Stack Analysis: Autodesk — Rebuilding Design with AI-Powered Foundational Models

Autodesk: Reshaping Creation with Intelligent Design

Autodesk, a global leader in design and make software for architecture, engineering, construction, media, and entertainment industries, is strategically embedding AI throughout its product suite. Their commitment to generative design, automated workflows, and intelligent assistance positions them as a key player in the AI-driven future of creation. They are moving beyond simple automation towards truly intelligent systems that augment and enhance the creative process.

Core AI/ML Stack

Autodesk leverages a mixed approach to AI/ML, blending open-source frameworks with custom-built solutions tailored to their specific domain. Their core stack includes:

• Models: Primarily transformer-based architectures, including custom variants of Vision Transformers (ViTs) and Graph Neural Networks (GNNs) optimized for CAD data. They are increasingly employing diffusion models for generative design tasks. For time-series data in construction and manufacturing, they use LSTM and Transformer models.
• Frameworks: While still supporting TensorFlow 3.x for legacy projects, the focus is shifting towards PyTorch 3.1, leveraging its flexibility and dynamic computation graphs for research and development. JAX is used for high-performance numerical computation and some model training, particularly for generative models.
• Training Infrastructure: Autodesk operates a hybrid cloud model, utilizing both AWS SageMaker and a private cluster of NVIDIA H200 GPUs. They are experimenting with Cerebras Systems' Wafer Scale Engine (WSE-3) for large-scale model training. Distributed training is managed with Ray and Horovod.

Hardware & Compute Infrastructure

Autodesk's compute infrastructure is a mix of cloud-based resources and on-premise data centers. Key components include:

• Data Centers: Two primary data centers located in Oregon and Ireland, housing high-density server racks equipped with NVIDIA A100 and H200 GPUs. A third smaller datacenter is located in Singapore for APAC operations.
• Chip Architecture: Primarily NVIDIA GPUs (A100, H200) for model training and inference. They are also experimenting with AMD Instinct MI400 series GPUs for cost optimization in specific workloads. They have not publicly disclosed any custom ASIC development.
• Cloud vs. On-Prem: A strategic balance, with burst capacity and experimental workloads handled on AWS. Core model training and sensitive data processing remain on-premise.
• Networking Fabric: High-bandwidth, low-latency networking is critical. They utilize InfiniBand (200Gbps) for inter-node communication within their GPU clusters.

Software Platform & Developer Tools

Autodesk is actively fostering a strong developer ecosystem around its AI capabilities:

• APIs: Robust REST APIs for accessing AI-powered features across their product line. These APIs are versioned and well-documented, encouraging third-party integrations. They offer Python and C++ SDKs.
• Developer Platforms: Autodesk Forge remains the core platform for building custom applications on top of their ecosystem. AI features are increasingly integrated into Forge's API surface.
• Open-Source Contributions: Autodesk contributes to several open-source projects, particularly in the areas of CAD data processing and model visualization. They actively contribute to Open3D and related libraries.
• Key Internal Tools: They have developed internal tools for data labeling, model versioning (using MLflow), and automated model deployment (leveraging Kubernetes).

Data Pipeline & Storage

Autodesk's data pipeline is designed to handle the immense scale and complexity of CAD data:

• Data Lakes: A large, centralized data lake built on Apache Hadoop and Spark, storing both structured and unstructured data from various sources (design files, sensor data, usage logs). They are migrating to a cloud-native data lake solution on AWS S3 with Apache Iceberg for improved data governance and querying.
• Streaming: Apache Kafka is used for real-time data ingestion from connected devices and applications. This data is used for real-time monitoring and predictive maintenance.
• ETL Pipelines: Apache Airflow orchestrates complex ETL pipelines for data cleaning, transformation, and feature engineering. Custom data loaders are developed for efficient handling of CAD data formats.

Key Products & How They're Built

• Fusion 360 Generative Design: Powered by custom GNNs trained on a vast dataset of CAD models and simulation results. The system leverages reinforcement learning to optimize designs based on user-defined constraints and objectives. The backend uses Python and C++, deployed on Kubernetes for scalability.
• Autodesk Construction Cloud Insight: Uses time-series models to predict potential project delays and cost overruns based on historical data and real-time sensor readings from construction sites. Natural language processing (NLP) models analyze project documentation to identify potential risks. The frontend is built with React, communicating with a Python-based backend via REST APIs.

Competitive Moat

Autodesk's competitive moat stems from several factors:

• Proprietary Data: Decades of accumulated CAD data, simulation results, and user behavior data provide a significant advantage in training high-performance AI models. This data is difficult for competitors to replicate.
• Specialized AI/ML Talent: Autodesk has invested heavily in attracting and retaining top AI/ML talent, particularly in the areas of geometric deep learning and generative design.
• Deep Domain Expertise: Combining AI/ML expertise with deep understanding of the design and manufacturing industries enables them to build solutions that are highly tailored to specific user needs.

Stack Scorecard