Stack Analysis: Synopsys — Architecting the Future of Silicon with AI-Driven Design

Company Overview

Synopsys is the undisputed leader in Electronic Design Automation (EDA), providing software, IP, and services that power the design and manufacture of advanced semiconductors. As AI accelerates, Synopsys is incorporating machine learning into all aspects of chip design, from architecture exploration to layout optimization and verification, significantly impacting the pace and complexity of future chip development. Their AI-driven solutions are becoming indispensable for companies pushing the boundaries of Moore's Law and beyond.

Core AI/ML Stack

Synopsys has made a significant investment in its AI/ML infrastructure, built around a hybrid approach leveraging both industry-standard frameworks and internally developed tools optimized for EDA workloads. Key components include:

• Model Frameworks: Synopsys utilizes a blend of PyTorch 2.3 for rapid prototyping and experimentation, and JAX 0.4 for production-level training and inference due to its performance and automatic differentiation capabilities. They are also exploring custom neural architecture search (NAS) frameworks based on distributed reinforcement learning to optimize model structures for specific EDA tasks.
• Model Types: Graph Neural Networks (GNNs) are heavily used for analyzing circuit topologies and predicting performance metrics. Generative Adversarial Networks (GANs) are employed for design space exploration and generating novel chip architectures. Transformer models are used for code generation and bug detection in RTL.
• Training Infrastructure: Synopsys operates a large-scale, on-premise GPU cluster consisting of NVIDIA H200 Tensor Core GPUs interconnected with a high-bandwidth InfiniBand network. They are also leveraging AWS Trainium instances for specific model training workloads. Custom ASICs optimized for EDA-specific AI tasks are rumored to be in development, although concrete details remain scarce. They use Ray for distributed training and hyperparameter optimization.

Hardware & Compute Infrastructure

Synopsys maintains a significant on-premise data center footprint to handle the immense computational demands of chip design simulation and AI model training. Their infrastructure comprises:

• Data Centers: High-density, liquid-cooled data centers located in secure facilities.
• Chip Architecture: Primarily NVIDIA H200 and A100 GPUs. They are also actively evaluating AMD Instinct MI300 series accelerators.
• Cloud vs. On-Premise: Hybrid approach. Core simulation and model training are performed on-premise for performance and data security reasons. Cloud resources (AWS, Azure) are used for burst capacity and experimentation.
• Custom Silicon: While unconfirmed, strong indicators suggest Synopsys is investing in custom ASICs tailored for accelerating specific EDA AI tasks, particularly in areas like place and route optimization and functional verification.
• Networking Fabric: High-bandwidth InfiniBand HDR (200 Gbps) networking for GPU interconnectivity and data transfer within the on-premise cluster.

Software Platform & Developer Tools

Synopsys provides a comprehensive suite of developer tools and APIs to enable engineers to integrate AI models into their EDA workflows:

• APIs & SDKs: Well-defined Python and C++ APIs for accessing Synopsys's AI models and integrating them into custom design flows. An actively maintained SDK provides tools for data preprocessing, model deployment, and performance monitoring.
• Developer Platform: A cloud-based platform (Synopsys AI Studio) allows engineers to experiment with AI models, train custom models on their designs, and deploy them to on-premise or cloud environments.
• Open-Source Contributions: While primarily focused on commercial products, Synopsys has made contributions to open-source projects related to hardware description languages (HDLs) and verification, demonstrating a commitment to industry standards.
• Key Internal Tools: In-house developed tools for data annotation, model evaluation, and deployment management are critical components of their AI infrastructure. They use MLflow for model tracking and versioning.

Data Pipeline & Storage

Synopsys handles massive datasets generated during chip design and simulation. Their data pipeline architecture includes:

• Data Lakes: A centralized data lake built on Apache Iceberg stores design data, simulation results, and manufacturing data.
• Streaming: Apache Kafka is used for ingesting real-time data from simulation tools and hardware emulators.
• ETL Pipelines: Apache Spark is used for large-scale data processing and transformation. Custom ETL pipelines extract features relevant for AI model training and inference.
• Storage: A combination of object storage (AWS S3) and high-performance parallel file systems (Lustre) provides the necessary scalability and performance for storing and accessing large datasets.

Key Products & How They're Built

Synopsys is integrating AI into its core product offerings:

• DSO.ai (Design Space Optimization): Uses reinforcement learning to automate the place and route process, optimizing chip performance and power consumption. Built on PyTorch and deployed on a cluster of NVIDIA GPUs, DSO.ai drastically reduces design time and improves chip quality.
• VCS Functional Verification: Employs machine learning to identify potential bugs and vulnerabilities in hardware designs. GNNs analyze code dependencies to predict error propagation. This allows for more efficient testing and reduces the risk of costly hardware bugs.
• Fusion Compiler: Uses AI to optimize timing closure and power efficiency during the synthesis process, leading to faster design cycles and improved chip performance. Leverages JAX for high-performance optimization algorithms.

Competitive Moat

Synopsys's competitive advantage stems from several factors:

• Proprietary Data: Synopsys has access to an unparalleled dataset of chip designs and simulation results, providing a significant advantage in training accurate AI models. This data is carefully curated and annotated by domain experts.
• Custom Hardware: Potential development of custom ASICs could provide a performance advantage over competitors relying solely on general-purpose GPUs.
• Network Effects: As more customers use Synopsys's AI-powered tools, the models improve, attracting even more customers and creating a virtuous cycle.
• Talent: Synopsys has assembled a world-class team of AI researchers, EDA experts, and software engineers, creating a formidable barrier to entry.

Stack Scorecard