Modern physical design flows highly depend on design space exploration to find the commercial tools’ clock tree synthesis (CTS) parameters that lead to optimized clock trees. However, such exploration is often time-consuming and computationally inefficient. In this article, we overcome this drawback by proposing a novel framework named GAN-CTS, which utilizes conditional generative adversarial network (GAN) to predict and optimize CTS outcomes. Our framework is built upon three sequential learning stages. First, to precisely characterize distinct designs, we leverage transfer learning to extract netlist features directly from placement images. Second, we perform regression learning using various methods to predict the target CTS outcomes and demonstrate that the proposed multitask learning approach achieves better accuracy than the meta-modeling method adopted by previous works. To fully benefit from the predictions made by our framework, we further quantitatively interpret the importance of each CTS input parameter subject to various design objectives through attribution-based learning. Finally, generative adversarial learning is leveraged to optimize the target clock metrics with the guidance provided by the pretrained regression model. To substantiate the generality of our framework, we perform validations on four unseen netlists that are not utilized in the training process. The experimental results conducted on real-world designs demonstrate that our framework: 1) achieves an average prediction error of 3%; 2) improves the commercial tool’s auto-generated clock tree by 20.7% in clock power, 21.5% in clock wirelength, 36.1% in the worst skew; and 3) reaches an F1-score of 0.93 in the classification task of determining successful and failed CTS runs.

Tier partitioning is one of the most critical stages in monolithic 3-D (M3D) integrated circuits (ICs) implementation flows. It transforms 2-D netlists into 3-D by performing tier assignment for each design instance, which directly impacts the power, performance, and area (PPA) metrics of final 3-D full-chip designs. However, the current state-of-the-art tier partitioning approach named bin-based min-cut algorithm has fundamental flaws that lead to severe drawbacks, such as timing degradation, 3-D routing overhead, and redundant monolithic intertier vias (MIVs) insertion. To overcome these issues, in this article, we propose TP-GNN, an unsupervised graph learning-based tier partitioning framework that utilizes graph neural networks (GNNs) and advanced machine learning (ML) techniques to perform tier partitioning. The proposed framework comprehends design- and technology-related parameters properly so that it is generalizable to various netlists and technologies. In addition, it can be integrated with any style of M3D design flows that require tier assignments of standard cells. In the experiments, we validate the proposed framework on seven industrial designs with two different fashions of M3D implementation flows: 1) partitioning-first (Snap3D) and 2) partitioning-last (Shrunk2D and Compact2D) flows. We demonstrate that our framework, TP-GNN, significantly improves the 3-D quality of results (QoR) across most testing designs in a large margin compared with the bin-based min-cut tier partitioning algorithm. Specifically, in OpenPiton, an RISC-V-based multicore system, we observe 27.4%, 7.7%, and 20.3% improvements in performance, wirelength, and energy-per-cycle, respectively. Finally, we perform a case study by applying the proposed framework to a heterogeneous M3D design flow, Pin3D, on a commercial CPU design and observe that TP-GNN reaches better partitioning solutions than the existing partitioning approaches for heterogeneous 3-D ICs.

3D integration technology is one of the few options that can keep Moore's Law trajectory beyond conventional scaling. Existing 3D physical design flows fail to benefit from the full advantage that 3D integration provides. Particularly, current 3D partitioning algorithms do not comprehend technology and design-related parameters properly, which results in sub-optimal partitioning solutions. In this paper, we propose TP-GNN, an unsupervised graph-learning-based tier partitioning framework, to overcome this issue. Experimental results on 7 industrial designs demonstrate that our framework significantly improves the QoR of the state-of-the-art 3D implementation flows. Specifically, in OpenPiton, a RISC-V-based multi-core system, we observe 27.4%, 7.7% and 20.3% improvements in performance, wirelength, and energy-per-cycle respectively.

In advanced technology nodes, emerging die-to-wafer (D2W) integration technology is a promising "More Than Moore" lever for continued scaling of system capability and value. In D2W 3D IC implementation, the power delivery network (PDN) is crucial to meeting design specifications. However, determining the optimal PDN design is nontrivial. On the one hand, to meet the IR drop requirement, denser power mesh is desired. On the other hand, to meet the timing requirement for a high-utilization design, more routing resource should be available for signal routing. Moreover, additional competition between signal routing and power routing is caused by inter-tier vertical interconnects in 3D IC. In this paper, we propose a power delivery pathfinding methodology for emerging die-to-wafer integration, which seeks to identify an optimal or near-optimal PDN for a given design and PDN specification. Our pathfinding methodology exploits models for routability and worst IR drop, which helps reduce iterations between PDN design and circuit design in 3D IC implementation. We present validations with real design examples and a 28nm foundry technology.

In this paper, we propose a complete framework named GAN-CTS which utilizes conditional generative adversarial network (GAN) and reinforcement learning to predict and optimize clock tree synthesis (CTS) outcomes. To precisely characterize different netlists, we leverage transfer learning to extract design features directly from placement images. Based on the proposed framework, we further quantitatively interpret the importance of each CTS input parameter subject to various design objectives. Finally, to prove the generality of our framework, we conduct experiments on the unseen netlists which are not utilized in the training process. Experimental results performed on industrial designs demonstrate that our framework (1) achieves an average prediction error of 3%, (2) improves the commercial tool's auto-generated clock tree by 51.5% in clock power, 18.5% in clock wirelength, 5.3% in the maximum skew, and (3) reaches an F1-score of 0.952 in the classification task of determining successful and failed CTS processes.

Generative Adversarial Networks (GANs) are a frontier in deep learning. GANs consist of two models: generative and discriminative. While the discriminative model uses the conventional convolution, the generative model depends on a fundamentally different operator, called transposed convolution. This operator initially inserts a large number of zeros in its input and then slides a window over this expanded input. This zero-insertion step leads to a large number of ineffectual operations and creates distinct patterns of computation across the sliding windows. The ineffectual operations along with the variation in computation patterns lead to significant resource underutilization when using conventional convolution hardware. To alleviate these sources of inefficiency, this paper devises FlexiGAN, an end-to-end solution, that generates an optimized synthesizable FPGA accelerator from a high-level GAN specification. FlexiGAN is coupled with a novel template architecture that aims to harness the benefits of both MIMD and SIMD execution models to avoid ineffectual operations. To this end, the proposed architecture separates data retrieval and data processing units at the finest granularity of each compute engine. Leveraging this separation enables the architecture to use a succinct set of operations to cope with the irregularities of transposed convolution. At the same time, it significantly reduces the on-chip memory usage, which is generally limited in FPGAs. We evaluate our end-to-end solution by generating FPGA accelerators for a variety of GANs. These generated accelerators provide 2.4× higher performance than an optimized conventional convolution design. In addition, FlexiGAN, on average, yields 2.8× (up to 3.7×) improvements in Performance-per-Watt over a Titan X GPU.

Deep Convolutional Neural Networks (CNNs) perform billions of operations for classifying a single input. To reduce these computations, this paper offers a solution that leverages a combination of runtime information and the algorithmic structure of CNNs. Specifically, in numerous modern CNNs, the outputs of compute-heavy convolution operations are fed to activation units that output zero if their input is negative. By exploiting this unique algorithmic property, we propose a predictive early activation technique, dubbed SnaPEA. This technique cuts the computation of convolution operations short if it determines that the output will be negative. SnaPEA can operate in two distinct modes, exact and predictive. In the exact mode, with no loss in classification accuracy, SnaPEA statically re-orders the weights based on their signs and periodically performs a single-bit sign check on the partial sum. Once the partial sum drops below zero, the rest of computations can simply be ignored, since the output value will be zero in any case. In the predictive mode, which trades the classification accuracy for larger savings, SnaPEA speculatively cuts the computation short even earlier than the exact mode. To control the accuracy, we develop a multi-variable optimization algorithm that thresholds the degree of speculation. As such, the proposed algorithm exposes a knob to gracefully navigate the trade-offs between the classification accuracy and computation reduction. Compared to a state-of-the-art CNN accelerator, SnaPEA in the exact mode, yields, on average, 28% speedup and 16% energy reduction in various modern CNNs without affecting their classification accuracy. With 3% loss in classification accuracy, on average, 67.8% of the convolutional layers can operate in the predictive mode. The average speedup and energy saving of these layers are 2.02x and 1.89x, respectively. The benefits grow to a maximum of 3.59x speedup and 3.14x energy reduction. Compared to static pruning approaches, which are complimentary to the dynamic approach of SnaPEA, our proposed technique offers up to 63% speedup and 49% energy reduction across the convolution layers with no loss in classification accuracy.

Generative Adversarial Networks (GANs) are one of the most recent deep learning models that generate synthetic data from limited genuine datasets. GANs are on the frontier as further extension of deep learning into many domains (e.g., medicine, robotics, content synthesis) requires massive sets of labeled data that is generally either unavailable or prohibitively costly to collect. Although GANs are gaining prominence in various fields, there are no accelerators for these new models. In fact, GANs leverage a new operator, called transposed convolution, that exposes unique challenges for hardware acceleration. This operator first inserts zeros within the multidimensional input, then convolves a kernel over this expanded array to add information to the embedded zeros. Even though there is a convolution stage in this operator, the inserted zeros lead to underutilization of the compute resources when a conventional convolution accelerator is employed. We propose the GANAX architecture to alleviate the sources of inefficiency associated with the acceleration of GANs using conventional convolution accelerators, making the first GAN accelerator design possible. We propose a reorganization of the output computations to allocate compute rows with similar patterns of zeros to adjacent processing engines, which also avoids inconsequential multiply-adds on the zeros. This compulsory adjacency reclaims data reuse across these neighboring processing engines, which had otherwise diminished due to the inserted zeros. The reordering breaks the full SIMD execution model, which is prominent in convolution accelerators. Therefore, we propose a unified MIMD-SIMD design for GANAX that leverages repeated patterns in the computation to create distinct microprograms that execute concurrently in SIMD mode. The interleaving of MIMD and SIMD modes is performed at the granularity of single microprogrammed operation. To amortize the cost of MIMD execution, we propose a decoupling of data access from data processing in GANAX. This decoupling leads to a new design that breaks each processing engine to an access micro-engine and an execute micro-engine. The proposed architecture extends the concept of access-execute architectures to the finest granularity of computation for each individual operand. Evaluations with six GAN models shows, on average, 3.6x speedup and 3.1x energy savings over Eyeriss without compromising the efficiency of conventional convolution accelerators. These benefits come with a mere ≈7.8% area increase. These results suggest that GANAX is an effective initial step that paves the way for accelerating the next generation of deep neural models.

Monolithic 3-D (M3D) integrated circuits (ICs) are an emerging technology that offer much higher integration densities than previous 3-D IC approaches. In this paper, we present a complete netlist-to-layout design flow to design an M3D block, as well as to integrate 2-D and 3-D blocks into an M3D SoC. This design flow is based on commercial tools built for 2-D ICs, and enhanced with our 3-D specific methodologies. We use the OpenSPARC T2 SoC as a case study, implement it in a 28-nm fully depleted silicon on insulator foundry process, and demonstrate that we can achieve up to 12% and 8% power savings for a single block and SoC, respectively, when compared with their 2-D counterparts implemented using commercial tools.

Monolithic 3-D ICs (M3-D ICs) offer extremely high vertical interconnection density, significantly improving the power-performance envelope when compared to conventional 2-D ICs. However, process limitations lead to one tier having either degraded transistors or interconnects. This paper models the amount of degradation that can be expected at current and future nodes (45, 22, and 10 nm), develops a process development kit using these models to enable evaluation, and presents a block-level M3-D IC RTL-to-GDSII flow that is capable of mitigating some of this degradation. Experimental results indicate that at lower technology nodes, M3-D ICs offer more benefits. Results also indicate that the impact of transistor degradation is diminished at lower technology nodes while the impact of interconnect degradation becomes worse. Overall, perfect M3-D ICs close more than half the gap in the power-performance envelope between 2-D ICs and the “ideal” block-level design. While degraded tiers reduce the benefit of M3-D ICs, our degradation-aware floorplanner gives back up to 17% of the loss, and helps to obtain significant overall benefits compared to 2-D ICs.

In this paper, we present a comprehensive study on the impact of power delivery network (PDN) on full-chip wirelength, routability, power, and thermal effects in gate-level monolithic 3-D (M3-D) ICs across different technology nodes. Our studies show that PDN worsens routing congestion more severely in M3-D ICs than in 2-D designs due to the significant reduction in resources for 3-D connections. The relative impact worsens at advanced technology nodes due to higher congestion of interconnects. The increase in signal wirelength translates into additional net switching power dissipation, which significantly contributes to total power. This in turn aggravates thermal issues further in 3-D ICs. In addition, we observe that PDN tradeoffs among wirelength, power, and thermal are more pronounced in M3-D ICs than through silicon via-based 3-D and 2-D designs because of the higher integration density and the severe competition between signal and power connections. We also compare the impact of PDN on full chip routing in M3-D ICs versus face-to-face 3-D ICs. Lastly, we use various PDN design optimization techniques for M3-D ICs at different nodes and obtain up to 13.9% signal wirelength and 17.6% total power reduction under the given IR drop budget.

Heterogeneous MPSoCs typically integrate diverse cores, including application CPUs, GPUs, and HD coders. These cores commonly share an off-chip memory to save cost and energy, but their memory accesses often interfere with each other, leading to undesirable consequences like a slowdown of application performance or a failure to sustain real-time performance. The memory controller plays a central role in meeting the QoS needs of real-time cores while maximizing CPU performance. Previous QoS-aware memory controllers are based on a classic two-tier queuing architecture that buffers memory transactions at the first tier, followed by a second tier that buffers translated DRAM commands. In these designs, QoS-aware policies are used to schedule competing transactions at the first stage, but the translated DRAM commands are served in FIFO order at the second stage. Unfortunately, once the scheduled transactions have been forwarded to the command stage, newly arriving transactions that may be more critical cannot be served ahead of those translated commands that are already queued at the second stage. To address this, we propose a scalable memory controller architecture based on single-tier virtual queuing (STVQ) that maintains a single tier of request queues and employs an efficacious scheduler that considers both QoS requirements and DRAM bank states. In comparison with previous QoS-aware memory controllers, the proposed STVQ memory controller reduces CPU slowdown by up to 13.9% while satisfying all frame rate requirements. We propose further optimizations that can significantly increase row-buffer hits by up to 66.2% and reduce memory latency by up to 19.8%.

In heterogeneous MPSoCs, memory interference between the CPU and realtime cores is a critical impediment to system performance. Previous memory schedulers adopt the classic two-tier queuing system, but unfortunately the use of two-tier queuing deteriorates the QoS of scheduling policies. In this paper, we propose the Single-Tier Virtual Queuing (STVQ) memory controller for efficacious QoS-aware scheduling. The STVQ memory controller maintains single-tier transaction queues and employs separable allocation for transaction scheduling with high scalability. A multi-source realtime scheduling algorithm is further presented. The STVQ controller achieves up to 13.9% less CPU IPC slowdown than previous schedulers with no frame rate penalty on realtime cores.

In advanced technology nodes, physical design engineers must estimate whether a standard-cell placement is routable (before invoking the router) in order to maintain acceptable design turnaround time. Modern SoC designs consume multiple compute servers, memory, tool licenses and other resources for several days to complete routing. When the design is unroutable, resources are wasted, which increases the design cost. In this work, we develop machine learning-based models that predict whether a placement solution is routable without conducting trial or early global routing. We also use our models to accurately predict iso-performance Pareto frontiers of utilization, aspect ratio and number of layers in the back-end-of-line (BEOL) stack. Furthermore, using data mining and machine learning techniques, we develop new methodologies to generate training examples given very few placements. We conduct validation experiments in three foundry technologies (28nm FDSOI, 28nm LP and 45nm GS), and demonstrate accuracy ≥ 85.9% in predicting routability of a placement. Our predictions of Pareto frontiers in the three technologies are pessimistic by at most 2% with respect to the maximum achievable utilization for a given design in a given BEOL stack.

In this paper, we first present a comprehensive study of the unique thermal behavior in monolithic 3-D integrated circuits (ICs) in contrast to through silicon via-based 3-D ICs. In particular, we study the impact of the thin interlayer dielectric between the device tiers on vertical thermal coupling. We then study and compare the impact of different application-based package structures on the thermal behavior of monolithic 3-D ICs. With these unique properties and behavior, we develop a fast and accurate compact full-chip thermal analysis model based on nonlinear regression technique which adapts to the package structure during development and hence considers it during temperature evaluation. Our model is extremely fast and highly accurate with an error of less than 5%. This model is incorporated into a thermal-aware 3-D-floorplanner that runs without significant runtime overhead. We use the floorplanner with our package-aware thermal model and observe up to 22% reduction in the maximum temperature with insignificant area and performance overhead.

Quantification of three-dimensional integrated circuit (3DIC) benefits over corresponding 2DIC implementation for arbitrary designs remains a critical open problem, largely due to nonexistence of any "golden" 3DIC flow. Actual design and implementation parameters and constraints affect 2DIC and 3DIC final metrics (power, slack, etc.) in highly non-monotonic ways that are difficult for engineers to comprehend and predict. We propose a novel machine learning-based methodology to estimate 3DIC power benefit (i.e., percentage power reduction) based on corresponding golden 2DIC implementation parameters. The resulting 3D Power Estimation (3DPE) models achieve small prediction errors that are bounded by construction. We are the first to perform a novel stress test of our predictive models across a wide range of implementation and design-space parameters. Further, we explore model-guided implementation of designs in 3D to achieve minimum power: that is, our models recommend a most-promising set of implementation parameters and constraints, and also provide a priori estimates of 3D power benefits, based on a given design's post-synthesis and 2D implementation parameters. We achieve ≤10% error in power benefit prediction across various 3DIC designs.

Power delivery to the tier farthest away from the package in 3D VLSI is challenging. This is because the current provided by the package on the bottom is (1) first used by other tiers before it reaches the top, and (2) delivered using extremely small-size intra and inter-tier vias. Our solution is a tier partitioning method that assigns power hungry cells to the tier closer to the package, which is farther away from the heat spreader. Our study shows that this approach alleviates the IR-drop, power delivery network (PDN) resource usage, and power consumption in the top tier. Moreover, moving the cells to the bottom tier, unlike popular belief, does not cause any serious thermal issues. This is especially true in mobile applications, where heat is dissipated by both the heat spreader and printed circuit board. In summary, our tier-partitioning leads to 24.66% IR-drop reduction, 28.57% PDN resource reduction, and 4% wirelength reduction, with < 1°C increase in temperature.

As the semiconductor industry faces serious challenges extending the CMOS roadmap, traditional cost reduction benefits that accompanied power/performance/area (PPA) advantages of successive technology nodes have decreased due to a myriad of process integration challenges and increased variability, reliability, power and thermal constraints. 3D integration technologies have been pursued as a potential solution to help integrate more functions within a confined available dimensions of advanced mobile devices. 3D VLSI (3DV) is an emerging 3D integration technology that unlike packaging-driven 3D technologies (e.g., 2.5D, TSV-based 3D, etc.) can deliver orders of magnitude more integration densities due to extremely small sizes of vertical vias. In this paper, we describe the 3DV technology and its current benefits and challenges. We also survey recent literature that show the potential of 3DV to help continue Moore's law trajectory beyond 2D.

Monolithic 3D (M3D) is an emerging technology that enables integration density which is orders of magnitude higher than that offered by through-silicon-vias. In this paper, we demonstrate that a modified 2D placement technique coupled with a post-placement partitioning step is sufficient to produce high-quality M3D placement solutions. We also present a commercial router-based monolithic intertier via insertion methodology that improves the routability of M3D ICs. We demonstrate that, unlike in 2D ICs, the routing supply and demand in M3D ICs are not completely independent of each other. We develop a routing demand model for M3D ICs, and use it to develop an O(N) min-overflow partitioner that enhances routability by off-loading demand from one tier to another. This technique reduces the routed wirelength and the power delay product by up to 7.44% and 4.31%, respectively. This allows a two-tier M3D IC to achieve, on average, 19.9% and 11.8% improvement in routed wirelength and power delay product over 2D, even with reduced metal layer usage.

In this paper we study the power vs. performance tradeoff in block-level monolithic 3D IC designs. Our study shows that we can close the power-performance gap between 2D and a theoretical lower bound by up to 50%. We model the inter-tier performance variations caused by a low temperature manufacturing process on the non-bottom tiers. We also model an alternate manufacturing process, where highly resistive tungsten interconnects are used on the bottom tier to withstand a high temperature process on the non-bottom tiers. We propose a variation-aware floorplanning technique that makes our design more tolerant to these variations. We demonstrate that our design methods can help us obtain high quality designs even under inter-tier performance variations.

In this paper, we present a comprehensive study of the unique thermal behavior in monolithic 3D ICs. In particular, we study the impact of the thin inter-layer dielectric (ILD) between the device tiers on vertical thermal coupling. In addition, we develop a fast and accurate compact full-chip thermal analysis model based on non-linear regression technique. Our model is extremely fast and highly accurate with an error of less than 5%. This model is incorporated into a thermal-aware 3D-floorplanner that runs without significant runtime overhead. We observe up to 22% reduction in the maximum temperature with insignificant area and performance overhead.

In this paper, we present a comprehensive study on the impact of power delivery network (PDN) on full-chip wirelength, routability, power, and thermal effects in monolithic 3D ICs. Our studies first show that the full PDN worsens routing congestion more severely in monolithic 3D ICs than in 2D designs due to the significant reduction in resources for 3D connections. The increase in signal wirelength translates into additional net switching power dissipation, which significantly contributes to total power. This in turn aggravates thermal issues in 3D ICs. In addition, we observe that PDN tradeoffs among wirelength, power, and thermal are more pronounced in monolithic 3D ICs than TSV-based 3D and 2D designs. This is because of the higher integration density and the severe competition between signal and power connections. Lastly, we develop various PDN design optimization techniques for monolithic 3D ICs and obtain up to 8% signal wirelength and 5% maximum temperature reduction under the given IR drop budget.

In a gate-level monolithic 3D IC (M3D), all the transistors in a single logic gate occupy the same tier, and gates in different tiers are connected using nano-scale monolithic inter-tier vias. This design style has the benefit of the superior power-performance quality offered by flat implementations (unlike block-level M3D), and zero total silicon area overhead compared to 2D (unlike transistor-level M3D). In this paper we develop, for the first time, a complete RTL-to-GDSII design flow for gate-level M3D. Our tool flow is based on commercial tools built for 2D ICs and enhanced with our 3D specific methodologies. We use this flow along with a 28nm PDK to build layouts for the OpenSPARC T2 core. Our simulations show that at the same performance, gate-level M3D offers 16% total power reduction with 0% area overhead compared to commercial quality 2D IC designs.

Monolithic 3D is an emerging technology that enables integration density which is orders of magnitude higher than that offered by through-silicon-vias (TSV). In this paper we demonstrate that a modified 2D placement technique, coupled with a post-placement partitioning step, is sufficient to produce high quality monolithic 3D placement solutions. We also present a commercial router based monolithic inter-tier via (MIV) insertion methodology that dramatically improves the routability of monolithic 3D-ICs. We develop a routing demand model for monolithic 3D-ICs, and use it to develop an O(N) min-overflow partitioner that enhances routability by off-loading demand from one tier to another. This technique reduces the routed wirelength and the power delay product (PDP) by up to 4% and 4.33% respectively, under the same half-perimeter wirelength. This allows a two-tier monolithic 3D-IC to achieve, on average, 19.2% and 12.1% improvement in routed wirelength and PDP over 2D, even with reduced metal layer usage.

Three dimensional integrated circuits (3D-ICs) have emerged as a promising solution to continue device scaling. They can be realized using Through Silicon Vias (TSVs), or monolithic integration using Monolithic Inter-tier vias (MIVs), an emerging alternative that provides much higher via densities. In this paper, we provide a framework for floorplanning existing 2D IP blocks into 3D-ICs using MIVs. We take the floorplanning solution all the way through place-and-route and report post-layout metrics for area, wirelength, timing, and power consumption. Results show that the wirelength of TSV-based 3D designs outperform 2D designs by upto 14% in large-scale circuits only. MIV-based 3D designs, however, offer an average wirelength improvement of 33% for a wide range of benchmark circuits. We also show that while TSV-based 3D cannot improve the performance and power unless the TSV capacitance is reduced, MIV-based 3D offers significant reduction of upto 33% in the longest path delay and 35% in the inter-block net power.

As industry moves towards multicore chips, networks-on-chip (NoCs) are emerging as the scalable fabric for interconnecting the cores. With power now the first-order design constraint, early-stage estimation of NoC power has become crucially important. In this work, we present ORION 2.0, an enhanced NoC power and area simulator, which offers significant accuracy improvement relative to its predecessor, ORION 1.0.

Over the course of this decade, uniprocessor chips have given way to multi-core chips which have become the primary building blocks of today's computer systems. The presence of multiple cores on a chip shifts the focus from computation to communication as a key bottleneck to achieving performance improvements. As industry moves towards many-core chips, networks-on-chip (NoCs) are emerging as the scalable fabric for interconnecting the cores. With power now the first-order design constraint, early-stage estimation of NoC power has become crucially important. Existing power models (e.g., ORION 2.0 [12], Xpipes [7], etc.) are based on certain router microarchitecture and circuit implementation. Therefore, when validated against different NoC prototypes - different router implementations - we saw significant deviation (up to 40% on average) that can lead to erroneous NoC design choices. This has prompted our development of a new, accurate architecture- and circuit implementation-independent router power and area modeling methodology with complete portability across existing NoC component libraries. Also, validation against a range of implemented router designs confirms substantial improvement in accuracy over existing models.

Networks-on-chip (NoCs) are becoming increasingly important in general-purpose and application-specific multi-core designs. Although uniform router configurations are appropriate for generalpurpose NoCs, router configurations for application-specific NoCs can be non-uniformly optimized to application-specific traffic characteristics. In this paper, we specifically consider the problem of virtual channel (VC) allocation in application-specific NoCs. Prior solutions to this problem have been average-rate driven. However, average-rate models are poor representations of real application traffic, and can lead to designs that are poorly matched to the application. We propose an alternate trace-driven paradigm in which configuration of NoCs is driven by application traces. We propose two simple greedy trace-driven VC allocation schemes. Compared to uniform allocation, we observe up to 51 % reduction in the number of VCs under a given average packet latency constraint, or up to 74 % reduction in average packet latency with same number of VCs. Our results suggest that average-rate driven methods cannot effectively select appropriate links for VC allocation because they fail to consider the impact of traffic bursts. As a case study, we compare our proposed approach with an existing average-rate driven method [9] and observe up to 35 % reduction in the number of VCs for a given target latency. 1.

As industry moves towards many-core chips, networks-on-chip (NoCs) are emerging as a scalable communication fabric for interconnecting the cores. With increasing core counts, there is a corresponding increase in communication demands in multi-core designs to facilitate high core utilization, and a consequent critical need for high-performance NoCs. Another megatrend in advanced technologies is that power has become the most critical design constraint. In this paper, we focus on trace-driven virtual channel (VC) allocation in application-specific NoCs. We propose a new significant VC failure metric to capture the impact of VCs on network performance and efficiently drive NoC optimization. Our proposed metaheuristics achieve up to 38% reduction in the number of VCs under a given average packet latency constraint. In addition, compared to a recently proposed trace-driven VC allocation approach, we obtain up to an O(|L|) speedup, where |L| is total number of links in the network, with no degradation in the quality of results.

We propose new accurate predictive models for the delay, power, and area of buffered interconnects to enable a more effective system-level design exploration with existing and future nanometer technology processes. We show that our models are significantly more accurate than previous models - essentially matching sign-off analyses. We integrate our models in the COSI-OCC communication synthesis infrastructure and show how they impact the feasibility and optimality of the network-on-chip architectures that are synthesized by this tool.

As industry moves towards many-core chips, networks-on-chip (NoCs) are emerging as the scalable fabric for interconnecting the cores. With power now the first-order design constraint, early-stage estimation of NoC power has become crucially important. ORION was amongst the first NoC power models released, and has since been fairly widely used for early-stage power estimation of NoCs. However, when validated against recent NoC prototypes - the Intel 80-core Teraflops chip and the Intel scalable communications core (SCC) chip - we saw significant deviation that can lead to erroneous NoC design choices. This prompted our development of ORION 2.0, an extensive enhancement of the original ORION models which includes completely new subcomponent power models, area models, as well as improved and updated technology models. Validation against the two Intel chips confirms a substantial improvement in accuracy over the original ORION. A case study with these power models plugged within the COSI-OCC NoC design space exploration tool confirms the need for, and value of, accurate early-stage NoC power estimation. To ensure the longevity of ORION 2.0, we will be releasing it wrapped within a semi-automated flow that automatically updates its models as new technology files become available.

Accurate modeling of delay, power, and area of interconnections early in the design phase is crucial for effective system-level optimization. Models presently used in system-level optimizations, such as network-on-chip (NoC) synthesis, are inaccurate in the presence of deep-submicron effects. In this paper, we propose new, highly accurate models for delay and power in buffered interconnects; these models are usable by system-level designers for existing and future technologies. We present a general and transferable methodology to construct our models from a wide variety of reliable sources (Liberty, LEF/ITF, ITRS, PTM, etc.). The modeling infrastructure, and a number of characterized technologies, are available as open source. Our models comprehend key interconnect circuit and layout design styles, and a power-efficient buffering technique that overcomes unrealities of previous delay-driven buffering techniques. We show that our models are significantly more accurate than previous models for global and intermediate buffered interconnects in 90 nm and 65 nm foundry processes - essentially matching signoff analyses. We also integrate our models in the COSI-OCC synthesis tool and show that the more accurate modeling significantly affects optimal/achievable architectures that are synthesized by the tool. The increased accuracy provided by our models enables system-level designers to obtain better assessments of the achievable performance/power/area tradeoffs for (communication-centric aspects of) system design, with negligible setup and overhead burdens.

We survey recent research and practice in the area of chemical-mechanical polishing (CMP) fill synthesis, in terms of both problem formulations and solution approaches. We review the CMP as the planarization technique of choice for multilevel very large-scale integration metallization processes. Post-CMP wafer topography varies according to pattern density. We review density-analysis methods and density-control objectives that are used in today's fill-synthesis algorithms. In addition, we discuss the concept of design-driven fill synthesis that seeks to optimize CMP fill with respect to objectives beyond mere density uniformity. Design-driven fill synthesis minimizes the impact of CMP fill on design performance and parametric yield while still satisfying the density criteria that are motivated by manufacturing models. We conclude with a discussion of where CMP fill synthesis may be naturally integrated within future design and manufacturing flows.

Depth of focus is the major contributor to lithographic process margin. One of the major causes of focus variation is imperfect planarization of fabrication layers. Presently, optical proximity correction (OPC) methods are oblivious to the predictable nature of focus variation arising from wafer topography. As a result, designers suffer from manufacturing yield loss as well as loss of design quality through unnecessary guardbanding. In this paper, the authors propose a novel flow and method to drive OPC with a topography map of the layout that is generated by chemical-mechanical polishing simulation. The wafer topography variations result in local defocus, which the authors explicitly model in the OPC insertion and verification flows. In addition, a novel topography-aware optical rule check to validate the quality of result of OPC for a given topography is presented. The experimental validation in this paper uses simulation-based experiments with 90-nm foundry libraries and industry-strength OPC and scattering bar recipes. It is found that the proposed topography-aware OPC (TOPC) can yield up to 67% reduction in edge placement errors. TOPC achieves up to 72% reduction in worst case printability with little increase in data volume and OPC runtime. The electrical impact of the proposed TOPC method is investigated. The results show that TOPC can significantly reduce timing uncertainty in addition to process variation