Case Study: AMD achieves zero-copy GraphRAG and multi-hop enterprise context with PuppyGraph

AMD builds zero-copy GraphRAG on Iceberg with PuppyGraph for petabyte-scale data

AMD, a major semiconductor company, faced a challenge in making its enterprise data, consolidated on an Apache Iceberg foundation, usable as connected context for AI applications. Reconstructing complex relationships across many tables through joins became too difficult for LLMs to handle, preventing them from providing grounded answers. The company needed a way to enable multi-hop queries for GraphRAG without the burden of creating and maintaining a separate, synchronized copy of its petabyte-scale data in a dedicated graph store.

The vendor PuppyGraph provided a solution by adding a zero-copy knowledge graph layer directly over AMD's existing Iceberg tables. This allowed for multi-hop relationship queries without data loading or separate synchronization pipelines. The result was an architectural extension that made connected context available to chatbots and agents while preserving data security and access controls. This enabled AMD to support a permission-aware chatbot and lay the groundwork for future multimodal GraphRAG, all without building a parallel data stack.

AMD