“One of the most significant benefits we observed after implementing our AI factory was a substantial increase in developer productivity,” said Bryan Donovan, Technical Lead of AI Foundations at Lockheed Martin. “Previously, setting up a development environment on local laptops or workstations was a challenging task due to stringent security constraints, often taking several weeks to complete. Now, with the AI factory powered by DGX SuperPOD, developers can request GPU resources and have their environments up and running in just a few minutes.”
Access to GPUs with high-speed interconnects offered by the DGX SuperPOD has enabled teams to leverage capabilities that were previously unaffordable for individual programs. Training processes that once took weeks are now reduced to days. Forrest commented, “Additionally, our AI factory enables us to share best practices and expertise across Lockheed Martin. We have codified AI best practices into a set of reusable components, pipelines, and playbooks, further improving developer productivity.”
Centralizing compute via the Lockheed Martin AI Factory has led to significant cost savings. Tokens are created when a model breaks down text into smaller units like words, parts of words, or characters, and they are used in the training process to convert textual data into numerical data that the model can understand and learn from. “Using APIs via public cloud requires purchasing tokens to cover the computational costs, but with our on-prem AI factory, we handle tokenization, training, and deployment in-house,” said Forrest. “Our DGX SuperPOD helps us process over one billion tokens per week, enabling fine-tuning, retrieval augmented generation (RAG), or inference on our large language models (LLMs). This solution avoids the escalating costs and significant limitations of fees based on token usage.”
In addition to their DGX SuperPOD, the team still leverages burst-scaling to cloud environments for temporal needs and for non-classified training and data handling. The on-premises DGX systems allow Lockheed Martin to control specific data types and telemetry in a secure environment. “With approved internal data centers, developers can immediately utilize sensitive data, ensuring security and agility across all use cases without waiting for additional approvals. This allows for greater flexibility and faster deployment of AI solutions,” said Donovan.
“Lockheed Martin has a tremendous amount of diverse data types being processed for various AI applications, including time series data, signal intelligence, geospatial data, text, image, and video. The DGX SuperPOD supports a wide range of use cases from data analytics to large-scale inference, on the same platform, enhancing efficiency and reducing costs,” Forrest added.