of high leakage, parasitic, and timing jitter. With power-intent-driven flows and architecture-conscious flows, co-optimization has brought forth new strategies of clock and power gating. Clock gating minimizes dynamic power by disabling idle logic, and power gating minimizes leakage by disconnecting dead blocks. The integration of these techniques using UPF-based intent models and EDA tools with the assistance of the AI can lead to the saving of over 40 % of power without timing closure effect. Vo’s comparative study presents power gating as a practical leakage-reduction technique, with the main goal of cutting static power while managing trade-offs between wake-up time, power loss, and power-delay product across different gating approaches. The technique focuses primarily on static power, though it has dynamic and transitional implications during sleep-entry and wake-up events. Power gating is implemented at the circuit and block level through switch structures and gating topologies, with explicit timing effects because wake-up latency, charge equalization, and recovery behavior directly influence circuit responsiveness. A key strength is its comparison of different power-gating implementations using concrete metrics, making it valuable for design trade-offs. However, its scope is limited to power-gating variants rather than covering the full low- power taxonomy, and reported results hould be applied cautiously to modern process nodes without additional validation. The technique works best for leakage-sensitive blocks and systems that frequently switch between active and sleep states [8].

      The MLCAD survey by Rapp et al. frames AI and machine learning-assisted optimization as a broad EDA technique rather than a single circuit-level power-saving method, aiming to improve design-space exploration and predictive optimization across power, performance, and timing objectives. The technique indirectly addresses both dynamic and static power since ML models guide CAD decisions affecting multiple design metrics simultaneously. It operates at the EDA, design- space, and runtime optimization level, with generally positive timing impact through better prediction quality, faster convergence, and improved optimization guidance rather than direct transistor-level delay manipulation. Its main strength is breadth and relevance to modern CAD workflows, particularly for context-aware automation. As a survey paper, it provides strong support for trends and methodologies but lacks specific quantitative guarantees for individual power techniques. It’s most useful for understanding modern AI-assisted optimization approaches across placement, routing, timing, and power- aware CAD tasks [25]. Chandrakasan, Sheng, and Brodersen offer the foundational perspective on low-power CMOS design, where the core technique is voltage scaling alongside broader circuit and architectural strategies, with the primary objective of reducing power consumption while maintaining required throughput. The paper mainly addresses dynamic power, though it examines total power decomposition and low- power design trade-offs more broadly. The work spans architecture, logic, circuit, and technology considerations, with timing impact central to the discussion because lowering supply voltage reduces speed and drive capability, creating an explicit power-delay trade-off. A major strength is its rigorous first-principles treatment of low-power design and clear formulation of voltage-scaling benefits and penalties. The limitation is that it represents foundational work from 1992 and isn’t a direct source for modern node-specific implementation details like FinFET or gate-all-around constraints or advanced- node signoff practices. It’s best used as a theoretical foundation for low-power CMOS design and voltage-scaling-based optimization [26].

      Carver et al. describe power-intent modeling using the Si2 Common Power Format, where the technique focuses on formal specification of low-power design intent throughout the implementation and verification flow rather than direct power reduction. The primary goal is to unify and communicate power domains, states, switching behavior, and related constraints consistently across tools, indirectly targeting both dynamic and static power by enabling implementation of multiple power-saving techniques like power gating and DVFS. It operates at the EDA flow integration and methodology level, with indirect timing impact emerging through correct power-aware synthesis, implementation, and signoff interactions. Key strengths include cross-tool consistency, automation support, and reduced ambiguity when propagating low-power intent. Limitations include methodology learning curve, tool dependence, and being stronger for flow semantics than quantitative savings claims. It’s most appropriate as a methodology reference for power- aware RTL-to-signoff flows [27].

      Sachid, Khandelwal, and Hu examine body biasing in SOI FinFETs, where the technique modulates threshold voltage to improve the leakage-speed trade-off in low-power applications. The primary power type addressed is static leakage power, while timing and performance are affected through threshold

      adjustment. The work focuses on device-level SOI FinFET physics and gate-length dependence, though findings inform circuit-level body-bias strategies. The timing impact is indirect but meaningful because body bias changes threshold voltage and influences switching speed, with effectiveness depending on geometry, particularly gate length. Its strength is providing modern, device-grounded evidence for body-bias behavior in FinFET-era devices rather than relying on planar CMOS assumptions. The limitation is its device-focused scope rather than providing a complete digital implementation flow reference, so it should be paired with circuit and system-level sources for production methodology claims. It works best in FinFET low-power device and circuit design contexts where evaluating body-bias effectiveness is important [28].

      Kao, Chandrakasan, and Antoniadis present MTCMOS optimization through sleep-transistor sizing, aiming to achieve high performance in active mode and low leakage in sleep mode using threshold-partitioned design. The technique primarily addresses static power through leakage reduction, though implementation choices also influence dynamic behavior and active-mode performance. It operates at the transistor and circuit level, focusing on virtual-ground behavior and sleep-transistor sizing. The timing impact is explicit and closely tied to sizing decisions: undersized sleep devices increase resistance and degrade delay, while oversized devices create area and switching penalties. Its major strength is direct, practical treatment of MTCMOS trade-offs and a sizing- oriented design methodology. The limitation is that it reflects older process assumptions and doesn’t provide a complete modern signoff-flow or standard-cell-library methodology reference. It’s best used as a primary technical reference for understanding MTCMOS fundamentals and optimizing the leakage-delay trade-off [29]. Jairam et al. address clock gating as an ASIC-flow power optimization technique, aiming to reduce clock-related switching activity without altering functional behavior. The technique targets dynamic power, with implementation spanning system, RTL, gate-level, and design-flow integration, making it particularly valuable for practical ASIC methodology discussions. The timing impact depends on implementation details: clock-gating insertion interacts with clock-tree structure, gating cells, and downstream loads, meaning skew, hold time, and verification considerations become important in practice. Strengths include strong industry relevance, practical flow perspective, and useful quantitative insights about clock-network power composition and gating opportunities. The limitation is its tutorial or proceedings style rather than tightly controlled experimental format with universally applicable savings percentages. It’s most suitable for ASIC and SoC design discussions about clock-gating insertion and optimization methodology [30].

      Park et al. provide a modern DVFS-focused perspective, where the specific objective is accurate modeling of delay and energy overhead during DVFS transitions so runtime managers can make correct break-even decisions. The technique addresses both dynamic and static energy behavior during transitions, including losses from voltage regulator behavior, clock generation, lock times, and transition intervals. It operates at the system and runtime power management level with hardware-interface and platform-overhead modeling, with explicit timing impact because transition latency significantly affects scheduler decisions, energy efficiency, and thermal

      management behavior. Its major strength is realism for modern microprocessors and detailed decomposition of DVFS overhead components, making it highly suitable for production-oriented discussions of DVFS limitations. The limitation is its focus on overhead modeling rather than providing a comprehensive survey of all DVFS governors and policies. It’s best used in modern processor DVFS control and runtime scheduling contexts where transition overhead cannot be ignored [31].

   4. Clock Gating Mechanisms and Implementation Strategies

      Clock gating is one of the most widely used and effective techniques for reducing dynamic power consumption in modern integrated circuits. It operates by disabling clock propagation to idle or inactive logic blocks, thereby preventing unnecessary switching activity. Because the clock distribution network often contributes a significant portion of active switching power, effective clock gating can produce substantial energy savings. The technique can be applied at multiple stages of the design flow, including register-transfer level (RTL), synthesis and clock-tree synthesis (CTS), and physical design. However, successful clock-gating implementation requires carefully designed control logic to ensure glitch-free operation, proper synchronization of enable signals, and preservation of timing integrity during clock activation and deactivation. As shown in Fig. 2, a latch-based clock-gating structure is commonly used to stabilize the enable signal before generating the gated clock (GCLK), thereby reducing the risk of glitches. With the advancement in technology to sub-7 nm and below, clock gating is quite a complex activity because timing sensitivity is more sensitive, interconnect parasites are more common, and clock skew variation is more widespread. Gating controls, positioning, and insertion delay placement must be done carefully by designers to prevent timing integrity. Moreover, hierarchical clock gating, in which it is done at multiple architectural levels, is now also typical of complex SoCs. Both the synthesis-time and the CTS-time gating offer local control on the fine-grains and the module-level optimization on the coarse-grains. State-of-the-art design flows are also exploited by power-intent specifications (through UPF) and machine- learning-based prediction to determine the best gating opportunities.

      Clock-gating techniques can be understood as a progression from simple logic-level gating to increasingly sophisticated approaches that incorporate architectural awareness and runtime adaptability. At the most basic level, combinational clock gating is implemented in RTL or during logic synthesis using straightforward combinational control, such as an AND or NAND gate combined with an enable signal, to block clock propagation when a block isn’t actively being used. This method is relatively straightforward to design and can deliver moderate dynamic power savings, but it’s susceptible to glitches and spurious pulses if the enable signal isn’t properly synchronized or latched. For this reason, it’s better suited to smaller logic blocks and carefully controlled low-power IP applications. Sequential clock gating improves on this by registering or latching the enable signal before it controls the clock gate, which enables glitch-free operation at the expense of a small increase in area and control overhead. This latch- or flip-flop-based approach is commonly used in datapath units and control finite state machines, and when properly integrated into synthesis and timing-aware design

      flows, it typically delivers higher savings than purely combinational gating. At the standard-cell level, integrated clock-gating cells represent the industrialized implementation of sequential gating. These are pre-characterized latch-based clock-gating cells from the standard-cell library that are typically inserted or optimized by EDA tools, rather than being simply “provided by the foundry” as is sometimes assumed. In practice, these cells reduce the manual effort required for implementation because they come with timing characterization and library optimization already done, making them the preferred choice in larger system-on-chips and mobile processors. However, ensuring enable correctness, maintaining verification quality, and achieving proper timing integration still demand careful design review. As the scope of clock gating expands, hierarchical clock gating applies enable control across multiple structural levels, such as block, cluster, and subsystem levels, which allows for larger aggregate power savings by suppressing switching activity more broadly throughout the clock tree. The increased savings potential comes with higher design complexity, though, because managing multiple enable domains, handling cross-block dependencies, and dealing with clock-distribution interactions all increase the effort required for timing closure and verification. It’s important to distinguish hierarchical clock gating from dynamic or adaptive clock gating: hierarchical clock gating describes where gating is applied meaning multi-level structural granularitywhile dynamic or

      Figure 2: Conceptual Diagram of a Clock Gating Cell

      adaptive clock gating describes how gating decisions are made meaning runtime or activity-aware control policies. These approaches aren’t mutually exclusive and can be combined in practical low- power system-on-chip implementations. Data-driven clock gating and dynamic or adaptive clock gating take the concept further by basing clock suppression on activity prediction, operand behavior, or runtime workload conditions rather than relying solely on static enable signals. In data-driven approaches, factors like signal transitions, operand correlations, or learned activity patterns can be used to selectively reduce clocking in arithmetic or digital signal processing units, though this typically requires additional monitoring and control logic and may introduce timing sensitivity in the control path. Dynamic or adaptive approaches generalize this to system-level policies where synthesis guidance, runtime analysis, or learned models help determine when and where clocking should be reduced. The benefits of these approaches are highly context-dependent and sensitive to implementation details. As a result, any reported power savings associated with these clock-gating styles should be understood as indicative rather than universal, since actual gains can vary considerably based on workload characteristics, clock-tree topology, gating granularity, enable signal quality, and technology or library choices. Overall, the clock-gating taxonomy presented here is technically sound and can be comprehensively summarized using the existing literature, particularly with a practical ASIC clock-gating methodology as an anchor and supporting references for robustness- aware and learning-driven variants [5], [11], [15], [18], [25], [30].

   5. Power Gating Mechanisms and Retention Strategies

   Power gating is one of the most effective methods for cutting leakage power in idle or low-activity blocks, especially at advanced technology nodes where standby power can make up a substantial portion of total energy use. The technique puts selected logic blocks into a low-leakage sleep state by disconnecting them from the main power rail using high- threshold-voltage sleep transistors. This approach works particularly well in multi-core and heterogeneous system-on- chips where activity depends on workload, allowing selective domain power-down to meaningfully improve energy efficiency. Depending on how fine-grained the control needs to be and what performance constraints exist, power gating can be applied broadly at the subsystem or domain level, or more precisely at the block or cell level. A typical power- gating structure places a header or footer sleep transistor between the logic block and the power rail, which creates a virtual supply or ground rail, as shown in Fig. 3.

   A major challenge with power gating is handling wake-up latency, inrush current, and state retention during power transitions. To maintain functional correctness and protect data integrity, designers use retention registers or flip-flops to store critical state information while the rest of the logic is powered down. These retention elements connect to an always-on supply rail and restore the saved state when the domain powers back up. Reliable power-state transitions also require proper sequencing of the sleep signal, isolation control, and save/restore operations to prevent glitches, floating outputs, and race conditions. In modern digital implementationflows, these interactions are typically modeled using power-intent formats like the Unified Power Format, which allows synthesis, place-and-route, and verification tools to validate low-power behavior throughout the design process.

   Figure 3: Conceptual Diagram of a Power-Gated Functional Block

   Figure 4 illustrates the concept of a retention-based storage strategy with a primary storage element in the switchable domain and a smaller retention latch in an always-on domain. Before the block powers down, the state gets copied from the main storage element into the retention latch through a save

   operation. During the power-off period, the retention latch holds onto this state while the switchable domain stays disabled. Once power comes back, the saved value transfers back to the main storage element through a restore operation, letting normal operation resume with minimal recovery overhead. Together, power-gating and state-retention strategies form a core part of modern low-power system-on- chip design, enabling aggressive leakage reduction while keeping critical context intact in multi-domain systems [16], [27].

   Figure 4: Retention Flip-Flop and State Preservation During Power- Down

4. Conclusion

The transition to the more advanced semiconductor manufacturing technologies has exacerbated the problem of the tradeoff between timing performance and power efficiency, and low-power design has become a major factor in whether SoCs can be brought to fruition. This paper has reviewed power gating and clock gating development and operation as one of the steps that have helped to make timing- aware power optimization in sub-7 nm technologies possible. Clock gating has been applied both in minimizing dynamic power, i.e., minimizing extraneous switching activity, and in power gating, minimizing leakage, i.e., by selectively turning domains off and maintaining states. These techniques will constitute a complementary system of holistic energy management at the design levels. Most present design streams can incorporate these techniques to machine-learning-driven prediction, intent-modeling-driven on UPFs, and a machine- learned timing-sensitive EDA-generation, which can be more flexible and as-you-scale optimized. Nonetheless, it can be seen that making the right trade between timing closure, area overhead, and power savings itself is a design-time trade-off, which is to take into consideration physical effects, e.g., IR drop, process variation, and wake-up latency, in a holistic way. Since the semiconductors are still being further scaled to 3 nm and further beyond will be convergence of AI-based power management, predictive timing modeling, and dense gating designs, which will represent the next generation of low-power design methodology, and will guarantee performance integrity and energy sustainability in the next generation of integrated circuits.

REFERENCES

1. Radamson, H. H., Miao, Y., Zhou, Z., Wu, Z., Kong, Z., Gao, J., … & Wang, G. (2024). CMOS scaling for the 5 nm node and beyond: Device, process and technology. Nanomaterials, 14(10), 837.
2. Jacob, A. P., Xie, R., Sung, M. G., Liebmann, L., Lee, R. T., & Taylor,

   B. (2017). Scaling challenges for advanced CMOS devices. International Journal of High Speed Electronics and Systems, 26(01n02), 1740001.

3. Rabaey, J. M., Chandrakasan, A., & Nikolic, B. (2002). Digital integrated circuits (Vol. 2). Englewood Cliffs: Prentice Hall.
4. Kang, S. M., & Leblebici, Y. (2003). CMOS digital integrated circuits

   (pp. 116-117). New York, NY, USA:: MacGraw-Hill.

5. Keating, M., Flynn, D., Aitken, R., Gibbons, A., & Shi, K. (2007). Low power methodology manual: for system-on-chip design. Boston, MA: Springer US.
6. Kao, J. T., & Chandrakasan, A. P. (2002). Dual-threshold voltage techniques for low-power digital circuits. IEEE Journal of Solid-state circuits, 35(7), 1009-1018.
7. Calimera, A., Macii, A., Macii, E., & Poncino, M. (2014). Power-gating for leakage control and beyond. In Circuit Design for Reliability (pp. 175-205). New York, NY: Springer New York.
8. Vo, M. (2018). Comparative study on power gating techniques for lower power delay product, smaller power loss, faster wakeup time. EAI Endorsed Transactions on Industrial Networks and Intelligent Systems, 5(15).
9. Patil, S., Rawat, A., & Ganguly, U. (2021). An accurate process-induced variability-aware compact model-based circuit performance estimation for design-technology co-optimization. IEEE Transactions on Electron Devices, 69(1), 45-50.
10. Xie, Z., Ren, H., Khailany, B., Sheng, Y., Santosh, S., Hu, J., & Chen,

    Y. (2020, January). PowerNet: Transferable dynamic IR drop estimation via maximum convolutional neural network. In 2020 25th Asia and South Pacific Design Automation Conference (ASP-DAC) (pp. 13-18). IEEE.

11. Huang, G., Hu, J., He, Y., Liu, J., Ma, M., Shen, Z., … & Wang, Y.

    (2021). Machine learning for electronic design automation: A survey. ACM Transactions on Design Automation of Electronic Systems (TODAES), 26(5), 1-46.

12. Zhao, Z., Zhang, S., Liu, G., Feng, C., Yang, T., Han, A., & Wang, L. (2022). Machine-learning-based multi-corner timing prediction for faster timing closure. Electronics, 11(10), 1571.
13. . (2023). Synthesis of Clock Gating Based on Accurate and Learning Driven Power Analyses (Doctoral dissertation, ).
14. Neto, W. L., Li, Y., Gaillardon, P. E., & Yu, C. (2022). FlowTune: End- to-end automatic logic optimization exploration via domain-specific multiarmed bandit. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 42(6), 1912-1925.
15. Won, D., Kim, S., & Kim, T. (2023, August). Machine Learning Driven Synthesis of Clock Gating. In 2023 IEEE/ACM International Symposium on Low Power Electronics and Design (ISLPED) (pp. 1-6). IEEE.
16. C. Yeh, Y. -C. Chen and J. -S. Wang, “Towards Process Variation- Aware Power Gating,” in IEEE Transactions on Very Large Scale

    Integration (VLSI) Systems, vol. 20, no. 11, pp. 1929-1937, Nov. 2012, doi: 10.1109/TVLSI.2011.2169435

17. Kwon, Y., & Shin, Y. (2022, September). Fast prediction of dynamic IR-drop using a recurrent u-net architecture. In Proceedings of the 2022 ACM/IEEE Workshop on Machine Learning for CAD (pp. 71-76).
18. L. Lai, V. Chandra, R. Aitken and P. Gupta, “BTI-Gater: An Aging- Resilient Clock Gating Methodology,” in IEEE Journal on Emerging and Selected Topics in Circuits and Systems, vol. 4, no. 2, pp. 180-189, June 2014, doi: 10.1109/JETCAS.2014.2315882.
19. Narendra, S. G., & Chandrakasan, A. (2006). Leakage in nanometer CMOS technologies. Boston, MA: Springer US.
20. Hu, H. H., Zeng, Y. W., & Chen, K. M. (2018). Improving the gate- induced drain leakage and on-state current of fin-like thin film transistors with a wide drain. Applied Sciences, 8(8), 1406.
21. Sitik, C., Filippini, L., Salman, E., & Taskin, B. (2014, July). High performance low swing clock tree synthesis with custom D flip-flop design. In 2014 IEEE Computer Society Annual Symposium on VLSI (pp. 498-503). IEEE.
22. Avadhani MD, S. (2020). Future of Timing Analysis in VLSI Circuits.

    International Journal of Electrical Engineering and Technology, 11(4).

23. Parihar, S., Pahw, G., Mohammad, B., Chauhan, Y. S., & Amrouch, H. (2024). Novel Trade-offs in 5 nm FinFET SRAM Arrays at Extreme Low Temperatures. IEEE Transactions on Quantum Engineering.
24. A. Chandrakasan and R. Brodersen, Minimizing power consumption in digital CMOS circuits, Proceedings of the IEEE, vol. 83, no. 4, pp. 498523, 1995.
25. M. Rapp et al., “MLCAD: A Survey of Research in Machine Learning for CAD Keynote Paper,” in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, vol. 41, no. 10, pp. 3162- 3181, Oct. 2022.
26. Anantha P. CHANDRAKASAN, Samuel SHENG, Robert W. BRODERSEN, “Low-Power CMOS Digital Design” in IEICE TRANSACTIONS on Electronics, vol. E75-C, no. 4, pp. 371-382, April 1992
27. S. Carver, A. Mathur, L. Sharma, P. Subbarao, S. Urish and Q. Wang, “Low-Power Design Using the Si2 Common Power Format,” in IEEE Design & Test of Computers, vol. 29, no. 2, pp. 62-70, April 2012
28. A. B. Sachid, S. Khandelwal and Chenming Hu, “Body-bias effect in SOI FinFET for low-power applications: Gate length dependence,” Proceedings of Technical Program – 2014 International Symposium on VLSI Technology, Systems and Application (VLSI-TSA),

    Hsinchu, Taiwan, 2014

29. James Kao, Anantha Chandrakasan, and Dimitri Antoniadis. 1997. Transistor sizing issues and tool for multi-threshold CMOS technology. In Proceedings of the 34th annual Design Automation Conference (DAC ’97).
30. Jairam, S., Rao, M., Srinivas, J., Vishwanath, P., Udayakumar, H., & Rao, J. C. (2008, August). Clock gating for power optimization in ASIC design cycle theory & practice. In ISLPED
31. S. Park et al., “Accurate Modeling of the Delay and Energy Overhead of Dynamic Voltage and Frequency Scaling in Modern Microprocessors,” in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, vol. 32, no. 5, pp. 695-708, May 2013