AI PCs from Intel provide an ideal platform for graph neural network (GNN) workloads, with powerful acceleration from built-in NPUs. GNNs are crucial for tasks like retrieval augmented generation (RAG) in large language models (LLMs) and event-based vision tasks. However, running GNNs involves irregular memory access and control-heavy computations, leading to high inference latency. Discover how GraNNite, a hardware-aware framework developed by Intel, optimizes GNN operation on Intel® NPUs for unparalleled performance and efficiency.

GranNNite follows these stages to optimize GNN operation:

• Enabling NPU mapping while enhancing dynamic graph handling
• Optimizing control-heavy operations and exploiting sparsity
• Balancing efficiency and accuracy with approximation techniques, including quantization

The course targets intermediate to expert developers.

The session covers these topics:

• Demystifies GNN operation on NPUs and provides best practices to address the challenge.
• Covers the three stages to enable and optimize GNNs on NPUs.
• Explores key tools to boost performance and efficiency, including GraphSplit, EffOps, and QuantGr.
• Shows GraNNite’s improvements over default NPU execution.
• Provides insights for balancing accuracy, performance, and energy efficiency for GNN deployment.