TGS Scales Seismic AI Training on AWS, Cuts 6-Month Process to 5 Days

Revolutionizing Seismic AI with AWS

TGS, a leading geoscience data provider, has achieved a remarkable breakthrough in training its seismic foundation models (SFMs) by partnering with the AWS Generative AI Innovation Center (GenAIIC). By leveraging Amazon SageMaker HyperPod, TGS slashed the training time for their Vision Transformer-based SFMs from an astounding six months down to just five days. This dramatic acceleration not only enhances efficiency but also expands the models' capacity to analyze vastly larger seismic volumes than ever before.

The Challenge of Geological Insights

For the energy sector, understanding complex 3D seismic data is crucial for identifying geological structures vital for exploration and production. TGS develops advanced SFMs to provide these insights, but training these sophisticated models comes with significant challenges. Managing terabytes of proprietary 3D seismic data, ensuring efficient computation for intensive volumetric data, and expanding the analytical "context window" to capture both local details and broader geological patterns simultaneously, all posed considerable hurdles. TGS recognized the need for a robust, scalable infrastructure to modernize their operations and accelerate model development.

A Breakthrough with Amazon SageMaker HyperPod

The collaboration between TGS and the AWS GenAIIC focused on optimizing TGS's SFM training infrastructure on AWS. The core of their solution involved Amazon SageMaker HyperPod, a purpose-built service designed for distributed training of large-scale machine learning models. The results were transformational:

• Time Reduction: Training time plummeted from six months to just five days.
• Scalability: The solution achieved near-linear scaling for distributed training.
• Expanded Context: Models can now analyze seismic volumes larger than previously possible, offering richer geological understanding.

The technical backbone of this achievement is impressive. The training infrastructure utilized 16 Amazon EC2 P5 instances. Each instance was configured with 8 NVIDIA H200 GPUs, boasting 141GB HBM3e memory per GPU, along with 192 vCPUs, 2048 GB system RAM, and cutting-edge 3200 Gbps EFAv3 networking for ultra-low latency communication.

Optimizing the Data Pipeline

Handling massive datasets efficiently was key. TGS's training data, comprising terabytes of proprietary 3D seismic information in their open-source MDIO format, streams directly from Amazon S3 to the training nodes. This direct streaming approach, instead of relying on intermediate storage layers, proved critical for maintaining high throughput and ensuring GPUs were never idle. This efficient data pipeline played a pivotal role in enabling the rapid training cycles and expanded context windows, as detailed in the comprehensive article on Scaling Seismic Foundation Models on AWS.

Why This Matters for AI and Energy Exploration

This collaboration between TGS and AWS sets a new benchmark for large-scale AI model training in geoscience. For AI practitioners, it demonstrates the power of specialized infrastructure like Amazon SageMaker HyperPod and high-performance EC2 instances (P5 with H200 GPUs) in tackling computationally intensive tasks. It highlights how optimizing data pipelines and leveraging distributed training can unlock unprecedented efficiency and analytical capabilities, especially when dealing with foundation models and massive datasets. For the energy sector, this means TGS can develop and refine their SFMs much faster, leading to quicker, more accurate insights into subsurface geology, ultimately accelerating energy exploration and production workflows. The ability to process larger seismic volumes translates directly into a more comprehensive and nuanced understanding of geological formations, reducing risks and improving decision-making.

For more in-depth technical details on this groundbreaking work, dive into the official blog post: Scaling Seismic Foundation Models on AWS.

Read more: Scaling Seismic Foundation Models on AWS to understand how this partnership is transforming geoscience AI.