AWS for Industries
How Liftoff improved conversion performance and reduced infrastructure costs with Cortex using AWS Graviton
Liftoff is a provider of programmatic advertising technology for mobile applications. AI and machine learning have been part of Liftoff since its beginning. ML models are instrumental in powering Liftoff’s advertising platform and its ability to process, analyze, and apply data to deliver performance for their advertisers.
Mobile apps have become an integral part of our daily lives, making the mobile ad ecosystem large and complex. To scale, apps require consistent and sustainable returns from advertiser spend. They need to model a variety of user actions, everything from installations to making a deposit to signing up for free trials.
To better serve its partners, Liftoff evolved its machine learning models to tackle an evolving set of challenges. The result is Cortex™. Cortex makes over two billion predictions per second, connecting advertisers with audiences, ensuring that the relevant ad is served to the right user, and ultimately improving the return on investment (ROI) for their partners.
In this blog we explore how Liftoff built and scaled Cortex, their AI-powered growth platform, on AWS using the latest neural network technology and AWS Graviton Processors. We discuss how Cortex addresses the challenges of training hundreds of models per day, with over 1 Petabyte of new data daily. All, while lowering infrastructure costs and improving customer conversion performance.
AWS Graviton Processors
AWS Graviton processors are custom-designed by AWS to enable the best price performance for workloads in Amazon EC2. Amazon EC2 instances powered by AWS Graviton processors provide customers up to 40% better price performance over comparable x86-based instances for a wide variety of workloads. Amazon EC2 Graviton instances also use up to 60% less energy than comparable x86-based instances. AWS Graviton is optimized for most modern and cloud-based workloads such as web and application servers, machine learning, databases, containerized workloads, and video encoding.
AWS Graviton4 memory-optimized instances, delivers up to 30% better compute performance, 50% more cores, and 75% more memory bandwidth compared to Graviton3. Graviton4 processors feature advanced security capabilities with always-on memory encryption and enhanced hardware telemetry. These processors power the new Amazon EC2 R8g instances, making them ideal for demanding workloads such as large-scale machine learning inference, in-memory databases, and CPU-intensive application servers. Graviton4’s improved performance and efficiency help customers accelerate their cloud applications while reducing their carbon footprint.
Visit the AWS Graviton Technical Guide to learn more about running workloads on AWS Graviton.
The Birth of Cortex
Prior to the development of Cortex, Liftoff’s advertising technology was built on logistic regression models. While these models served their purpose initially, they had inherent limitations:
- Limited ability for processing complex, non-linear relationships in user data
- Challenges in handling diverse ad formats and user contexts efficiently
- Scalability issues when dealing with rapidly growing datasets
- Increasing infrastructure costs as operations expanded
As Liftoff’s business grew, it became evident that a more optimal solution was necessary to meet the demands of scale and performance in the competitive mobile advertising landscape.
To address these challenges, Liftoff developed Cortex, an AI-powered growth platform. The primary objectives for Cortex were:
- Data Processing at Scale: Handle substantially larger and more diverse datasets, accommodating the growing complexity of user interactions and ad formats.
- Enhanced Model Performance: Significantly improve prediction accuracy and performance, leading to more effective ad targeting and better outcomes for advertisers.
- Rapid Deployment Capability: Increase the frequency of model updates and deployments, allowing for more agile responses to market trends and user behavior changes.
- Cost Optimization: Reduce infrastructure costs while scaling operations, ensuring sustainable growth and improved ROI.
Cortex represents a significant evolution in Liftoff’s technological approach, leveraging advanced machine learning techniques, neural networks, and cloud computing capabilities to overcome the limitations of their previous system.
Training large neural networks requires a significant amount of computational power. At Liftoff, a typical training job uses an Apache Spark Cluster of between 32 to 128 Amazon EC2 Graviton instance for dataset generation. Liftoff realized a significant reduction in the cost of training jobs by running their Apache Spark clusters to EC2 instances powered by AWS Graviton processors (m8gd.8xlarge, m7gd.8xlarge, and m6gd.8xlarge).
Figure 1. shows the high-level architecture of Cortex.
Figure 1: Cortex Architecture
Optimizing Cortex for AWS Graviton
AWS Graviton, unlike x86-based processors, does not use Simultaneous Multi-Threading (SMT) or hyper-threading. This means that a vCPU is a physical core leading to better isolation and improved performance for workloads such as Cortex.
Liftoff’s engineering team further maximized the benefits of AWS Graviton by implementing several optimizations. First, they utilized the ARM NEON instructions to accelerate floating-point operations. They also used the latest compilers with the ARM Neoverse compiler flags.
Servicing Mobile Ads at Scale with AWS Graviton
Cortex reacts quickly to shifting market conditions and campaign fluctuations, resulting in more efficient ad spend and performance. Making two billion predictions per second to serve mobile ads in real-time at scale requires high-performance computing. To do this, Liftoff also migrated their bidding servers from Intel-based Amazon EC2 c5 instances to AWS Graviton-based Amazon EC2 c7g and c8g instances. By doing this, they realized both an overall performance improvement along with a reduction in cost. Liftoff’s bidding servers are written in Golang. Golang’s cross-compilation capabilities made the migration seamless. For certain performance-critical calculations, the engineering team relied on optimized Go assembly code. Using AWS Graviton was relatively easy and saw a significant improvement over pure Golang.
Conclusion
In this blog post, we discussed how Liftoff Mobile, a leader in programmatic advertising for mobile applications, built their AI-powered growth platform, Cortex, on AWS using AWS Graviton EC2 instance types.
By doing so, Liftoff improved customer conversion performance while deploying more powerful models and lowered their infrastructure cost. Liftoff officially launched Cortex in October 2024.
Advertisers can learn how to use Cortex to improve their Return on Ad Spend (ROAS) and increase revenue by visiting ‘Introducing Cortex, Liftoff’s Next Generation ML Platform’ or get started with Liftoff.
To learn more about AWS Graviton Processors and how other customers are achieving better both price-performance and meeting their sustainability targets, visit the following resources: