AWS re:Invent 2025 - FINRA: Accelerate Massive Data Processing with NVIDIA on AWS EMR (AIM279)

AWS re:Invent 2025 - FINRA: Accelerate Massive Data Processing with NVIDIA on AWS EMR

Introduction to GPU Acceleration for Apache Spark

• Apache Spark is a widely-used data processing framework across enterprises and organizations
• Data volumes have been growing exponentially in recent years, especially for AI and machine learning workloads
• To handle this data growth, enterprises are turning to GPU acceleration to speed up Spark workloads

The GPU Acceleration Stack

• The solution introduces a plugin that can be easily integrated into existing Spark workflows
• This plugin leverages the NVIDIA GPU layer running on the cloud or on-premises
• This GPU acceleration can provide significant performance improvements and cost savings

Real-World Use Cases

• Companies across various industries have publicly shared their success with this GPU-accelerated Spark technology
• One example is in fraud detection, where billions of records need to be processed using time series analysis - an ideal workload for GPUs
• This resulted in a 14x speedup and 90% cost savings compared to CPU-only processing

FINRA's Journey with GPU-Accelerated Spark

Background on FINRA

• FINRA is a not-for-profit organization responsible for market integrity and investor protection
• They operate over 1 PB of storage in the AWS cloud, processing massive datasets for regulatory compliance and fraud detection

Evaluating GPU Acceleration

• FINRA initially used Apache Hive for their SQL queries, then transitioned to Apache Spark
• When introduced to GPU-accelerated Spark, they ran tests on the TPCDS 9B benchmark
• This resulted in a 50% performance improvement and 50% cost reduction compared to CPU-only Spark

Applying to Production Workloads

• FINRA then applied the GPU-accelerated Spark to their production trading application workloads
• Again, they saw around 50% performance improvements and 45% cost reductions
• However, the initial GPU runs were not optimal, requiring collaboration with NVIDIA to identify and resolve bottlenecks

Integrating GPU Spark into the Data Pipeline

• FINRA's data pipeline involves decompressing, type conversion, and parquet conversion of 100,000 daily CSV files
• By transitioning this pipeline to use GPU-accelerated Spark, they achieved consistent runtime and cost savings
• This required some code changes to leverage Spark DataFrames instead of the less GPU-friendly Dataset API

Lessons Learned and the Path Forward

• Not every workload will see immediate benefits from GPU acceleration - identifying bottlenecks is crucial
• FINRA has established a process to validate CPU vs GPU performance for their workloads
• While CPU remains the default, GPU acceleration is now a strategic part of FINRA's big data technology stack for the future

Key Takeaways

• GPU acceleration can provide significant performance improvements and cost savings for large-scale Spark workloads
• Integrating GPU-accelerated Spark requires some upfront effort to identify and resolve bottlenecks, but can lead to transformative results
• FINRA's experience demonstrates the real-world business impact of this technology, from regulatory compliance to fraud detection
• As GPU hardware and software continue to evolve, GPU acceleration is becoming a strategic part of enterprise big data architectures