Stack Analysis: Adobe — Weaving AI into the Creative Fabric

Company Overview

Adobe is a leading provider of creative software, document management tools, and digital marketing solutions. They are at the forefront of integrating AI and machine learning into their products, enhancing user experiences, and automating complex tasks for creative professionals. Their AI-powered features are increasingly critical to maintaining their competitive edge in the rapidly evolving digital landscape.

Core AI/ML Stack

Adobe's AI/ML stack is a hybrid approach, blending industry-standard open-source frameworks with proprietary tools tailored to their specific needs. They heavily utilize PyTorch 3.2 for model development and training, leveraging its flexibility and active community support. For large language models used in features like enhanced text prompting in Photoshop and Illustrator, they've adopted the JAX framework, taking advantage of its superior performance on TPUs. They also maintain an internal framework, 'Project Firefly ML,' built on top of TensorFlow, focused on optimizing image and video processing models.

For model training, Adobe relies on a combination of NVIDIA A200 and H100 GPU clusters hosted both on-premise and within AWS SageMaker and Google Cloud AI Platform. Their internal 'MercuryTrain' platform orchestrates distributed training across these resources, managing data parallelism and model parallelism. They are also experimenting with custom ASICs, developed in collaboration with TSMC, designed for accelerated inferencing of their most popular AI features, particularly those involving generative AI and real-time video analysis.

Hardware & Compute Infrastructure

Adobe operates a distributed hybrid cloud infrastructure. Core services and large-scale training jobs are run on AWS and Google Cloud. They maintain several on-premise data centers in California and Ireland, primarily for handling sensitive user data and running latency-critical applications. The company is actively transitioning towards a predominantly cloud-based infrastructure. Their custom ASICs are deployed on-premise in specialized inference servers optimized for low latency and high throughput. Their networking fabric utilizes RDMA over Converged Ethernet (RoCE) to ensure high-bandwidth, low-latency communication within the GPU clusters.

Software Platform & Developer Tools

Adobe provides a rich set of APIs and SDKs for developers to integrate AI-powered features into their own applications and workflows. The Adobe Creative Cloud SDK allows third-party developers to build plugins and extensions for Photoshop, Illustrator, and other Adobe applications. They also offer the Adobe Experience Platform (AEP) Launch SDK for integrating data collection and marketing automation tools. Adobe is a significant contributor to open-source projects, particularly in the areas of image processing and machine learning. Key internal tools include 'Phoenix', a model serving platform based on gRPC, and 'Hydra', a data pipeline management system built on Apache Beam.

Data Pipeline & Storage

Adobe's data pipeline is designed to handle massive volumes of user data generated across their various products. They use a combination of data lakes and streaming platforms to ingest, process, and store this data at scale. Their primary data lake is built on Apache Hadoop and Apache Spark, with data stored in Apache Parquet format. They utilize Apache Kafka and Apache Flink for real-time data streaming, enabling features such as personalized recommendations and anomaly detection. Their ETL pipelines are orchestrated using Apache Airflow and are designed to handle both batch and streaming data. For long-term storage, they leverage Amazon S3 and Google Cloud Storage.

Key Products & How They're Built

• Photoshop: AI features in Photoshop, such as Content-Aware Fill, Sky Replacement, and Neural Filters, are powered by models trained using PyTorch and 'Project Firefly ML'. These models are served using the 'Phoenix' platform and often accelerated by their custom ASICs. The models are continuously refined using data from user interactions and feedback.
• Adobe Sensei: Sensei acts as the AI engine across Adobe’s products. Under the hood, it utilizes a combination of the aforementioned frameworks and infrastructure, exposing APIs for various AI capabilities. Their generative AI feature is powered by large language models running on JAX and TPUs.
• Adobe Experience Cloud: This platform leverages AI for personalized marketing campaigns and customer experience optimization. The data ingested by AEP is processed using Spark and Flink, and models for customer segmentation and predictive analytics are trained using PyTorch and TensorFlow.

Competitive Moat

Adobe's competitive moat lies in a combination of factors. Their vast proprietary dataset of creative assets (images, videos, fonts) is a significant advantage, providing them with a unique training ground for their AI models. The deep integration of AI into their existing product ecosystem creates strong network effects, making it difficult for competitors to displace them. Finally, their investment in custom silicon and specialized AI engineering talent gives them a performance edge in critical areas such as generative AI and real-time video processing.

Stack Scorecard