แนวคิดหลัก
Efficiently accelerating Graph Neural Networks on real Processing-In-Memory systems is crucial for improving performance and resource utilization in ML models.
บทคัดย่อ
The content discusses the importance of accelerating Graph Neural Networks (GNNs) on real Processing-In-Memory (PIM) systems. It introduces PyGim, an ML framework designed to optimize GNN execution on PIM systems. The article outlines intelligent parallelization techniques for memory-intensive kernels of GNNs tailored for real PIM systems and provides recommendations for software, system, and hardware designers. The evaluation on a real-world PIM system demonstrates the superior performance of PyGim compared to traditional CPU counterparts.
- Introduction to Graph Neural Networks (GNNs) and their significance in ML models.
- Challenges in GNN execution on traditional CPU and GPU systems.
- The concept of Processing-In-Memory (PIM) systems and their potential to alleviate data movement bottlenecks.
- PyGim framework for accelerating GNNs on real PIM systems.
- Evaluation results showcasing the performance benefits of PyGim on a real-world PIM system.
สถิติ
"PyGim outperforms its state-of-the-art CPU counterpart on Intel Xeon by on average 3.04×."
"PyGim achieves higher resource utilization than CPU and GPU systems."
คำพูด
"Graph Neural Networks (GNNs) are emerging ML models that provide high accuracy in node classification and link prediction."
"Our work provides useful recommendations for software, system, and hardware designers."