The Next Wave of EDA: Exploring Machine Learning and Open-source Philosophies for Physical Design

Abstract

The slowdown in Moore's law, coupled with the insatiable demand for compute from today's emerging applications, has created the need for hardware systems far beyond current capabilities. Electronic design automation (EDA) tools are now challenged to build chips that not only compensate for slow down in scaling (design-equivalent scaling) but also provide high performance for both ML and non-ML applications, which use a variety of new architectural techniques and operate under stringent performance constraints. However, conventional EDA tools and EDA research are plagued by several challenges. First, they involve computationally expensive analysis and optimizations which increase design time-to-market. Second, they are heuristic-based and often tradeoff speed for accuracy which results in suboptimal solutions. Third, the closed culture within the community, expensive EDA tools, and complexity of EDA tools that demands expert users have created a high barrier to entry slowing down research. This thesis explores machine learning (ML) and open-source philosophies to address these challenges in EDA tools for physical design. Every stage of physical design involves the use of optimization algorithms, which invoke computationally expensive analysis, to minimize power and area and maximize the performance of the circuit. Three critical analyses that are performed several times during physical design are power delivery network (PDN) analysis, thermal analysis, and timing analysis. Optimization algorithms require these analyses to be fast and accurate to generate solutions that can drive design-equivalent scaling. While these analyses and optimization problems have been explored since the early 1980s, the advent of ML brings in novel solutions and an open culture that creates a new wave in EDA. The first part of this thesis shows how ML transcends conventional EDA tool challenges for power-related problems by (i) enabling vast speedups for PDN and thermal analysis with fast inference with high accuracy,(ii) providing optimized correct-by-construction PDN solutions, and (iii) aiding in PDN benchmark and dataset generation. To speed up conventional PDN and thermal analysis, this thesis maps them into image-to-image and sequence-to-sequence translation tasks, which allows leveraging a class of ML models with an encoder-decoder-based generative (EDGe) architecture. Once trained, these ML models are transferable across designs synthesized within the same technology and packing solution. The networks perform millisecond analyses with negligibly small errors against commercial tools that require several hours. Further, these EDGe networks are leveraged to develop MAVIREC, which can rapidly sample a large set of input vectors and provide recommendations for voltage drop analysis. While analysis diagnoses problems, it must be coupled with optimization techniques to solve these problems. The thesis proposes OpeNPDN, an ML-based methodology for PDN optimization that employs a set of predefined templates that serve as potential building blocks for the PDN. The optimization algorithm involves building a PDN that meets voltage drop and electromigration constraints while minimizing the wiring resources used. A convolutional neural network (CNN) is trained to select an appropriate PDN template for each region on the chip based on power, bump, macro, and congestion distributions. The CNN-based optimization rapidly frees thousands of routing tracks in congestion-critical regions, while staying within the voltage drop and electromigration (EM) current density limits. Although ML has found success in the problems described above, a major challenge has been the unavailability of benchmarks for evaluating these solutions. The thesis develops BeGAN, an ML-based methodology for synthesizing synthetic PDN benchmarks that obfuscate intellectual property information. The approach applies generative adversarial networks (GANs) and transfer learning techniques to create realistic PDN benchmarks from a small set of real circuit data. BeGAN generates thousands of PDN benchmarks released in the public domain in four open-source technologies. The second part of the thesis develops ML-based solutions for timing analysis and optimization. A critical part of timing closure is accurate crosstalk-aware timing analysis. However, existing crosstalk-aware static timing analysis (STA) techniques are limited due to their imprecise use of heuristics. The thesis proposes XT-PRAGGMA, a tool that uses GPU-accelerated dynamic gate-level simulations and ML to eliminate inaccuracies in crosstalk-aware STA tools and accurately predict crosstalk-induced delays. The proposed technique is fast and reduces falsely-reported total negative slack by 70% compared to traditional STA-based techniques. In addition to accurate analysis techniques, timing optimization algorithms are critical to successful timing closure. Logic gate sizing is one such NP-hard-constrained optimization problem that aims to minimize the circuit implementation cost (typically, area or power), subject to timing constraints. This thesis presents RL-LR-Sizer, a tool that applies Lagrangian relaxation techniques to a novel reinforcement learning (RL) framework to perform autonomous gate sizing. The RL framework trains a relational graph convolutional network agent to perform timing optimization. RL-LR-Sizer outperforms conventional algorithms and moves the Pareto optimal front of the area-delay tradeoff curve to the left on designs in a 45nm technology. The last part of the thesis aims to address challenges related to high barriers to entry and the closed culture in the EDA community, which hinder reproducible research and the adoption of ML-based techniques. It contributes to the open ecosystem by developing novel open-source EDA tools. The first is TherMOS, an open-source thermal model to estimate self heating effects in advanced transistors. Modern transistors such as FinFETs and gate-all-around FETs suffer from excessive heat confinement due to their small size and three-dimensional geometries. This results in device self-heating, which can reduce speed, increase leakage, and accelerate aging. The thesis uses TherMOS to characterize the temperature for both the 7nm FinFET and 5nm GAAFET sub-structures and analyzes the impact of temperature on circuit performance and reliability. The second is RTA-simulator, an open-source pattern-dependent transient rapid thermal annealing (RTA) simulator for analyzing RTA-induced variations due to differences in die layout patterns. Unlike prior art, RTA-simulator considers the temperature-dependent thermal conductivity of silicon. The thesis leverages RTA-simulator to analyze RTA effects on a 16KB SRAM and suggest strategies for RTA-induced variation mitigation for a 7nm FinFET SRAM design. The third is PDNSim, an open-source PDN analysis tool integrated with OpenROAD, for estimating voltage drop and EM current densities. With scaling, power densities and wire parasitics are on an increasing trend making PDN analysis an integral part of physical design. For the OpenROAD project, PDNSim serves two purposes. First, it performs static voltage drop estimation and current density estimation at various stages of the physical design flow after placement. Second, it checks the connectivity of the power grid to ensure that every standard cell receives power (voltage) from the pin (bump) and that all power stripes are connected. The thesis demonstrates the results of PDNSim on several designs in both open-source and closed-source technologies, which are being run as a part of the OpenROAD project regression tests.